CN-121979970-A - Knowledge question and answer method and device

Abstract

The embodiment of the application provides a knowledge question-answering method and device. The method comprises the steps of using a knowledge base to check a target question-answer text in a question-answer knowledge base by an agent to obtain a check result, sending the check result to a maintainer to conduct content correction on the question-answer knowledge base by the maintainer, using a question-answer case learning agent to analyze the question-answer text in a question-answer system and generating a question-answer case base according to the analysis result, wherein the question-answer case base comprises a question-answer error case and a question-answer correct case, when a question text input by a user is received, searching a target question-answer case with similarity higher than a similarity threshold value from the question-answer case base, optimizing a question-answer system according to system optimization item parameters in the target question-answer case, and generating an answer corresponding to the question text according to the question-answer knowledge base and the target question-answer case after optimization.

Inventors

• ZHANG JINGWEI
• QI CHEN
• Jin Qiuxiao
• GENG RUIYING
• MA XINHONG

Assignees

• 中国民生银行股份有限公司

Dates

Publication Date

20260505

Application Date

20251211

Claims (16)

1. 1. A knowledge question-answering method, the method comprising: Using a knowledge base checking agent to check a target question-answer text in a question-answer knowledge base to obtain a check result, and sending the check result to a maintainer so that the maintainer can correct the content of the question-answer knowledge base; Analyzing a question and answer text in a question and answer system by using a question and answer case learning agent, and generating a question and answer case library according to an analysis result, wherein the question and answer case library comprises question and answer error cases and question and answer correct cases; When a question text input by a user is received, searching a target question-answer case with similarity higher than a similarity threshold value from the question text library; Optimizing a question-answering system according to system optimization item parameters in the target question-answering case, and generating an answer corresponding to the question text according to the question-answering knowledge base and the target question-answering case after optimization.
2. 2. The method of claim 1, wherein the using the knowledge base checking agent to check the target question-answer text in the question-answer knowledge base to obtain a check result comprises: checking the intelligent agent by using the knowledge base, and determining whether the checking condition of the question-answer knowledge base is reached; When determining that the checking condition of the question-answering knowledge base is met, acquiring a target question-answering text from the historical question-answering text of the question-answering system by a question acquisition tool, wherein the target question-answering text refers to a question-answering text with a question-answering frequency greater than a first frequency threshold value in the historical question-answering text of the question-answering system, and the first frequency threshold value is dynamically set according to the total amount of the knowledge base, the equipment capacity and the task processable time; traversing the question-answer knowledge base by using a knowledge base traversing tool aiming at each target question-answer text to obtain target information fragments associated with the target question-answer text, and constructing a temporary sub-knowledge base according to the target information fragments; performing multi-dimensional detection on the temporary sub-knowledge base to obtain a multi-dimensional detection result; and summarizing the multi-dimensional detection results to generate a detection result comprising a problem abstract, a knowledge problem list and corresponding optimization suggestions, wherein the knowledge problem list is attached with problem source associated information.
3. 3. The method of claim 2, wherein said determining that the check condition of the question-answer knowledge base is reached comprises: Under the condition that the interval time length for checking the question-answer knowledge base reaches the set time length, determining that the checking condition of the question-answer knowledge base is reached; Determining that the checking condition of the question-answer knowledge base is met under the condition that the number of newly added error cases in the question-answer case base reaches a number threshold value, wherein the newly added error cases refer to newly added error cases of the question-answer knowledge base after the last checking; And determining the checking condition reaching the question-answering knowledge base under the condition that the variation amplitude of a pre-stored specified question list in the question-answering system is larger than an amplitude threshold, wherein questions with question-answering frequencies larger than a second frequency threshold are recorded in the specified question list.
4. 4. The method according to claim 2, wherein the performing the multi-dimensional detection on the temporary sub-knowledge base to obtain a multi-dimensional detection result includes: performing contradiction expression detection on the information in the temporary sub-knowledge base, identifying contradictory information, and associating and recording source information and complete content information of each contradiction information to obtain a contradiction expression detection result; Detecting the outdated content of the information containing the timeliness identifier in the temporary sub-knowledge base, comparing the timeliness identifier with the current date, screening and recording the outdated information to obtain an outdated content detection result; Carrying out text expression detection on the information in the temporary sub-knowledge base according to a preset user visual angle to obtain an expression detection result; When judging that the information in the temporary sub-knowledge base does not meet the reply requirement of the target question-answer text by using a preset model, acquiring missing key information of the temporary sub-knowledge base output by the preset model, and taking the missing key information as a knowledge gap detection result; And taking the contradictory expression detection result, the expiration content detection result, the expression detection result and the knowledge gap detection result as the multi-dimensional detection result.
5. 5. The method of claim 1, wherein the learning by questioning and answering case agent comprises a miscase learning agent and a positive case learning agent, The learning agent using the question and answer case analyzes the question and answer text in the question and answer system, and generates a question and answer case library according to the analysis result, comprising: Using the wrong case learning agent to carry out mining analysis on wrong questions and answers of the questions and answers text, wrong reason annotation information of the wrong questions and answers text and correct answer annotation information to obtain wrong questions and answers; Analyzing correct question-answering texts marked in the question-answering texts by using the positive case learning intelligent agent to obtain question-answering correct cases; And generating the question-answer case library according to the question-answer error case and the question-answer correct case.
6. 6. The method of claim 5, wherein the miscase learning agent comprises a master coordination agent, a research agent, and a regression testing agent, The learning agent for wrong cases carries out mining analysis on wrong questions and answers of the questions and answers text, wrong reason annotation information of the wrong questions and answers text and correct answer annotation information to obtain wrong questions and answers cases, and the learning agent comprises the following components: Determining, by the master coordination agent, an error type corresponding to the error question-answer text based on the error cause annotation information, the associated interaction information of the error question-answer text, and the correct answer annotation information; analyzing the error type to the error question-answer text by the research intelligent agent by taking the correct answer labeling information as a standard to obtain a fault cause analysis result; Invoking the regression testing agent, taking the correct answer labeling information as a testing evaluation benchmark, testing the optimization strategy in the error factor analysis result, and generating an evaluation report containing a strategy effect and a risk evaluation result; and calling the main coordination intelligent agent, collecting the error cause analysis result and the evaluation report based on the correct answer labeling information to obtain improved suggestion information, and integrating the improved suggestion information and the wrong question-answer text to obtain the question-answer wrong case.
7. 7. The method of claim 6, wherein the research agent comprises a knowledge base content review agent, a search process simulation and analysis agent, a generation link analysis agent, The step of analyzing the error type to the error question-answer text by the research intelligent agent with the correct answer labeling information as a standard to obtain error cause analysis results, comprises the following steps: invoking the knowledge base content examination agent, and when the answer label corresponding to the question-answer text is wrong, examining the content in the question-answer knowledge base to obtain a content examination result; invoking the retrieval process to simulate and analyze an agent, and analyzing a retrieval system corresponding to the wrong question-answer text when the answer context corresponding to the question-answer text is incomplete or irrelevant to the question-answer text, so as to obtain an analysis and check result; Invoking the generation link analysis intelligent agent, and carrying out generation link analysis on the large model when the error type is the error type of the large model or the error type of answer synthesis so as to obtain a generation link analysis result; And determining the content examination result, the analysis and inspection result and the generation link analysis result as the error cause analysis result.
8. 8. The method of claim 7, wherein the invoking the knowledge base content review agent to review the content in the knowledge base for content review results when the answer label corresponding to the question-answer text is wrong comprises: Invoking the knowledge base content to examine an agent, and extracting a target knowledge point from the context information corresponding to the answer when the answer corresponding to the question-answer text is marked as error; Invoking a knowledge base retrieval tool, and retrieving the question-answer knowledge base by taking the target knowledge point as a retrieval word to obtain a corresponding retrieval result; comparing the target knowledge points with the search result, identifying mutually contradictory information, and associating and recording source information of each contradictory information; Generating error cause analysis conclusion and corresponding contradiction description information based on the contradiction information and the source information; and taking the error cause analysis conclusion, the contradiction description information and each contradiction information as the content examination result.
9. 9. The method according to claim 7, wherein the analyzing the search system corresponding to the wrong question-answer text to obtain the analysis and inspection result includes: Invoking the retrieval process to simulate and analyze an agent, and invoking a question rewriting tool to rewrite an original question text corresponding to the question and answer text when the answer context corresponding to the question and answer text is incomplete or irrelevant to the question and answer text, and using a knowledge base retrieval tool to simulate and retrieve the rewritten question text to obtain a simulated retrieval result; using the knowledge base searching tool to search the original problem text by using different searching parameters to obtain a plurality of searching results; determining a fault cause analysis result and description information corresponding to the fault cause analysis result according to the simulation search result and the plurality of search results, and taking the fault cause analysis result and the description information as the analysis and inspection result; And if the simulated search result and the plurality of search results do not meet the requirement of the original problem text, searching the original problem text by using a knowledge base traversing tool, and determining missing information associated with the original problem text in the question-answer knowledge base according to the content returned by searching to obtain the analysis and inspection result.
10. 10. The method of claim 7, wherein the invoking the generation link analysis agent performs generation link analysis on the large model to obtain a generation link analysis result when the error type is a large model error type or an answer synthesis error type, comprising: invoking the generating link analysis agent, and when the error type is a large model error type or an answer synthesis error type, analyzing the prompting word of the generating link of the original problem text by using a prompting word analysis tool to obtain a prompting word analysis result; analyzing the model reasoning process of the large model by using a large model analysis tool to obtain a model capacity analysis result; and determining the generating link analysis result according to the prompt word analysis result and the model capacity analysis result.
11. 11. The method of claim 6, wherein the invoking the regression testing agent to test the optimization strategy in the miscause analysis result with the correct answer label information as a test evaluation benchmark to generate an evaluation report comprising a strategy effect and a risk evaluation result comprises: invoking the regression testing intelligent agent to obtain an optimization strategy to be tested, wherein the optimization strategy to be tested is a strategy in an error question-answer case; Simultaneously inputting each question in the test set into two parallel question-answering systems, wherein one question-answering system of the two question-answering systems applies the optimizing strategy to be tested, and the other question-answering system does not apply the optimizing strategy to be tested; Obtaining output results of two question-answering systems corresponding to each question, and comparing the output results with standard answers in the test set to obtain a strategy effect of the optimization strategy to be tested; Performing risk assessment on the optimization strategy to be tested according to the comparison result to obtain a risk assessment result; And generating the evaluation report according to the strategy effect and the risk evaluation result.
12. 12. The method according to claim 1, wherein the method further comprises: when determining that the clustering analysis conditions of the question-answer case library are met, using the question-answer case learning agent to perform clustering processing on question-answer error cases in the question-answer case library to obtain clustered error cases of the question-answer error cases; Performing error commonality analysis on the clustering error cases to obtain error commonality characteristics of the clustering error cases; and generating a general solution corresponding to the clustering error case according to the error commonality characteristic.
13. 13. The method of claim 5, wherein the analyzing, using the positive learning agent, the correct question-answer text marked in the question-answer text to obtain the question-answer correct case includes: using the positive learning agent to perform strategy stripping test and/or substitution contrast test on the marked correct question-answer text to obtain a test result; analyzing and obtaining successful attribution conclusion information corresponding to the correct question-answer text according to the test result, and generating an optimization method according to the successful attribution conclusion information; and generating the question-answering correct case according to the original question text of the correct question-answering text and the optimization method.
14. 14. The method of claim 1, wherein upon receiving a question text entered by a user, searching for a target question-answer case from the question-answer case library that has a similarity to the question text above a similarity threshold, comprising: when the question text is received, a mixed retrieval algorithm and a reordering algorithm are used for retrieving and reordering the question text, and a standard answer corresponding to the question text is obtained; And when the standard answer does not meet the answer requirement of the question text, searching a target question-answer case with similarity higher than a similarity threshold value with the question text from the question-answer case library.
15. 15. The method according to claim 1, wherein optimizing the question-answering system according to the system optimization term parameters in the target question-answering case, and generating an answer corresponding to the question text according to the question-answering knowledge base and the target question-answering case after optimization, comprises: When the target question-answering case is a question-answering error case, optimizing a system module matched with the system optimization term parameter in the question-answering system according to the system optimization term parameter in the target question-answering case, wherein the system module comprises at least one of a search engine, a large model and a knowledge base; when the question text is a real-time question text, identifying the user intention of the user according to the user portrait of the user and the question text, generating an initial answer according to the user intention, a question-answer knowledge base retrieval result and reference information in a target question-answer case, and processing the initial answer by combining the user image to generate a target answer of the question text; When the question text is a text of a non-real-time question, generating an initial answer based on the question text, a question and answer knowledge base search result and reference information in a target question and answer case, generating an answer tracing report by a large model according to process information for generating an answer, integrating the initial answer and the answer tracing report into a final answer, and using the answer tracing report for performing later quality inspection.
16. 16. A knowledge question-answering apparatus, the apparatus comprising: The system comprises an inspection result acquisition module, a maintenance personnel and a content correction module, wherein the inspection result acquisition module is used for inspecting a target question-answer text in a question-answer knowledge base by using a knowledge base inspection agent to obtain an inspection result, and sending the inspection result to the maintenance personnel so as to carry out content correction on the question-answer knowledge base by the maintenance personnel; The case library generation module is used for analyzing the question-answer text in the question-answer system by using the question-answer case learning agent and generating a question-answer case library according to the analysis result, wherein the question-answer case library comprises question-answer error cases and question-answer correct cases; The target case searching module is used for searching a target question-answer case with similarity higher than a similarity threshold value from the question-answer case library when receiving a question text input by a user; And the answer generation module is used for optimizing the question-answer system according to the system optimization item parameters in the target question-answer case, and generating an answer corresponding to the question text according to the question-answer knowledge base and the target question-answer case after optimization.

Description

Knowledge question and answer method and device Technical Field The application relates to the technical field of knowledge question and answer, in particular to a knowledge question and answer method and device. Background With the popularization of the mobile internet and the rapid development of artificial intelligence technology, the requirements of customers on bank customer service are also increasing increasingly, and all-weather, instant and accurate solutions to the problems are expected. Current banking question-answering systems generally lack the ability to efficiently optimize and iterate knowledge. Most banking systems record only negative feedback given by the user (e.g., click "not help" or score, etc.), and then manually analyze these negative feedback cases periodically in an attempt to find the root cause of the problem and repair it. This process is not only labor intensive, but also depends on personal ability both on the accuracy of the positioning and the effectiveness of the repair measures. The supervision fine tuning training can improve the question-answering capability of the large model in the vertical field, and some prior art utilizes the large model to automatically extract supervision fine tuning data from the existing business documents and the historical dialogue records. However, the requirements for bank hardware investment (such as display card server resources) are high whether the training data is generated or the large model supervised fine tuning training is performed. The graphics card server resources are intense, business knowledge is frequently changed, and the large model base model is frequently updated, so that incremental training for improving the large model business knowledge question-answering capability is high in cost. Thus, the scheme of automatically generating large model incremental training data from system history data and performing incremental training is not practical for most banking question-and-answer scenarios. Disclosure of Invention In order to solve the technical problems, the embodiment of the application provides a knowledge question answering method and device. In a first aspect, an embodiment of the present application provides a knowledge question-answering method, where the method includes: Using a knowledge base checking agent to check a target question-answer text in a question-answer knowledge base to obtain a check result, and sending the check result to a maintainer so that the maintainer can correct the content of the question-answer knowledge base; Analyzing a question and answer text in a question and answer system by using a question and answer case learning agent, and generating a question and answer case library according to an analysis result, wherein the question and answer case library comprises question and answer error cases and question and answer correct cases; When a question text input by a user is received, searching a target question-answer case with similarity higher than a similarity threshold value from the question text library; Optimizing a question-answering system according to system optimization item parameters in the target question-answering case, and generating an answer corresponding to the question text according to the question-answering knowledge base and the target question-answering case after optimization. In a second aspect, an embodiment of the present application provides a knowledge question-answering apparatus, where the apparatus includes: The system comprises an inspection result acquisition module, a maintenance personnel and a content correction module, wherein the inspection result acquisition module is used for inspecting a target question-answer text in a question-answer knowledge base by using a knowledge base inspection agent to obtain an inspection result, and sending the inspection result to the maintenance personnel so as to carry out content correction on the question-answer knowledge base by the maintenance personnel; The case library generation module is used for analyzing the question-answer text in the question-answer system by using the question-answer case learning agent and generating a question-answer case library according to the analysis result, wherein the question-answer case library comprises question-answer error cases and question-answer correct cases; The target case searching module is used for searching a target question-answer case with similarity higher than a similarity threshold value from the question-answer case library when receiving a question text input by a user; And the answer generation module is used for optimizing the question-answer system according to the system optimization item parameters in the target question-answer case, and generating an answer corresponding to the question text according to the question-answer knowledge base and the target question-answer case after optimization. In a third aspect, an embodiment of the present applic