A dynamic question-answer matching method and system based on context awareness

Abstract

The present invention relates to the field of artificial intelligence and natural language processing technology, in particular to a context-aware dynamic question-answer matching method and system. The method comprises: obtaining multi-round conversation history and user background information, generating a semantic understanding model of the current conversation; matching the most relevant answer from the question-answer library based on the semantic understanding model and the current user input, and dynamically adjusting the matching strategy; obtaining real-time feedback data from the user, updating the personalized recommendation algorithm based on changes in user interests, and generating a personalized answer recommendation list; dynamically updating the question-answer library content based on the personalized recommendation list and user behavior data; inputting user feedback and updated question-answer library data into a personalized recommendation system to generate a final personalized answer recommendation. Through context awareness and real-time dynamic adjustment, the present invention significantly improves the matching accuracy of the question-answer system, enhances the user's interactive experience, and enhances the adaptability and flexibility of the system.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and natural language processing, and in particular to a context-aware dynamic question-answer matching method and system. Background Art

[0002] With the widespread application of intelligent question-answering systems, traditional question-answering matching methods mainly rely on static matching based on keywords, which cannot effectively consider users' historical conversations and personalized needs, resulting in problems such as inaccurate semantic understanding, irrelevant answer recommendations, and poor timeliness in complex and changing application scenarios. In addition, existing question-answering systems are unable to adjust and optimize recommendation strategies in real time, especially when users' interests and needs change, and are unable to respond dynamically, resulting in a decline in user experience. Therefore, how to dynamically adjust personalized question-answering recommendation strategies based on users' real-time feedback data, behavior analysis data, and contextual information, thereby improving users' interactive experience and the relevance and timeliness of answer recommendations, has become a key technical issue that needs to be urgently addressed in intelligent question-answering systems. Summary of the invention

[0003] The present invention provides a context-aware dynamic question-answer matching method and system to solve the problem of how to dynamically adjust the personalized question-answer recommendation strategy based on the user's real-time feedback data, behavior analysis data and context information, thereby improving the user's interactive experience and the relevance and timeliness of answer recommendations.

[0004] On the one hand, in order to solve the above technical problems, the present invention provides a context-aware dynamic question and answer matching method, including: obtaining multi-round conversation history and user background information, performing context-aware analysis, and generating a semantic understanding model of the current conversation; matching the most relevant answers from the question and answer library based on the semantic understanding model and the current user input, and dynamically adjusting the matching strategy to generate preliminary answer candidates; obtaining real-time feedback data and behavior analysis data from users, updating a personalized recommendation algorithm based on changes in user interests and needs, and generating a personalized answer recommendation list; dynamically updating the question and answer library content according to the personalized recommendation list and user behavior data to generate an updated question and answer library; inputting user feedback and updated question and answer library data into a personalized recommendation system, readjusting the recommendation strategy, and generating a final personalized answer recommendation.

[0005] Optionally, the acquiring of multiple rounds of conversation history and user background information, performing context-aware analysis, and generating a semantic understanding model of the current conversation includes:

[0006] Extract the most recent Round of conversation records, where the conversation records are represented as a collection ,in For the The content of the round of dialogue, including user input text and system-generated answers; the data field includes the dialogue text , Timestamp , and user behavior annotation ;

[0007] Record of the conversation Perform text cleaning, including removing special characters, correcting spelling errors, and unifying language style;

[0008] For the text sequence after text cleaning, word segmentation technology is used to decompose it into a set of vocabulary units. ,in Indicates the number of word segments in this round of dialogue;

[0009] The word segmentation result Convert to embedding vector ;

[0010] Finally, a multi-round dialogue embedding matrix is ​​formed , as the input of the context-aware analysis module;

[0011] Obtain user static information from the user profile database, including interest preferences , Language style , Occupational category , forming a static feature vector set ;

[0012] Extract user dynamic features from real-time interaction logs, including current device type , Geographical location , and conversation sentiment distribution , generating dynamic feature vectors ;

[0013] Concatenate the dynamic feature vector with the static feature vector to form a comprehensive user background feature vector .

[0014] The comprehensive feature vector For normalization, use the standardization formula:

[0015] ;

[0016] in, is the feature mean, is the standard deviation, and the normalized features are obtained ;

[0017] Embedding multi-turn dialogue into a matrix and user background feature vector Input context-aware model , the model input is:

[0018] ;

[0019] in, represents the joint input data for contextual analysis;

[0020] based on The model performs semantic decomposition and generates semantic vectors ,in For the Dimensional semantic features;

[0021] Furthermore, the context weight distribution is calculated , calculated using the attention mechanism:

[0022] ;

[0023] in, Indicates The weight of the turn, Represents the dot product operation;

[0024] Semantic vector and weight Perform weighted aggregation to generate the semantic understanding vector of the current conversation:

[0025] ;

[0026] Finally, we get the semantic understanding model of the current conversation , used in the subsequent question-answer matching step.

[0027] Optionally, based on the semantic understanding model and the current user input, the most relevant answer is matched from the question and answer library, and the matching strategy is dynamically adjusted to generate preliminary answer candidates, including:

[0028] Get the current input text from the user , for the Perform text cleaning, including removing noise characters and unifying grammatical rules;

[0029] Using the same semantic embedding model Encoding, generating input semantic vector , specifically:

[0030] ;in represents the semantic embedding function;

[0031] Generation-based semantic understanding model ,calculate and The correlation , using the cosine similarity formula:

[0032] ;

[0033] in, is the relevance score of semantic matching, represents the dot product, Represents the L2 norm of the vector;

[0034] The correlation Map to The specific normalization formula is:

[0035] ;

[0036] in, and are the maximum and minimum correlations in the matching process, respectively;

[0037] Extract all question texts from the Q&A database , forming a set of problems ,in The total number of questions in the Q&A database. The same embedding model is used as for S130 and S210 Encode the problem set Q and generate an embedding matrix:

[0038] ;in, ;

[0039] For the user's current input semantic vector , calculate them one by one with The similarity of all question embedding vectors in , forming a similarity score vector , the specific calculation formula is:

[0040] ;

[0041] in, For user input and questions similarity;

[0042] Based on similarity score Before selection The questions with the highest similarity constitute the initial candidate question set , the corresponding answer set is ;

[0043] and It will be used as input to the dynamic adjustment strategy module;

[0044] Combine the generated context weights , adjust the candidate answer set The priority, specifically, the matching score for each candidate answer , recalculate the weighted score :

[0045] ;

[0046] in, For the The semantic contribution value of a turn-based conversation is generated based on the semantic understanding model.

[0047] Based on weighted score For the candidate answer set Re-sort and generate the final sorted answer candidate set ;

[0048] Filter below the set threshold To ensure the high relevance of the preliminary candidate answers, the final ranked candidate answer set Output, providing input for the subsequent answer verification module.

[0049] Optionally, obtain real-time feedback data and behavior analysis data from users, update the personalized recommendation algorithm based on changes in user interests and needs, and generate a personalized answer recommendation list, including:

[0050] The user behavior sequence is obtained by collecting data in real time through the user interface, including the user's click behavior, input revision data, dwell time, mouse trajectory, and option selection records in the current question-answering system interaction. ;

[0051] in: Indicates Data points of each interaction behavior, which may include timestamp, interaction type, and feedback content; Indicates the time period of the current session;

[0052] Extract feature vectors from the above user behavior sequence , including interest preference indicators , interaction frequency index and historical satisfaction ratings These features serve as the basic input for subsequent analysis and are passed to S320 for dynamic evaluation of interest changes;

[0053] Based on the extracted feature vector , evaluate the model using context variation , calculate the user interest change coefficient and the demand offset parameter , the calculation formula is as follows:

[0054] ;

[0055] ;

[0056] in: is the interest preference index of the previous round; Rate current historical satisfaction; Representation characteristics The weight of is dynamically adjusted by the feature importance analysis module;

[0057] Integration , generate a user demand change model , used to update the personalized recommendation algorithm;

[0058] Based on the user demand change model , dynamically adjust the weight matrix of the recommendation algorithm and priority strategy, retraining the personalized recommendation model , the specific steps include:

[0059] according to Demand offset parameter in , update the feature weights of the recommendation model: ,in is the learning rate, which controls the weight update amplitude;

[0060] Use the latest recommendation algorithm to sort the candidate answers in the question and answer database and generate a personalized recommendation list .

[0061] in, represents the i-th candidate answer, arranged in descending order of relevance; is the number of recommended answers, which is controlled by the threshold set by the system. Passed to subsequent modules for further optimization and output.

[0062] Optionally, dynamically updating the content of the question and answer library according to the personalized recommendation list and the user behavior data to generate an updated question and answer library includes:

[0063] Obtain personalized recommendation lists and user behavior data to evaluate the timeliness and relevance of question-answer pairs. This is achieved by combining personalized recommendation lists with user behavior data. and user behavior data , for each question-answer pair in the question-answer database First, from the recommended list Extract key features , which represents the user’s current topic interests. Extract interaction frequency from and user satisfaction ratings Distribution characteristics of . Construct a comprehensive evaluation function:

[0064] ;

[0065] in, For question and answer The timestamp of its generation, Score historical matches, is the current timestamp, and are the global distributions of user satisfaction and interaction frequency, is the weight parameter, Represents the normalized functions of time correlation, topic correlation, matching score correlation, satisfaction correlation and interaction frequency correlation respectively. , construct the filtering rules, which will satisfy:

[0066] A collection of question-answer pairs As the preliminary result of high-relevance question-answer pairs, question-answer pairs below this threshold are classified into the low-relevance set. ;

[0067] Based on the evaluation results of the question-answer pairs, identify the question-answer pairs with low relevance or low timeliness, and dynamically update the content of the question-answer library. The implementation method is as follows: for the collection The question-answer pairs in the and behavioral data Conduct in-depth analysis to focus on identifying the root causes of low timeliness or low relevance. For low timeliness question-answer pairs, use the formula: ;in, is the time attribute update function, Represents a dynamic correction parameter to correct the utility of generating timestamps. , combined with the theme correction function and content optimization functions , perform the following update:

[0068] ;

[0069] If the relevance evaluation value of the updated question-answer pair If the value is still below the minimum threshold, the question-answer pair will be permanently removed from the question-answer database. The weight parameters and topic matching of the question-answer pairs in the database are further optimized through feedback data to ensure the content accuracy of the question-answer database.

[0070] Dynamically integrate the newly generated question-answer pairs with the updated question-answer library to form the final version of the question-answer library. The implementation method is: and the updated set of question and answer pairs Conduct dynamic integration;

[0071] Using semantic similarity function ,calculate and Content similarity between question and answer pairs , merge high similarity question and answer pairs to avoid content duplication. The formula is:

[0072] ;

[0073] in, Represents the semantic similarity function. Reassign priority weights to the question-answer pairs in the integrated question-answer library , the weight calculation formula is:

[0074] ;

[0075] in, represents the correlation evaluation value, is the feedback weight adjustment function;

[0076] Build a new version of the Q&A library , and optimize the index structure to ensure that the retrieval efficiency meets the system requirements. Dynamic availability can be maintained through an incremental update mechanism.

[0077] Optionally, the user feedback and the updated question-answer database data are input into the personalized recommendation system, the recommendation strategy is readjusted, and the final personalized answer recommendation is generated, including:

[0078] The user's feedback data in the previous round of interaction and updated Q&A library Input into the personalized recommendation system, user feedback data Including satisfaction , interaction frequency , and timing characteristics , used to evaluate users’ current acceptance of recommended content and changes in demand;

[0079] Through the feedback enhancement mechanism, the recommendation weight of each question-answer pair is re-evaluated based on the following formula:

[0080]

[0081] in, is the relevance score of the question-answer pair, represents the feedback enhancement function based on user satisfaction, represents the feedback enhancement function based on the interaction frequency, is the weight parameter, U and F are the distribution data of global user satisfaction and interaction frequency;

[0082] Combination The result of the calculation is to reclassify the question-answer pairs and mark them as "high priority set" and "low priority collection" , while generating a priority index For subsequent recommendation strategy adjustments;

[0083] According to the significant pattern changes in user feedback, the parameters of the personalized recommendation algorithm are dynamically optimized. The evolution of user interests over time is modeled through the time decay model, and the formula is as follows:

[0084] ;

[0085] in, is the updated user interest feature vector, is the current interest feature, is the new interest feature in the feedback feature, is the time attenuation coefficient. Combined with the user behavior frequency and satisfaction , dynamically adjust the parameter set of the recommendation algorithm :

[0086] ;

[0087] in, is the learning rate It is an objective optimization function based on feedback data and question-answer database to minimize the recommendation error;

[0088] During the adjustment process, we further optimized topic matching and content diversity to ensure the accuracy and coverage of recommended content, including dynamic update of topic adaptation distribution and diversity indicators, and constructed content coverage function. and diversity optimization function , respectively calculate the recommended content in the topic coverage and similarity constraints The following distribution:

[0089] ,

[0090] Combination and , conduct secondary screening of recommended content;

[0091] Based on the adjusted recommendation strategy and a set of high-priority question-answer pairs , generate the final personalized answer recommendation list To ensure the dynamic adaptability of recommended content, the final ranking weight is constructed by combining user historical behavior data:

[0092] ;

[0093] in, It is the parameter for assigning sort weights. Generate recommendation results in descending order and integrate the final recommendation list Dynamic Q&A Database Priority index , the recommendation results are returned to users through an incremental recommendation mechanism, and a real-time feedback mechanism is supported to ensure the interactivity and adaptability of the system.

[0094] On the other hand, the present application provides a context-aware dynamic question-answer matching system, comprising:

[0095] Data collection module, which collects multi-round conversation history, user background information, user real-time feedback data and behavior analysis data in real time;

[0096] The data processing module cleans and standardizes the collected data.

[0097] The feature extraction and selection module further extracts features that can effectively describe user needs and context information based on the cleaned and standardized data obtained from the data processing module.

[0098] The question-answer matching and recommendation module matches answers from the question-answer library based on the semantic understanding model and the current user input, and dynamically adjusts the matching strategy; based on user feedback and behavior data, it updates the personalized recommendation algorithm and generates a personalized answer recommendation list;

[0099] The personalized recommendation update module updates the personalized recommendation algorithm by integrating user feedback data, historical behavior and the output results of the question-answer matching module.

[0100] The system monitoring and adaptive optimization module is responsible for continuously monitoring the operating status of the system.

[0101] The user interaction and visualization module displays the system's question-and-answer results and recommended content through an intuitive user interface, and supports multi-scenario simulation and interactive operations.

[0102] The key innovative features of the present invention include:

[0103] (1) Context-aware model: By introducing a context-aware mechanism and combining multi-round dialogue history and user background information, the question-answer matching strategy is dynamically adjusted to improve the accuracy of semantic understanding and the relevance of question-answer matching.

[0104] (2) Dynamic Q&A database updates: The Q&A database is automatically updated based on user feedback and new data to ensure that the content of the Q&A database is always the latest and most relevant, solving the problem that the content of traditional Q&A databases is static and not adaptable to changes.

[0105] (3) Personalized recommendation algorithm: Based on the user's interests and historical behavior, the personalized recommendation algorithm can respond to changes in user needs in real time, provide more accurate answer recommendations, and significantly improve the user's interactive experience.

[0106] The present invention achieves a high degree of personalization and accuracy in the question-and-answer system by introducing a dynamic question-and-answer matching method based on context perception. By combining multi-round conversation history and user background information, the present invention can effectively understand the user's intentions and needs, automatically adjust the question-and-answer matching strategy, eliminate the semantic bias in the traditional static question-and-answer system, and thus improve the matching accuracy. The system uses real-time feedback data and behavioral analysis data to dynamically update the recommendation algorithm and question-and-answer library content to ensure the timeliness and relevance of the questions and answers. This dynamic update process avoids the problem of outdated content in the traditional question-and-answer library, improves the flexibility and adaptability of the system, and ensures that efficient and accurate answer recommendations can still be provided in a rapidly changing environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0107] Figure 1 A flowchart of a context-aware dynamic question-answer matching method provided in an embodiment of the present application;

[0108] Figure 2 A structural block diagram of a context-aware dynamic question-answer matching system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0109] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit the present application; the terms "including" and "having" and any variations thereof in the specification and claims of the present application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of the present application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0110] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0111] Example 1: Reference Figure 1 , is a flow chart of a context-aware dynamic question-answer matching method provided by an embodiment of the present invention. The flow chart may at least include steps S100-S500:

[0112] S100, obtaining multi-round conversation history and user background information, performing context-aware analysis, and generating a semantic understanding model for the current conversation;

[0113] S200, matching answers from the question and answer library based on the semantic understanding model and current user input, and dynamically adjusting the matching strategy to generate preliminary answer candidates;

[0114] S300, obtaining real-time feedback data and behavior analysis data from users, updating personalized recommendation algorithms based on changes in user interests and needs, and generating personalized answer recommendation lists;

[0115] S400, dynamically updating the content of the question and answer library according to the personalized recommendation list and the user behavior data, and generating an updated question and answer library;

[0116] S500: Input the user feedback and the updated question and answer database data into the personalized recommendation system, readjust the recommendation strategy, and generate the final personalized answer recommendation.

[0117] Step S100 at least includes steps S110-S130:

[0118] S110: Obtain multiple rounds of conversation history.

[0119] Extract the most recent Round conversation record, record is represented as a collection ,in The i-th round of dialogue content, including user input text and system-generated answers. The data field includes the dialogue text , Timestamp , and user behavior annotation (e.g. "satisfied" or "unresolved").

[0120] Record of the conversation Perform text cleaning, including removing special characters, correcting spelling errors, and unifying the language style.

[0121] For the preprocessed text sequence, word segmentation technology is used to decompose it into a set of vocabulary units. ,in Indicates the number of word segments in this round of dialogue.

[0122] The word segmentation result Convert to embedding vector (Use a pre-trained embedding model such as or .

[0123] Finally, a multi-round dialogue embedding matrix is ​​formed , as the input of the context-aware analysis module.

[0124] S120: Obtain user background information.

[0125] Obtain user static information from the user profile database, including interest preferences , Language style , Occupational category Etc., constitute a static feature vector set .

[0126] Specifically, a rule matching method is used to transform unstructured text information into numerical features.

[0127] Extract user dynamic features from real-time interaction logs, including current device type , Geographical location , and conversation sentiment distribution , generating dynamic feature vectors .

[0128] Concatenate the dynamic feature vector with the static feature vector to form a comprehensive user background feature vector .

[0129] The comprehensive feature vector For normalization, use the standardization formula:

[0130] ;

[0131] in, is the feature mean, is the standard deviation, and the normalized features are obtained .

[0132] S130: Perform context-aware analysis.

[0133] Embedding multi-turn dialogue into a matrix and user background feature vector Input context-aware model , the model input is:

[0134] ;

[0135] in, Represents the joint input data for contextual analysis.

[0136] based on The model performs semantic decomposition and generates semantic vectors ,in For the Dimensional semantic features.

[0137] Furthermore, the context weight distribution is calculated , calculated using the attention mechanism:

[0138] ;

[0139] in, Indicates The weight of the turn, Represents a dot product operation.

[0140] Semantic vector and weight Perform weighted aggregation to generate the semantic understanding vector of the current conversation:

[0141] ;

[0142] Finally, we get the semantic understanding model of the current conversation , used in the subsequent question-answer matching step.

[0143] Step S200 at least includes steps S210-S230:

[0144] S210: Calculate the correlation between the current input semantic vector and the semantic understanding model.

[0145] Get the current input text from the user , for the Perform text cleaning, including removing noise characters and unifying grammatical rules.

[0146] Using the same semantic embedding model (e.g. )right Encoding, generating input semantic vector , specifically:

[0147] ;

[0148] in represents the semantic embedding function.

[0149] Combined with the semantic understanding model generated by S130 ,calculate and The correlation , using the cosine similarity formula:

[0150] ;

[0151] in, is the relevance score of semantic matching, represents the dot product, Represents the L2 norm of a vector.

[0152] The correlation Map to The specific normalization formula is:

[0153] ;

[0154] in, and are the maximum and minimum correlations in the matching process respectively.

[0155] S220: Match candidate answers in the question and answer database.

[0156] Extract all question texts from the Q&A database , forming a set of problems ,in is the total number of questions in the question-answering database. The same embedding model is used as for S130 and S210 Encode the problem set Q and generate an embedding matrix:

[0157] ;

[0158] in, .

[0159] For the user's current input semantic vector , calculate them one by one with The similarity of all question embedding vectors in , forming a similarity score vector , the specific calculation formula is:

[0160] ;

[0161] in, For user input and questions The similarity.

[0162] Based on similarity score , select before The questions with the highest similarity constitute the initial candidate question set , the corresponding answer set is

[0163] and It will be used as input to the dynamic adjustment strategy module.

[0164] S230: Dynamically adjust the matching strategy.

[0165] Combined with the context weight generated by S130 , adjust the candidate answer set Specifically, for each candidate answer, the matching score , recalculate the weighted score :

[0166] ;

[0167] in, For the The semantic contribution value of a turn-based conversation is generated based on the semantic understanding model.

[0168] Based on weighted score For the candidate answer set Re-sort and generate the final sorted answer candidate set .

[0169] Furthermore, filtering below the set threshold answers to ensure high relevance of preliminary candidate answers.

[0170] The final sorted set of candidate answers Output, providing input for the subsequent answer verification module.

[0171] Step S300 at least includes steps S310-S330:

[0172] S310: Collect user real-time feedback data and behavior analysis data.

[0173] The user behavior sequence is obtained by collecting data in real time through the user interface, including the user's click behavior, input revision data, dwell time, mouse trajectory, and option selection records in the current question-answering system interaction. .

[0174] in: Indicates Data points of each interaction behavior, which may include timestamp, interaction type, and feedback content; Indicates the time period of the current session.

[0175] Furthermore, feature vectors are extracted from the above user behavior sequences. , including interest preference indicators , interaction frequency index and historical satisfaction ratings These features serve as the basic input for subsequent analysis and are passed to S320 for dynamic evaluation of interest changes.

[0176] S320: Evaluate changes in user interests and needs.

[0177] Feature vector extracted based on S310 , evaluate the model using context variation , calculate the user interest change coefficient and the demand offset parameter The calculation formula is as follows:

[0178] ;

[0179] ;

[0180] in: is the interest preference index of the previous round; Rate current historical satisfaction; Representation characteristics The weight of is dynamically adjusted by the feature importance analysis module.

[0181] Furthermore, by integrating , generate a user demand change model and transmits it to S330 for updating the personalized recommendation algorithm.

[0182] S330: Update the personalized recommendation algorithm and generate a recommendation list.

[0183] Based on the user demand change model , dynamically adjust the weight matrix of the recommendation algorithm and priority strategy, retraining the personalized recommendation model The specific steps include:

[0184] according to Demand offset parameter in , update the feature weights of the recommendation model: ,in is the learning rate, which controls the amplitude of weight update.

[0185] Use the latest recommendation algorithm to sort the candidate answers in the question and answer database and generate a personalized recommendation list .

[0186] in, represents the i-th candidate answer, arranged in descending order of relevance; The number of recommended answers is controlled by the threshold set by the system.

[0187] Will Passed to subsequent modules for further optimization and output.

[0188] Step S400 at least includes steps S410-S430:

[0189] S410: Obtain personalized recommendation lists and user behavior data, and evaluate the timeliness and relevance of question and answer pairs.

[0190] Combined with personalized recommendation list and user behavior data , for each question-answer pair in the question-answer database First, from the recommended list Extract key features , which represents the user’s current topic interests. Extract interaction frequency and user satisfaction ratings Distribution characteristics of . Construct a comprehensive evaluation function:

[0191] ;

[0192] in, For question and answer The timestamp of its generation, Score historical matches, is the current timestamp, and are the global distributions of user satisfaction and interaction frequency, is the weight parameter, Represents the normalized functions of time correlation, topic correlation, matching score correlation, satisfaction correlation and interaction frequency correlation respectively. , construct the filtering rules, which will satisfy:

[0193] ;

[0194] A collection of question-answer pairs As the preliminary result of high-relevance question-answer pairs, question-answer pairs below this threshold are classified into the low-relevance set. .

[0195] S420: Based on the question-answer pair evaluation results, identify question-answer pairs with low relevance or low timeliness, and dynamically update the question-answer library content.

[0196] Furthermore, for the collection The question-answer pairs in the and behavioral data Conduct in-depth analysis to focus on identifying the root causes of low timeliness or low relevance. For low timeliness question-answer pairs, use the formula:

[0197] ;

[0198] in, is the time attribute update function, Represents a dynamic correction parameter to correct the utility of generating timestamps. , combined with the theme correction function and content optimization functions , perform the following update:

[0199] .

[0200] If the relevance evaluation value of the updated question-answer pair If the value is still below the minimum threshold, the question-answer pair will be permanently removed from the question-answer database. The question-answer pairs in the database are further optimized through feedback data to further optimize their weight parameters and topic matching to ensure the content accuracy of the question-answer database.

[0201] S430: Dynamically integrate the newly generated question-answer pairs with the updated question-answer library to form a final version of the question-answer library.

[0202] The new question-answer pairs generated from the recommender system and the updated set of question and answer pairs Dynamic integration.

[0203] Using semantic similarity function ,calculate and Content similarity between question and answer pairs , merge high similarity question and answer pairs to avoid content duplication. The formula is:

[0204] ;

[0205] in, Represents the semantic similarity function. Reassign priority weights to the question-answer pairs in the integrated question-answer library , the weight calculation formula is:

[0206] ;

[0207] in, represents the correlation evaluation value, is the feedback weight adjustment function.

[0208] Build a new version of the Q&A library , and optimize the index structure to ensure that the retrieval efficiency meets the system requirements. Dynamic availability can be maintained through an incremental update mechanism.

[0209] Step S500 at least includes steps S510-S530:

[0210] S510: Obtain user feedback and updated question and answer database data.

[0211] The user's feedback data in the previous round of interaction and updated Q&A library Input to the personalized recommendation system. User feedback data Including satisfaction , interaction frequency , and timing characteristics , used to evaluate users’ current acceptance of recommended content and changes in demand.

[0212] Through the feedback enhancement mechanism, the recommendation weight of each question-answer pair is re-evaluated based on the following formula:

[0213] ;

[0214] in, is the relevance score of the question-answer pair, represents the feedback enhancement function based on user satisfaction, represents the feedback enhancement function based on the interaction frequency, is the weight parameter, U and F are the distribution data of global user satisfaction and interaction frequency.

[0215] Combination The result of the calculation is to reclassify the question-answer pairs and mark them as "high priority set" and "low priority collection" , while generating a priority index For use in subsequent recommendation strategy adjustments.

[0216] S520: Dynamically adjust the recommendation strategy based on the feedback data.

[0217] According to the significant pattern changes in user feedback, the parameters of the personalized recommendation algorithm are dynamically optimized. The evolution of user interests over time is modeled through the time decay model, and the formula is as follows:

[0218] ;

[0219] in, is the updated user interest feature vector, is the current interest feature, is the new interest feature in the feedback feature, is the time attenuation coefficient. Combined with the user behavior frequency and satisfaction , dynamically adjust the parameter set of the recommendation algorithm :

[0220] ;

[0221] in, is the learning rate It is an objective optimization function based on feedback data and question-answer database, used to minimize the recommendation error.

[0222] During the adjustment process, we further optimize topic matching and content diversity to ensure the accuracy and coverage of recommended content, including dynamic update of topic adaptation distribution and diversity indicators. and diversity optimization function , respectively calculate the recommended content in the topic coverage and similarity constraints The following distribution:

[0223] ;

[0224] Combination and , conduct secondary screening of recommended content.

[0225] S530: Generate a final personalized answer recommendation.

[0226] Based on the adjusted recommendation strategy and a set of high-priority question-answer pairs , generate the final personalized answer recommendation list To ensure the dynamic adaptability of recommended content, the final ranking weight is constructed by combining user historical behavior data:

[0227] ;

[0228] in, It is the parameter for assigning sort weights. Generate recommendation results by sorting in descending order.

[0229] Put together a final list of recommendations Dynamic Q&A Database Priority index , the recommendation results are returned to users through an incremental recommendation mechanism, and a real-time feedback mechanism is supported to ensure the interactivity and adaptability of the system.

[0230] The key innovative features of the present invention include:

[0231] (1) Context-aware model: By introducing a context-aware mechanism and combining multi-round dialogue history and user background information, the question-answer matching strategy is dynamically adjusted to improve the accuracy of semantic understanding and the relevance of question-answer matching.

[0232] (2) Dynamic Q&A database updates: The Q&A database is automatically updated based on user feedback and new data to ensure that the content of the Q&A database is always the latest and most relevant, solving the problem that the content of traditional Q&A databases is static and not adaptable to changes.

[0233] (3) Personalized recommendation algorithm: Based on the user's interests and historical behavior, the personalized recommendation algorithm can respond to changes in user needs in real time, provide more accurate answer recommendations, and significantly improve the user's interactive experience.

[0234] The present invention achieves a high degree of personalization and accuracy in the question-and-answer system by introducing a dynamic question-and-answer matching method based on context perception. By combining multi-round conversation history and user background information, the present invention can effectively understand the user's intentions and needs, automatically adjust the question-and-answer matching strategy, eliminate the semantic bias in the traditional static question-and-answer system, and thus improve the matching accuracy. The system uses real-time feedback data and behavioral analysis data to dynamically update the recommendation algorithm and question-and-answer library content to ensure the timeliness and relevance of the questions and answers. This dynamic update process avoids the problem of outdated content in the traditional question-and-answer library, improves the flexibility and adaptability of the system, and ensures that efficient and accurate answer recommendations can still be provided in a rapidly changing environment.

[0235] Embodiment 2: Figure 2 FIG. 4 is a block diagram showing a dynamic question-answer matching system based on context awareness according to an embodiment of the present invention. Figure 2 As shown, the system may include:

[0236] The data collection module 10 obtains information related to user interaction from multiple data sources in real time. Specifically, the module 10 extracts key data from multi-dimensional information such as user input data, user behavior data, system logs, and real-time feedback data. This module continuously collects interaction data such as specific questions, click behaviors, and dwell time input by users through efficient sensors or interfaces. It also collects feedback data from users during the use of the system, such as user evaluation of recommended answers, operation paths, and satisfaction with responses. Through these data, the module 10 can provide the system with comprehensive contextual information and provide necessary data support for subsequent data processing and question-answer matching.

[0237] The data processing module 20 cleans and standardizes the collected data. Module 20 first filters the raw data for noise, removes invalid information and abnormal data, and ensures the validity and consistency of the data; then, it supplements the missing key data through the missing value filling algorithm to improve the integrity of the data; finally, it unifies the data format through the data standardization method to ensure that data from different sources can be analyzed and processed under the same standard. After these processes, the data will be stored in an efficient database to ensure that it can be quickly accessed by subsequent modules and support real-time queries. The output of the data processing module 20 is high-quality, structured user behavior data and interaction information, which provides an accurate data foundation for the question-answer matching process.

[0238] The feature extraction and selection module 30 further extracts features that can effectively describe user needs and context information based on the cleaned and standardized data obtained from the data processing module 20. Specifically, the module 30 extracts feature vectors from the user's input content (such as the keywords and context of the question) and user behavior data (such as click behavior, search history, etc.). These features can reflect the user's interests, needs, and preferences for answers. At the same time, the module 30 will also screen and optimize the extracted features according to the machine learning algorithm, eliminate redundant features, and retain the most predictive and relevant features, providing high-quality input data for subsequent model training.

[0239] The question-answer matching and recommendation module 40 is one of the core modules of the entire system. It is responsible for semantic analysis of the input based on the current question input by the user and the feature vector obtained from the feature extraction module 30, using the pre-trained semantic understanding model, and matching the most relevant answer from the question-answer library. This module sorts the candidate answers in the question-answer library by calculating the semantic similarity between the question and the answer, combined with a context-aware dynamic matching strategy. Specifically, module 40 uses a deep learning model for semantic matching, and dynamically adjusts the matching strategy based on the results of user behavior analysis to ensure that the recommended answers can accurately meet the user's needs. At the same time, module 40 further adjusts the recommendation strategy based on the user's personalized preferences and historical behavior, thereby generating a personalized answer recommendation list.

[0240] The personalized recommendation update module 50 continuously updates the personalized recommendation algorithm by integrating user feedback data, historical behavior and the output results of the question-answer matching module 40. In this process, the module 50 analyzes the user's real-time feedback (such as the user's click-through rate and satisfaction score for the recommended answer) and behavioral data (such as click path, browsing time, etc.), identifies changes in user interests and needs, and dynamically adjusts the recommendation strategy. These updates will affect the subsequent recommended content and ranking, allowing the system to adapt to changes in user interests in real time. Specifically, the module 50 will adjust the weight of the recommendation model based on the feedback data, so that the recommendation algorithm is more accurate in the personalized recommendation process, thereby further improving the user's interactive experience.

[0241] The system monitoring and adaptive optimization module 60 is responsible for continuously monitoring the operating status of the system to ensure the stability and efficiency of the system in various environments. This module analyzes the system performance based on real-time data inflow, system operating load, and user interaction data, automatically adjusts computing resources and optimizes algorithm parameters, thereby ensuring that the system can maintain an efficient response speed when processing large amounts of data. In addition, module 60 will also perform adaptive learning based on system operating data and optimize model parameters, so that the entire recommendation process can always be kept in the optimal state.

[0242] The user interaction and visualization module 70 displays the system's question-answering results and recommended content through an intuitive user interface, supporting multi-scenario simulation and interactive operations. Module 70 not only provides a feedback channel for user input, but also allows users to understand the results of the system's recommendations and the logic behind the data in a visual form. Through this module, users can easily view recommended answers and further filter them according to their personal needs, greatly improving the user experience.

[0243] Beneficial results of the embodiment:

[0244] (1) Improving the accuracy of question-answer matching: Through real-time feedback and context-aware analysis, the system can dynamically adjust the question-answer matching strategy based on the user’s input and historical conversation information. This mechanism can effectively eliminate the semantic bias in traditional static question-answer matching methods, improve the relevance and accuracy of questions and answers, and ensure that users receive more accurate and timely answers.

[0245] (2) Enhanced user interaction experience: The system uses a personalized recommendation algorithm to automatically generate a personalized answer recommendation list based on the user's interests, historical behavior, and real-time feedback. Compared with traditional methods, this algorithm can respond to changes in user needs in real time and provide users with tailored answers, thereby significantly improving user satisfaction, interaction effects, and overall interaction experience.

[0246] (3) Ensure the timeliness of the Q&A database: The system can timely supplement and modify the information in the Q&A database based on user feedback by updating and dynamically adjusting the content of the Q&A database in real time. Compared with the traditional static Q&A database, the present invention can ensure that the content of the Q&A database is always up to date and most relevant, solving the problem that the original Q&A database is difficult to cope with rapidly changing information needs, and significantly improving the adaptability and flexibility of the system.

[0247] (4) Improving the adaptability of the system: The present invention can flexibly adjust the question-answering recommendation strategy by combining contextual information with real-time feedback. Whether in a complex and ever-changing dialogue environment or facing different user needs, the system can efficiently optimize and adjust the strategy, thereby ensuring the provision of stable and efficient question-answering services in a variety of application scenarios, and enhancing the system's adaptability and resilience.

[0248] Obviously, the embodiments described above are only some embodiments of the present application, rather than all embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application is described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions recorded in the aforementioned specific implementation methods, or to perform equivalent replacement of some of the technical features therein. Any equivalent structure made using the contents of the specification and drawings of this application, directly or indirectly used in other related technical fields, is similarly within the scope of patent protection of this application.

Claims

1. A context-aware dynamic question-answer matching method, characterized in that: The following steps are involved: Obtain multi-round conversation history and user background information, perform context-aware analysis, and generate a semantic understanding model for the current conversation; Based on the semantic understanding model and the current user input, the most relevant answer is matched from the question and answer library, and the matching strategy is dynamically adjusted to generate preliminary answer candidates; Obtain users' real-time feedback data and behavior analysis data, update personalized recommendation algorithms based on changes in user interests and needs, and generate personalized recommendation lists; Dynamically update the question and answer library content according to the personalized recommendation list and user behavior data to generate an updated question and answer library; Input user feedback and updated question-answer database data into the personalized recommendation system, readjust the recommendation strategy, and generate the final personalized answer recommendation; The process of acquiring multiple rounds of conversation history and user background information, performing context-aware analysis, and generating a semantic understanding model of the current conversation includes: Extract the most recent Round of conversation records, where the conversation records are represented as a collection ,in is the content of the i-th round of dialogue, i∈[1,n], including the user input text and the system-generated answer; the data field of the dialogue content includes the dialogue text , Timestamp , and user behavior annotation ; Record of the conversation Perform text cleaning, including removing special characters, correcting spelling errors, and unifying language style; For the text sequence after text cleaning, word segmentation technology is used to decompose it into a set of vocabulary units. ,in ; Gather vocabulary units Convert to embedding vector , thus forming a multi-round dialogue embedding matrix , as the input of the context-aware analysis module; Obtain user static information from the user profile database, including interest preferences , Language style , Occupational category , forming a static feature vector set ; Extract user dynamic features from real-time interaction logs, including current device type , Geographical location , and conversation sentiment distribution , generating dynamic feature vectors ; Concatenate the dynamic feature vector with the static feature vector to form a comprehensive user background feature vector ; Comprehensive user background feature vector For normalization, use the standardization formula: ； in, is the feature mean, is the standard deviation, and the normalized user background feature vector is obtained ; Embedding multi-turn dialogue into a matrix and user background feature vector Input context-aware model , the model input is: ； in, represents the joint input data for contextual analysis; based on The model performs semantic decomposition and generates semantic vectors ,in For the Dimensional semantic features; Furthermore, the context weight distribution is calculated , calculated using the attention mechanism: ； in, Indicates The weight of the turn, Represents the dot product operation, i∈[1,n]; Semantic vector and weight Perform weighted aggregation to generate the semantic understanding vector of the current conversation: ； Finally, we get the semantic understanding model of the current conversation , used in the subsequent question-answer matching step.

2. The context-aware dynamic question-answer matching method according to claim 1, characterized in that: Based on the semantic understanding model and the current user input, the most relevant answer is matched from the question and answer library, and the matching strategy is dynamically adjusted to generate preliminary answer candidates, including: Get the current input text from the user ,right Perform text cleaning, including removing noise characters and unifying grammatical rules; Using the same semantic embedding model Encoding, generating input semantic vector , specifically: ;in represents the semantic embedding function; Generation-based semantic understanding model ,calculate and The correlation , using the cosine similarity formula: ； in, is the relevance score of semantic matching, represents the dot product, Represents the L2 norm of the vector; The correlation Map to The specific normalization formula is: ； in, and are the maximum and minimum correlations in the matching process, is the normalized correlation; Extract all question texts from the Q&A database , forming a set of problems ,in For the mth question in the question-answering database; use the embedding model For the problem set Encode and generate an embedding matrix: ;in, ; For the user's current input semantic vector , calculate them one by one with The similarity of all question embedding vectors in , forming a similarity score vector , the specific calculation formula is: ； in, For user input and questions similarity; Based on similarity score , select before The questions with the highest similarity constitute the initial candidate question set , the corresponding answer set is ; and It will be used as input to the dynamic adjustment strategy module; Combine the generated context weights , adjust the candidate answer set The priority, specifically, the matching score for each candidate answer , recalculate the weighted score : ； in, For the The semantic contribution value of the turn-based dialogue is generated based on the semantic understanding model; Based on weighted score For the candidate answer set Re-sort and generate the final sorted answer candidate set ; Filter below the set threshold To ensure the high relevance of the preliminary candidate answers, the final ranked candidate answer set Output, providing input for the subsequent answer verification module.

3. The context-aware dynamic question-answer matching method according to claim 2, characterized in that: Obtain real-time feedback data and behavior analysis data from users, update personalized recommendation algorithms based on changes in user interests and needs, and generate personalized recommendation lists, including: The user behavior sequence is obtained by real-time acquisition of the user's click behavior, input revision data, dwell time, mouse trajectory, and option selection record data in the current question-answering system interaction. ; in: Indicates The data points of interaction behaviors, i∈[1,n], include timestamp, interaction type and feedback content; Indicates the time period of the current session; Extract feature vectors from the above user behavior sequence , including interest preference indicators , interaction frequency index and historical satisfaction ratings ;These features serve as the basic input for subsequent analysis; Based on the extracted feature vector , evaluate the model using context variation , calculate the user interest change coefficient and the demand offset parameter , the calculation formula is as follows: ； ； in: is the interest preference index of the previous round; Rate current historical satisfaction; Representation characteristics The weight of is dynamically adjusted by the feature importance analysis module; Integration , generate a user demand change model , used to update the personalized recommendation algorithm; Based on the user demand change model , dynamically adjust the weight matrix of the recommendation algorithm and priority strategy, retraining the personalized recommendation model , the specific steps include: according to Demand offset parameter in , update the feature weights of the recommendation model: ,in is the learning rate, which controls the weight update amplitude; Use the latest recommendation algorithm to sort the candidate answers in the question and answer database and generate a personalized recommendation list ; in, represents the kth candidate answer, arranged in descending order of relevance; is the number of recommended answers, which is controlled by the threshold set by the system. Passed to subsequent modules for further optimization and output.

4. The context-aware dynamic question-answer matching method according to claim 3, characterized in that: Dynamically updating the Q&A database content based on the personalized recommendation list and user behavior data to generate an updated Q&A database, including: Obtain personalized recommendation lists and user behavior data to evaluate the timeliness and relevance of question-answer pairs. This is achieved by combining personalized recommendation lists with user behavior data. and user behavior data , for each question-answer pair in the question-answer database Comprehensively evaluate the timeliness and relevance of Extract key feature sets , which represents the user's current topic interests; at the same time, from the behavioral data Extract interaction frequency from and user satisfaction ratings Distribution characteristics; construct a comprehensive evaluation function: ； in, For question and answer The timestamp of its generation, Score historical matches, is the current timestamp, and are the global distributions of user satisfaction and interaction frequency, is the weight parameter, Respectively represent the normalized functions of time correlation, topic correlation, matching score correlation, satisfaction correlation and interaction frequency correlation; based on , construct the filtering rules, which will satisfy: A collection of question-answer pairs As the preliminary result of high-relevance question-answer pairs, question-answer pairs below this threshold are classified into the low-relevance set. ; Based on the evaluation results of the question-answer pairs, identify the question-answer pairs with low relevance or low timeliness, and dynamically update the content of the question-answer library. The implementation method is as follows: for the collection The question-answer pairs in the and behavioral data Conduct in-depth analysis to focus on identifying the root causes of low timeliness or low relevance. For low timeliness question-answer pairs, use the formula: ;in, is the time attribute update function, Represents a dynamic correction parameter to correct the utility of generating timestamps; for low-correlation question-answer pairs , combined with the theme correction function and content optimization functions , perform the following update: ； If the relevance evaluation value of the updated question-answer pair If the value is still below the minimum threshold, the question-answer pair will be permanently removed from the question-answer database. The weight parameters and topic matching of the question-answer pairs in the database are further optimized through feedback data to ensure the content accuracy of the question-answer database. Dynamically integrate the newly generated question-answer pairs with the updated question-answer library to form the final version of the question-answer library. The implementation method is: and the updated set of question and answer pairs Conduct dynamic integration; Using semantic similarity function ,calculate and Content similarity between question and answer pairs , merge high similarity question and answer pairs to avoid content duplication. The formula is: ； in, Represents the semantic similarity function; redistributes priority weights to the question-answer pairs in the integrated question-answer library , the weight calculation formula is: in, represents the correlation evaluation value, is the feedback weight adjustment function; Build a new version of the Q&A library , and optimize the index structure to ensure that the retrieval efficiency meets the system requirements; the updated question and answer library Dynamic availability can be maintained through an incremental update mechanism.

5. The context-aware dynamic question-answer matching method according to claim 4, characterized in that: Input user feedback and updated question-answer database data into the personalized recommendation system, readjust the recommendation strategy, and generate the final personalized answer recommendation, including: The user's feedback data in the previous round of interaction and updated Q&A library Input into the personalized recommendation system, user feedback data Including satisfaction , interaction frequency , and timing characteristics , used to evaluate users’ current acceptance of recommended content and changes in demand; Through the feedback enhancement mechanism, the recommendation weight of each question-answer pair is re-evaluated based on the following formula: ； in, is the relevance score of the question-answer pair, represents the feedback enhancement function based on user satisfaction, represents the feedback enhancement function based on the interaction frequency is the weight parameter, U and F are the distribution data of global user satisfaction and interaction frequency; Combination The result of the calculation is to reclassify the question-answer pairs and mark them as "high priority set" and "low priority collection" , while generating a priority index For subsequent recommendation strategy adjustments; According to the significant pattern changes in user feedback, the parameters of the personalized recommendation algorithm are dynamically optimized; the evolution of user interests over time is modeled through the time decay model, and the formula is as follows: ； in, is the updated user interest feature vector, is the current interest feature, is the new interest feature in the feedback feature, is the time attenuation coefficient; combined with the user behavior frequency and satisfaction , dynamically adjust the parameter set of the recommendation algorithm , and then get the adjusted parameter set : ； in, is the learning rate; It is an objective optimization function based on feedback data and question-answer database to minimize the recommendation error; is the loss function with respect to the parameter set The gradient of For users Real-time feedback data, including click behavior and satisfaction ratings; For the updated Q&A database; During the adjustment process, we further optimized topic matching and content diversity to ensure the accuracy and coverage of recommended content, including dynamic update of topic adaptation distribution and diversity indicators, and constructed content coverage function. and diversity optimization function , respectively calculate the recommended content in the topic coverage and similarity constraints The following distribution: ； Combination and , conduct secondary screening of recommended content; Based on the adjusted recommendation strategy and a set of high-priority question-answer pairs , generate the final personalized answer recommendation list ; To ensure the dynamic adaptability of recommended content, the final ranking weight is constructed by combining user historical behavior data: ； in, is the parameter for allocating sorting weights; Generate recommendation results in descending order and integrate the final recommendation list Dynamic Q&A Database Priority index , the recommendation results are returned to users through an incremental recommendation mechanism, and a real-time feedback mechanism is supported to ensure the interactivity and adaptability of the system.

6. A system adapted to the context-aware dynamic question-answer matching method according to any one of claims 1 to 5, characterized in that: include: Data collection module, which collects multi-round conversation history, user background information, user real-time feedback data and behavior analysis data in real time; Data processing module, which cleans and standardizes the collected data; The feature extraction and selection module further extracts features that can effectively describe user needs and context information based on the cleaned and standardized data obtained from the data processing module; The question-answer matching and recommendation module matches answers from the question-answer library based on the semantic understanding model and the current user input, and dynamically adjusts the matching strategy; based on user feedback and behavior data, it updates the personalized recommendation algorithm and generates a personalized answer recommendation list; The personalized recommendation update module updates the personalized recommendation algorithm by integrating user feedback data, historical behavior, and the output results of the question-answer matching module; System monitoring and adaptive optimization module, responsible for continuously monitoring the operating status of the system; The user interaction and visualization module displays the system's question-and-answer results and recommended content through an intuitive user interface, and supports multi-scenario simulation and interactive operations.

Images