Multi-agent collaboratively-driven power marketing question and answer method

Abstract

The invention relates to a multi-agent collaboratively-driven power marketing question and answer method and device, computer equipment, a computer readable storage medium and a computer program product, and can be applied to the technical field of artificial intelligence. The method comprises the following steps: constructing a plurality of problem sequences according to historical interaction information; performing keyword extraction on the plurality of question sequences to obtain a first keyword set and a second keyword set; labeling the first keyword set and the second keyword set to obtain a problem data set; determining a target problem sequence; constructing a portrait vector, and constructing a category labeling sequence; inputting the portrait vector and the category labeling sequence into a knowledge domain supplementary category prediction model for processing to obtain knowledge domain supplementary category information; distributing the question task of the user to the intelligent agent of the corresponding knowledge field for processing, and obtaining a plurality of pieces of reply information; and carrying out fusion processing on the multiple pieces of reply information to obtain a query result of the user. The question and answer accuracy can be improved by adopting the method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a multi-agent collaborative driving power marketing question and answer method and device, computer equipment, computer readable storage medium and computer program product. BACKGROUND

[0002] At present, in the field of power marketing services, in order to improve service quality and operational efficiency, how to accurately understand and respond to user consultation needs is crucial.

[0003] In the traditional technology, the power marketing question and answer system usually adopts a rule engine-based response; However, this method only answers single-round questions, and when dealing with complex problems involving multiple knowledge fields, it cannot accurately understand the user's true intention, resulting in low accuracy of question and answer. SUMMARY

[0004] Therefore, it is necessary to provide a multi-agent collaborative driving power marketing question and answer method, device, computer equipment, computer readable storage medium and computer program product capable of improving the accuracy of question and answer.

[0005] In a first aspect, the present application provides a multi-agent collaborative driving power marketing question and answer method. The method comprises:

[0006] Obtaining historical interaction information and account attribute information of a user, and constructing a plurality of question sequences according to the historical interaction information; The plurality of question sequences include user questions and system replies;

[0007] Extracting keywords from the plurality of question sequences to obtain a first keyword set in the user questions and a second keyword set in the system replies;

[0008] According to the matching results between the first keyword set and the second keyword set and the knowledge base, the first keyword set and the second keyword set are subjected to knowledge domain category annotation processing to obtain a question data set;

[0009] According to the knowledge domain category annotation in the question data set, a target question sequence is determined from the plurality of question sequences;

[0010] According to the account attribute information, an image vector of the user is constructed, and according to the knowledge domain category annotation in the target question sequence, a category annotation sequence is constructed;

[0011] The image vector and the category annotation sequence are input into a knowledge domain supplementary category prediction model for knowledge domain supplementary category prediction processing to obtain knowledge domain supplementary category information;

[0012] According to the knowledge field category annotation in the target question sequence and the knowledge field supplementary category information, the user's question task is allocated to an agent corresponding to the knowledge field for processing to obtain a plurality of reply information;

[0013] The plurality of reply information is fused to obtain the user's query result.

[0014] In one of the embodiments, the knowledge field category annotation processing of the first keyword set and the second keyword set according to the matching result between the first keyword set and the second keyword set and the knowledge base obtains a question data set, which includes:

[0015] The keywords in the first keyword set are matched with the keywords in the knowledge base to obtain a first matching result, and the first keyword set is subjected to knowledge field category annotation processing according to the first matching result to obtain first category annotation information;

[0016] The keywords in the second keyword set are matched with the keywords in the knowledge base to obtain a second matching result, and the second keyword set is subjected to knowledge field category annotation processing according to the second matching result to obtain second category annotation information;

[0017] The question data set is determined according to the first category annotation information and the second category annotation information.

[0018] In one of the embodiments, the target question sequence is determined from the plurality of question sequences according to the knowledge field category annotation in the question data set, which includes:

[0019] The category coincidence degree corresponding to the knowledge field category annotation in the question data set is determined.

[0020] The target question sequence is determined from the plurality of question sequences according to the category coincidence degree.

[0021] In one of the embodiments, the portrait vector of the user is constructed according to the account attribute information, which includes:

[0022] The attribute feature field of the user is extracted from the account attribute information.

[0023] The attribute feature field is subjected to data preprocessing to obtain a target attribute feature field.

[0024] The portrait vector is constructed according to the target attribute feature field.

[0025] In one of the embodiments, the attribute feature field includes user type, region of affiliation, power service type, power consumption capacity, historical power consumption data, and payment method.

[0026] In one of the embodiments, the constructing the category label sequence according to the knowledge field category labels in the target question sequence includes:

[0027] obtaining the knowledge field category label corresponding to each round of question and answer in the target question sequence;

[0028] performing permutation and combination processing on the knowledge field category label corresponding to each round of question and answer in the target question sequence according to the time sequence of the target question sequence, to obtain the category label sequence.

[0029] In a second aspect, the application further provides a multi-agent collaborative driving power marketing question and answer device. The device includes:

[0030] an information acquisition module, configured to acquire historical interaction information and account attribute information of a user, and construct a plurality of question sequences according to the historical interaction information; the plurality of question sequences include user questions and system replies;

[0031] a sequence extraction module, configured to perform keyword extraction on the plurality of question sequences, to obtain a first keyword set in the user questions and a second keyword set in the system replies;

[0032] a label processing module, configured to perform knowledge field category label processing on the first keyword set and the second keyword set according to a matching result between the first keyword set and the second keyword set and a knowledge base, to obtain a question data set;

[0033] a sequence determination module, configured to determine a target question sequence from the plurality of question sequences according to the knowledge field category label in the question data set;

[0034] a vector construction module, configured to construct a portrait vector of the user according to the account attribute information, and construct a category label sequence according to the knowledge field category label in the target question sequence;

[0035] a sequence input module, configured to input the portrait vector and the category label sequence to a knowledge field supplementary category prediction model for knowledge field supplementary category prediction processing, to obtain knowledge field supplementary category information;

[0036] a task allocation module, configured to allocate a question task of the user to an agent in a corresponding knowledge field for processing according to the knowledge field category label in the target question sequence and the knowledge field supplementary category information, to obtain a plurality of reply information;

[0037] An information fusion module is configured to fuse the multiple reply information to obtain the query result of the user.

[0038] In a third aspect, the present application provides a computer device. The computer device comprises a memory and a processor. The memory stores a computer program. When the processor executes the computer program, the following steps are implemented:

[0039] Obtain historical interaction information and account attribute information of a user, and construct multiple question sequences according to the historical interaction information. The multiple question sequences comprise user questions and system replies.

[0040] Extract keywords from the multiple question sequences to obtain a first keyword set in the user questions and a second keyword set in the system replies.

[0041] According to the matching results between the first keyword set, the second keyword set and a knowledge base, perform knowledge domain category labeling processing on the first keyword set and the second keyword set to obtain a question data set.

[0042] According to the knowledge domain category labeling in the question data set, determine a target question sequence from the multiple question sequences.

[0043] Construct a portrait vector of the user according to the account attribute information, and construct a category labeling sequence according to the knowledge domain category labeling in the target question sequence.

[0044] Input the portrait vector and the category labeling sequence into a knowledge domain supplementary category prediction model to perform knowledge domain supplementary category prediction processing, and obtain knowledge domain supplementary category information.

[0045] According to the knowledge domain category labeling in the target question sequence and the knowledge domain supplementary category information, assign a question task of the user to an agent in a corresponding knowledge domain for processing to obtain multiple reply information.

[0046] Fuse the multiple reply information to obtain the query result of the user.

[0047] In a fourth aspect, the present application provides a computer readable storage medium. The computer readable storage medium stores a computer program. When the computer program is executed by a processor, the following steps are implemented:

[0048] Obtain historical interaction information and account attribute information of a user, and construct multiple question sequences according to the historical interaction information. The multiple question sequences comprise user questions and system replies.

[0049] perform keyword extraction on the plurality of question sequences to obtain a first keyword set in the user question and a second keyword set in the system reply;

[0050] perform knowledge domain category annotation processing on the first keyword set and the second keyword set according to a matching result between the first keyword set and the second keyword set and a knowledge base to obtain a question data set;

[0051] determine a target question sequence from the plurality of question sequences according to the knowledge domain category annotation in the question data set;

[0052] construct a portrait vector of the user according to the account attribute information, and construct a category annotation sequence according to the knowledge domain category annotation in the target question sequence;

[0053] input the portrait vector and the category annotation sequence into a knowledge domain supplementary category prediction model to perform knowledge domain supplementary category prediction processing to obtain knowledge domain supplementary category information;

[0054] assign a question task of the user to an agent of a corresponding knowledge domain for processing according to the knowledge domain category annotation in the target question sequence and the knowledge domain supplementary category information to obtain a plurality of reply information;

[0055] perform fusion processing on the plurality of reply information to obtain a query result of the user.

[0056] In a fifth aspect, the present application further provides a computer program product. The computer program product comprises a computer program which, when executed by a processor, implements the following steps:

[0057] obtain historical interaction information and account attribute information of a user, and construct a plurality of question sequences according to the historical interaction information; the plurality of question sequences comprise user questions and system replies;

[0058] perform keyword extraction on the plurality of question sequences to obtain a first keyword set in the user question and a second keyword set in the system reply;

[0059] perform knowledge domain category annotation processing on the first keyword set and the second keyword set according to a matching result between the first keyword set and the second keyword set and a knowledge base to obtain a question data set;

[0060] determine a target question sequence from the plurality of question sequences according to the knowledge domain category annotation in the question data set;

[0061] According to the account attribute information, an image vector of the user is constructed, and a category label sequence is constructed according to a knowledge field category label in the target question sequence;

[0062] The image vector and the category label sequence are input into a knowledge field supplementary category prediction model for knowledge field supplementary category prediction processing, and knowledge field supplementary category information is obtained;

[0063] According to the knowledge field category label in the target question sequence and the knowledge field supplementary category information, the user's question task is assigned to an agent in a corresponding knowledge field for processing, and a plurality of reply information is obtained;

[0064] The plurality of reply information is fused to obtain the user's query result.

[0065] The power marketing question and answer method and device driven by multiple agents, the computer device, the computer readable storage medium, and the computer program product obtain historical interaction information and account attribute information of a user, construct multiple question sequences according to the historical interaction information; the multiple question sequences include user questions and system replies; key words in the multiple question sequences are extracted to obtain a first key word set in the user questions and a second key word set in the system replies; knowledge domain category labeling processing is performed on the first key word set and the second key word set according to a matching result between the first key word set, the second key word set, and a knowledge base to obtain a question data set; a target question sequence is determined from the multiple question sequences according to knowledge domain category labeling in the question data set; an image vector of the user is constructed according to the account attribute information, and a category labeling sequence is constructed according to the knowledge domain category labeling in the target question sequence; the image vector and the category labeling sequence are input into a knowledge domain supplementary category prediction model for knowledge domain supplementary category prediction processing to obtain knowledge domain supplementary category information; a question task of the user is assigned to an agent in a corresponding knowledge domain for processing according to the knowledge domain category labeling in the target question sequence and the knowledge domain supplementary category information to obtain multiple reply information; the multiple reply information is fused to obtain a query result of the user. The scheme constructs multiple question sequences by obtaining historical interaction information of a user, extracts key words in the question sequences, and performs knowledge domain category labeling processing, which is conducive to identifying a knowledge domain that the user focuses on in historical interaction, determining a target question sequence from the multiple question sequences, which is conducive to screening a question domain that the user highly focuses on, constructing an image vector of the user and a category labeling sequence, and inputting them into a knowledge domain supplementary category prediction model for prediction processing, which is conducive to predicting knowledge domain supplementary category information that the user needs but has not explicitly proposed, assigning a question task to an agent in a corresponding knowledge domain for processing according to the knowledge domain category labeling in the target question sequence and the knowledge domain supplementary category information, which is conducive to providing accurate reply information for questions and potential needs proposed by the user, and fusing the multiple reply information, which is conducive to generating a comprehensive query result covering explicit needs and potential needs of the user, thereby improving the accuracy and integrity of question and answer. BRIEF DESCRIPTION OF DRAWINGS

[0066] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the embodiments or the related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort.

[0067] Figure 1 A flowchart of a power marketing question and answer method driven by multiple agents in cooperation in an embodiment;

[0068] Figure 2 A structural diagram of a power marketing question and answer system driven by multiple agents in cooperation in an embodiment;

[0069] Figure 3 A structural block diagram of a power marketing question and answer device driven by multiple agents in cooperation in an embodiment;

[0070] Figure 4 An internal structure diagram of a computer device in an embodiment. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical scheme and advantages of the present application clearer, further detailed description will be made to the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0072] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0073] In an exemplary embodiment, as shown in Figure 1 A power marketing question and answer method driven by multiple agents in cooperation is provided, and the present embodiment takes the method applied to a terminal as an example for illustration. It should be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is realized through interaction between the terminal and the server. The terminal can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, etc.; the server can be a standalone physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. In the present embodiment, the method includes the following steps:

[0074] Step S101, obtaining historical interaction information and account attribute information of a user, and constructing a plurality of question sequences according to the historical interaction information; the plurality of question sequences include user questions and system replies;

[0075] Step S102, performing keyword extraction on the plurality of question sequences to obtain a first keyword set in the user questions and a second keyword set in the system replies;

[0076] Step S103, according to the matching results between the first keyword set and the second keyword set and the knowledge base, the first keyword set and the second keyword set are processed for knowledge field category labeling to obtain a question data set;

[0077] Step S104, according to the knowledge field category labeling in the question data set, a target question sequence is determined from a plurality of question sequences;

[0078] Step S105, according to the account attribute information, a portrait vector of the user is constructed, and according to the knowledge field category labeling in the target question sequence, a category labeling sequence is constructed;

[0079] Step S106, the portrait vector and the category labeling sequence are input into a knowledge field supplementary category prediction model for knowledge field supplementary category prediction processing to obtain knowledge field supplementary category information;

[0080] Step S107, according to the knowledge field category labeling in the target question sequence and the knowledge field supplementary category information, the user's question task is assigned to an agent in the corresponding knowledge field for processing to obtain a plurality of reply information;

[0081] Step S108, the plurality of reply information is fused to obtain the user's query result.

[0082] The historical interaction information can be the interaction record generated by the user in the power marketing Q&A system and the system in multiple rounds of dialogue, including the user's initiated question and the system's reply content.

[0083] The account attribute information can be the user account registration information, including user type, belonging to the region, power service type, power capacity, historical power consumption data, payment method and other key attribute fields.

[0084] The question sequence can be a sequence established based on the historical interaction information, composed of user questions and system replies, taking the first user-initiated question as the starting point and the first system reply content as the ending point to establish question sequence 1, and then establishing question sequences in turn based on each round of user-initiated question and system reply content until the latest system reply content of the user interaction module to establish a complete interaction sequence set.

[0085] The first keyword set can be a class set generated by screening the keywords in the user question after keyword extraction of the question sequence.

[0086] The second keyword set can be a two-class set generated by screening the keywords in the system reply content after keyword extraction of the question sequence.

[0087] The keyword extraction can be based on a TF-IDF (term frequency-inverse document frequency) algorithm to extract keywords in the user question and system reply content.

[0088] The knowledge base can be a database storing knowledge related to the power marketing field and corresponding knowledge fields.

[0089] The knowledge field category labeling process can be to input the keywords in the first keyword set, call the retrieval agent, match the knowledge field related to the keywords in the knowledge base, label the knowledge field category of the first keyword set according to the matching result, and generate a class of matching information; input the keywords in the second keyword set, call the retrieval agent, match the knowledge field related to the keywords in the knowledge base, label the knowledge field category of the second keyword set according to the matching result, and generate a second class of matching information; the category labeling is generated by semantic matching and rule matching with the knowledge base to generate the category labeling of the knowledge field.

[0090] The question data set can be a data set generated by integrating the first class of matching information and the second class of matching information.

[0091] The first class of matching information can be matching information generated by labeling the knowledge field category of the first keyword set.

[0092] The second class of matching information can be matching information generated by labeling the knowledge field category of the second keyword set.

[0093] The target question sequence can be a high-attention question sequence determined from a plurality of question sequences.

[0094] The portrait vector can be a vector representation of a user portrait constructed according to account attribute information.

[0095] The category labeling sequence can be a sequence S constructed in time sequence by obtaining the knowledge field category labeling corresponding to each round of question and answer in the target question sequence.

[0096] The knowledge field supplementary category prediction model can be an auxiliary prediction model constructed based on a Transformer algorithm, which analyzes the time series trend of the user portrait vector and the category labeling sequence by inputting the user portrait vector and the category labeling sequence, and outputs a potential category that exists with the current user portrait vector and meets the relevance condition and is not involved. The auxiliary prediction model is trained in the training stage with historical high-attention question sequences and their subsequent appearing knowledge fields as supervision samples to learn the time dependence of the category sequence and the correlation between the category and the portrait.

[0097] The knowledge field supplementary category information can be the supplementary category output by the knowledge field supplementary category prediction model.

[0098] The intelligent agent can be a professional processing unit for different knowledge fields, and can process the assigned problem task.

[0099] The reply information can be a reply result generated by the intelligent agent corresponding to the knowledge field after processing the problem task.

[0100] The fusion processing can be to integrate multiple reply information, generate an output data set, and perform data fusion processing.

[0101] The query result can be the final result generated after the fusion processing of the multiple reply information.

[0102] Optionally, the terminal obtains historical interaction information and account attribute information of the user, establishes a question sequence 1 based on the historical interaction information, from the first user-initiated question to the first system reply content, and then sequentially establishes question sequences based on each round of user-initiated question and system reply content until the latest system reply content of the user interaction module, to establish a complete interaction sequence set and construct multiple question sequences including user questions and system replies; performs keyword extraction on the multiple question sequences, extracts keywords in the user questions and system reply contents based on the TF-IDF algorithm, filters the keywords in the user questions to generate a first keyword set, and filters the keywords in the system reply contents to generate a second keyword set; takes the keywords in the first keyword set as input, calls the retrieval agent, matches the knowledge fields related to the keywords in the knowledge base, labels the knowledge field categories of the first keyword set according to the matching results to generate a first matching information, takes the keywords in the second keyword set as input, calls the retrieval agent, matches the knowledge fields related to the keywords in the knowledge base, labels the knowledge field categories of the second keyword set according to the matching results to generate a second matching information, generates the category label of the knowledge field through semantic matching and rule matching with the knowledge base, integrates the first matching information and the second matching information to obtain a question data set; obtains the first matching information and the second matching information in the question sequence 1 and the first matching information and the second matching information in the question sequence 2, compares the first matching information in the question sequence 1 and the first matching information in the question sequence 2, obtains the number of coinciding knowledge field category labels, and if the number is greater than a preset value T, determines that the question sequence 1 is a high-attention sequence; according to the order of the question sequence numbers, a current question sequence and a question sequence corresponding to a question sequence number increased by one are grouped as an analysis unit for processing, until the analysis of each question sequence is completed, the question sequences are renumbered in the original order from small to large, and a target question sequence is determined from the multiple question sequences according to the knowledge field category labels in the question data set; extracts key attribute fields from the user account registration information, including user type, belonging region, power service type, power capacity, historical power consumption data, and payment method, performs normalization processing on the above attribute fields, cleans and removes redundant information and invalid information, maps the above standardized attributes to a user vector U, and the user vector U includes six feature dimensions of user type, belonging region, power service type, power capacity, historical power consumption data, and payment method; constructs a portrait vector of the user according to the account attribute information, obtains the knowledge field category label corresponding to each round of question and answer in the target question sequence, and constructs a category label sequence in chronological order.The image vector and the category label sequence are input into a knowledge field supplementary category prediction model, the knowledge field supplementary category prediction model is constructed based on a Transformer algorithm, the time sequence change trend of the user image vector and the category label sequence is analyzed through inputting the user image vector and the category label sequence, and potential categories that exist with the current user image vector and meet the correlation condition and are not involved are output, knowledge field supplementary category prediction processing is performed, and knowledge field supplementary category information is obtained; according to the knowledge field category label in the target question sequence and the knowledge field supplementary category information, the question task of the user is allocated to an agent in the corresponding knowledge field for processing, and a plurality of reply information is obtained; the plurality of reply information is integrated to generate an output data set, data fusion processing is performed, and a query result of the user is obtained.

[0103] In the above-mentioned power marketing question and answer method driven by multiple agents in cooperation, a plurality of question sequences are constructed by acquiring historical interaction information of the user, and key word extraction and knowledge field category label processing are performed on the question sequences, which is beneficial to identifying the knowledge field that the user pays attention to in the historical interaction, the target question sequence is determined from the plurality of question sequences, which is beneficial to screening the question field that the user pays high attention to, the image vector of the user and the category label sequence are constructed and input into the knowledge field supplementary category prediction model for prediction processing, which is beneficial to predicting the knowledge field supplementary category information that the user needs potentially but has not proposed explicitly, the question task is allocated to the agent in the corresponding knowledge field for processing according to the knowledge field category label in the target question sequence and the knowledge field supplementary category information, which is beneficial to providing accurate reply information for the question proposed by the user and the potential demand, and the plurality of reply information is integrated for fusion processing, which is beneficial to generating a comprehensive query result covering the explicit demand and the potential demand of the user, thereby improving the accuracy and completeness of the question and answer.

[0104] In one exemplary embodiment, according to the matching results between the first keyword set and the second keyword set and the knowledge base, the first keyword set and the second keyword set are subjected to knowledge field category label processing to obtain a question data set, including: matching the keywords in the first keyword set with the keywords in the knowledge base to obtain a first matching result, and performing knowledge field category label processing on the first keyword set according to the first matching result to obtain first category label information; matching the keywords in the second keyword set with the keywords in the knowledge base to obtain a second matching result, and performing knowledge field category label processing on the second keyword set according to the second matching result to obtain second category label information; and determining the question data set according to the first category label information and the second category label information.

[0105] The first matching result can be a matching result obtained by matching the keywords in the first keyword set with the keywords in the knowledge base.

[0106] The first category annotation information can be one type of matching information obtained by performing knowledge field category annotation processing on the first keyword set according to the first matching result.

[0107] The second matching result can be a matching result obtained by matching the keywords in the second keyword set with the keywords in the knowledge base.

[0108] The second category annotation information can be two types of matching information obtained by performing knowledge field category annotation processing on the second keyword set according to the second matching result.

[0109] Optionally, the terminal takes the keywords in the first keyword set as input, calls the retrieval agent, matches the knowledge related to the power marketing field stored in the knowledge base and the corresponding knowledge field, identifies the knowledge field involved in the keywords in the first keyword set through semantic matching and rule matching technology, obtains the first matching result, performs knowledge field category annotation processing on the keywords in the first keyword set according to the first matching result, and generates the first category annotation information; takes the keywords in the second keyword set as input, calls the retrieval agent, matches the knowledge related to the power marketing field stored in the knowledge base and the corresponding knowledge field, identifies the knowledge field involved in the keywords in the second keyword set through semantic matching and rule matching technology, obtains the second matching result, performs knowledge field category annotation processing on the keywords in the second keyword set according to the second matching result, and generates the second category annotation information; and integrates the first category annotation information and the second category annotation information to determine the problem data set.

[0110] The technical scheme provided by the embodiment is advantageous in identifying the knowledge field involved in the user question by matching the keywords in the first keyword set with the keywords in the knowledge base to obtain the first matching result, performing knowledge field category annotation processing on the first keyword set according to the first matching result to obtain the first category annotation information, identifying the knowledge field involved in the system reply by matching the keywords in the second keyword set with the keywords in the knowledge base to obtain the second matching result, performing knowledge field category annotation processing on the second keyword set according to the second matching result to obtain the second category annotation information, and determining the problem data set according to the first category annotation information and the second category annotation information, which is advantageous in comprehensively obtaining the knowledge field category annotation involved in the problem sequence, thereby improving the accuracy and completeness of the knowledge field identification.

[0111] In one exemplary embodiment, determining a target problem sequence from a plurality of problem sequences according to the knowledge field category annotation in the problem data set comprises: determining a category coincidence degree corresponding to the knowledge field category annotation in the problem data set; and determining the target problem sequence from the plurality of problem sequences according to the category coincidence degree.

[0112] The category coincidence degree can be the coincidence degree of the knowledge field category labels between different question sequences in the question data set.

[0113] Optionally, the terminal determines the category coincidence degree corresponding to the knowledge field category labels between different question sequences in the question data set, obtains the one-class matching information and the two-class matching information in the question sequence 1 and the one-class matching information and the two-class matching information in the question sequence 2, compares the one-class matching information in the question sequence 1 with the one-class matching information in the question sequence 2, obtains the number of coincident knowledge field category labels as the category coincidence degree, determines whether the question sequence is a high attention sequence according to the category coincidence degree, determines that the question sequence 1 is a high attention sequence if the category coincidence degree is greater than a preset value T, groups the current question sequence and the question sequence corresponding to the question sequence number incremented by one according to the order of the question sequence number to form an analysis unit for processing, until the analysis of each question sequence is completed, renumbers the question sequences in the original order of the question sequence number from small to large, and determines a target question sequence from the multiple question sequences according to the category coincidence degree.

[0114] The technical scheme provided in this embodiment is advantageous in identifying the knowledge field correlation between different question sequences by determining the category coincidence degree corresponding to the knowledge field category labels in the question data set, and advantageous in screening the question sequence that the user highly pays attention to and repeatedly involves in the historical interaction by determining a target question sequence from multiple question sequences according to the category coincidence degree, thereby being advantageous in identifying the core attention point and the real intention of the user and improving the pertinence and accuracy of subsequent question and answer processing.

[0115] In an example embodiment, according to the account attribute information, an image vector of the user is constructed, including: extracting attribute feature fields of the user from the account attribute information; performing data preprocessing on the attribute feature fields to obtain target attribute feature fields; and constructing the image vector according to the target attribute feature fields.

[0116] The attribute feature fields can be key attribute fields extracted from the account attribute information, including user type, belonging region, power service type, power consumption capacity, historical power consumption data, and payment method.

[0117] The data preprocessing can be normalization processing, cleaning, and eliminating redundant information and invalid information on the attribute feature fields.

[0118] The target attribute feature fields can be standardized attributes obtained by performing data preprocessing on the attribute feature fields.

[0119] Optionally, the terminal extracts attribute feature fields of the user from the account attribute information, including user type, belonging region, power service type, power consumption capacity, historical power consumption data, and payment method, performs data preprocessing on the attribute feature fields, performs normalization processing on the attribute feature fields, cleans and removes redundant information and invalid information, obtains target attribute feature fields, maps the target attribute feature fields into a user vector U, constructs a portrait vector according to the target attribute feature fields, and the portrait vector includes six feature dimensions of user type, belonging region, power service type, power consumption capacity, historical power consumption data, and payment method.

[0120] The technical scheme provided in the embodiment is advantageous in obtaining key attribute information of the user by extracting attribute feature fields of the user from the account attribute information, in standardizing attribute information of different types into a unified format by obtaining target attribute feature fields by performing data preprocessing on the attribute feature fields, and in converting the user attribute information into a vector representation that can be used for model calculation by constructing a portrait vector according to the target attribute feature fields, thereby being advantageous in providing accurate user feature input for subsequent knowledge field supplementary category prediction and improving prediction accuracy.

[0121] In an exemplary embodiment, the attribute feature fields include user type, belonging region, power service type, power consumption capacity, historical power consumption data, and payment method.

[0122] The user type can represent a category to which the user belongs, such as an industrial user or a residential user.

[0123] The belonging region can represent a geographical region where the user is located.

[0124] The power service type can represent a power supply service category used by the user, such as high-voltage power supply or low-voltage power supply.

[0125] The power consumption capacity can represent the power consumption equipment capacity declared by the user.

[0126] The historical power consumption data can represent power consumption information of the user in a historical period.

[0127] The payment method can represent a method of paying electricity fees by the user.

[0128] The technical scheme provided in the embodiment is advantageous in comprehensively obtaining key attribute information of the user in the power marketing field, thereby being advantageous in constructing a more complete and accurate user portrait vector and improving the accuracy of knowledge field supplementary category prediction, and further improving the accuracy of the question and answer.

[0129] In an example embodiment, according to the knowledge field category labels in the target question sequence, a category label sequence is constructed, including: obtaining the knowledge field category label corresponding to each round of question and answer in the target question sequence; and performing permutation and combination processing on the knowledge field category label corresponding to each round of question and answer in the target question sequence according to the time sequence of the target question sequence, to obtain the category label sequence.

[0130] The knowledge field category label corresponding to each round of question and answer can be the knowledge field category label information corresponding to each round of user question and system reply in the target question sequence.

[0131] The permutation and combination processing can be a processing procedure of sequentially arranging the knowledge field category label corresponding to each round of question and answer in the target question sequence according to the time sequence of the target question sequence.

[0132] Optionally, the terminal obtains the knowledge field category label corresponding to each round of question and answer in the target question sequence, extracts the knowledge field category label corresponding to the first round of question and answer, the knowledge field category label corresponding to the second round of question and answer, the knowledge field category label corresponding to the third round of question and answer, and the knowledge field category label corresponding to the last round of question and answer, and performs permutation and combination processing on the knowledge field category label corresponding to each round of question and answer in the target question sequence according to the time sequence of the target question sequence, sequentially arranges the knowledge field category label corresponding to the first round of question and answer, the knowledge field category label corresponding to the second round of question and answer, the knowledge field category label corresponding to the third round of question and answer, and the knowledge field category label corresponding to the last round of question and answer in time sequence, to obtain the category label sequence.

[0133] The technical scheme provided in this embodiment is advantageous in identifying each knowledge field involved by the user in the historical interaction by obtaining the knowledge field category label corresponding to each round of question and answer in the target question sequence, and is advantageous in constructing a sequence reflecting the time evolution law of the user's focus by performing permutation and combination processing on the knowledge field category label corresponding to each round of question and answer in the target question sequence according to the time sequence of the target question sequence, thereby being advantageous in providing time sequence feature input for the knowledge field category supplement prediction model and improving the accuracy of the category supplement prediction.

[0134] The following describes a multi-agent collaborative driving power marketing question and answer system provided by the present application with an example.

[0135] The power marketing question and answer system is an intelligent dialogue system for power users and marketing business personnel, providing power-related information consultation and business handling support, and realizing question and answer response, business guidance and service recommendation through natural language interaction.

[0136] The power marketing question and answer system has been widely applied to customer service and marketing links of power enterprises, and is in the stage of transformation from basic intelligence to deep intelligence. The current system is mostly based on rule engine and FAQ (frequently asked questions) template construction, which can support users to quickly query and respond to common business problems such as electricity price, electricity charge, payment method and power failure notification, and improve user service efficiency.

[0137] The existing power marketing question and answer system still has certain limitations in semantic understanding, especially in handling complex problems. Due to the lack of professional knowledge in the field of power marketing, some users have obstacles in language expression, which makes it difficult for them to accurately convey the core intent to the system in the multi-round question and answer process, thereby causing repeated questioning. The existing system usually only responds to single-round questioning, lacks comprehensive analysis of the context, and cannot conduct deep understanding and reasoning based on the user's historical interaction information and account portrait, thereby affecting the accuracy of the reply and the service experience.

[0138] The purpose of the embodiment is to solve the technical problem that the existing system usually only responds to single-round questioning, lacks comprehensive analysis of the context, and cannot conduct deep understanding and reasoning based on the user's historical interaction information and account portrait.

[0139] As shown in Figure 2 The embodiment provides a multi-agent collaborative driving power marketing question and answer system, which comprises a user interaction module, an intent analysis module, a multi-agent collaboration module and an information fusion and response module.

[0140] The user interaction module is used for obtaining the historical interaction information of the user and the corresponding account registration information and sending them to the intent analysis module.

[0141] The intent analysis module is used for receiving the historical interaction information and the account registration information.

[0142] The historical interaction information is preprocessed to construct a problem sequence composed of user questions and system replies. Key words are extracted from the problem sequence, the knowledge domain in the knowledge base is matched based on the key words, the category is labeled, and a problem data set is generated.

[0143] The user account registration information is filtered and structured vector is constructed to generate a user portrait.

[0144] The problem data set and the user portrait are sent to the multi-agent collaboration module.

[0145] The problem data set is generated, and the specific steps are as follows:

[0146] Based on historical interaction information, starting with the first question initiated by the user and ending with the first system response, a question sequence 1 is established. Subsequently, a question sequence is established sequentially for each round of user-initiated questions and system responses, until the latest system response from the user interaction module is reached, thus establishing a complete set of interaction sequences.

[0147] Keyword extraction is performed on each question sequence. Keywords from user questions are filtered to generate a first set, and keywords from system responses are filtered to generate a second set.

[0148] Keywords within a set are used as input to invoke a retrieval agent, which matches the knowledge domains involved in the keywords in the knowledge base. Based on the matching results, the set is labeled with knowledge domain categories, and a set of matching information is generated.

[0149] The keywords in the second set are used as input, and the retrieval agent is invoked to match the knowledge domains involved in the keywords in the knowledge base. Based on the matching results, the knowledge domain categories of the second set are labeled, and second-class matching information is generated.

[0150] Keyword extraction is based on the TF-IDF algorithm to extract keywords from user questions and system responses;

[0151] Category labeling generates category labels for the knowledge domain by semantic and rule matching with the knowledge base;

[0152] Integrate primary and secondary matching information to generate a problem dataset;

[0153] The steps for filtering user account registration information and building user profiles are as follows:

[0154] Extract key attribute fields from user account registration information, including user type, region, type of electricity service, electricity capacity, historical electricity consumption data, and payment method;

[0155] The above attribute fields are normalized to clean up and remove redundant and invalid information.

[0156] By using a multi-agent collaborative driving mechanism, combined with user historical interaction information and account registration information, the system achieves accurate positioning and dynamic supplementation of knowledge domain category tags based on user profiles, effectively overcoming the shortcomings of insufficient understanding of complex questions and inadequate utilization of context in the existing power marketing Q&A system.

[0157] The above standardized attributes are mapped to user vector U, which includes six feature dimensions: user type, region, type of electricity service, electricity capacity, historical electricity consumption data, and payment method.

[0158] A multi-agent collaborative module is configured to receive a question data set and a user portrait;

[0159] An intent direction analysis is performed on the question data set based on the user portrait, the question sequence is divided, and a high-attention question sequence is identified and generated;

[0160] Based on the user portrait, an auxiliary prediction model is constructed to make a supplementary prediction of the potential category of the high-attention question sequence;

[0161] According to the knowledge field category label in the high-attention question sequence, the question task is assigned to the agent of the corresponding knowledge field for processing, and multiple reply information is generated;

[0162] The multi-agent parallelly processes different knowledge field tasks, significantly shortens the response time, and improves the processing efficiency and professionalism;

[0163] The reply information is integrated to generate an output data set, which is sent to an information fusion and response module;

[0164] The question sequence is divided, and the specific operation is as follows:

[0165] Obtain the first-class matching information and the second-class matching information in the question sequence 1 and the first-class matching information and the second-class matching information in the question sequence 2,

[0166] Compare the first-class matching information in the question sequence 1 with the first-class matching information in the question sequence 2, obtain the number of coincident knowledge field category labels, and if the number is greater than a preset value T, determine that the question sequence 1 is a high-attention sequence, and the value of T can be set according to the actual business scenario;

[0167] Thereafter, according to the order of the question sequence number, the current question sequence and the question sequence corresponding to the sequence number increased by one are combined into a group as an analysis unit for processing until the analysis of each question sequence is completed;

[0168] According to the original question sequence number, the question sequence is renumbered from small to large to generate a high-attention question sequence;

[0169] The high-attention question sequence is supplemented, and the specific steps are as follows:

[0170] Obtain the knowledge field category label corresponding to each round of question and answer in the current high-attention question sequence, and construct a category label sequence S in chronological order;

[0171] Obtain the user portrait vector U;

[0172] The category label sequence S and the user portrait vector U are input into the auxiliary prediction model;

[0173] The auxiliary prediction model is based on a Transformer algorithm, analyzes the time series trend by inputting the user portrait vector U and the category label sequence S, and outputs potential categories that exist with the current user portrait vector U and meet the relevance condition and have not been involved;

[0174] The relevance condition is that the relevance of the output to the current user portrait is higher than a threshold T, and the potential knowledge domain category that has not appeared in the current high attention problem sequence is used as a supplementary category;

[0175] Through high attention problem sequence analysis and category supplement prediction, the user's potential demand can be predicted in advance, the number of repeated questions can be reduced, and the recognition accuracy of the system to the user's true intention can be improved;

[0176] The specific operation steps of the auxiliary prediction model are as follows:

[0177] Map each candidate category in the category label sequence to a semantic embedding vector, and concatenate the sequence embedding with the user portrait vector and the position encoding to construct the input tensor of the auxiliary prediction model;

[0178] Input the input tensor into the trained auxiliary prediction model. The auxiliary prediction model is based on the Transformer algorithm, and is trained in the training stage with historical high attention problem sequences and subsequent knowledge domains as supervision samples to learn the time dependence of the category sequence and the correlation between the category and the portrait;

[0179] The auxiliary prediction model outputs a set of candidate potential knowledge domain categories and corresponding confidence scores. The candidate set is filtered and sorted: first, eliminate the categories that have appeared in the current high attention problem sequence;

[0180] Then, sort them in descending order of confidence score, and select the categories with confidence scores greater than the threshold T as supplementary categories;

[0181] Write the supplementary categories back to the problem data set as input for subsequent task allocation and agent processing. At the same time, save the confidence and attention weight output by the model for result interpretability display and as a basis for user feedback and model retraining;

[0182] The auxiliary prediction model supports online inference and offline periodic fine-tuning. The model parameters are updated using user feedback and new interaction data to improve prediction accuracy;

[0183] If the model does not output a supplementary category that meets the conditions, the system can call a fallback mechanism based on knowledge base retrieval to obtain candidate supplementary categories;

[0184] The threshold T is automatically determined according to the historical data training results;

[0185] The correlation between the potential knowledge field category that has not appeared in the current high-attention problem sequence and the user portrait is calculated, and the specific steps are as follows:

[0186] The definition category is , (i = 1, 2, 3…, i is the number of supplementary categories), then The correlation between U and the category The calculation formula of

[0187] ;

[0188] Wherein, is the semantic embedding vector of the category , is the vector splicing function, is the sigmoid activation function, wherein x is the independent variable symbol; and are weight parameters obtained by model training. The setting of the interval and the size of the threshold is to facilitate comparison. The size of the threshold depends on the number of sample data and the base number set by the person skilled in the art for each group of sample data; as long as it does not affect the proportional relationship between the parameters and the quantized values.

[0189] The above formulas are all dimensionless values, and the formula is obtained by software simulation of a large amount of data to obtain the most real situation. The preset parameters in the formula are set by the person skilled in the art according to the actual situation.

[0190] The technical scheme provided by the embodiment has the beneficial effects that: the multi-agent collaborative driving power marketing question and answer system obtains the historical interaction information of the user, constructs a problem sequence and extracts keywords, matches related knowledge field category labels; obtains corresponding account registration information, and constructs a user portrait; through analysis and screening of the knowledge field category labels, the user portrait is used to complete the supplement of the screened knowledge field category labels, so as to perform directional positioning of the real intention of the user; through the multi-agent, different category labels are quickly processed in different knowledge fields, and reply information of each knowledge field is generated, and the final query result is generated after fusion; effectively overcome the defects of insufficient understanding of composite problems and insufficient use of context in the existing power marketing question and answer system; at the same time, the multi-agent parallelly processes different knowledge field tasks, significantly shortens the response time, improves the processing efficiency and professionalism; the overall scheme supports online learning and user feedback closed-loop optimization, has strong adaptability and scalability, and can continuously optimize the quality of question and answer in multiple rounds of interaction scenarios.

[0191]

[0192] ​It should be understood that although each step in the flowchart involved in each embodiment as described above is shown in sequence according to the direction of the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise explicitly stated herein, there is no strict order limitation for the execution of these steps, and these steps can be executed in other orders. Moreover, at least some of the steps in the flowchart involved in each embodiment as described above can include multiple steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be alternately or alternately executed with at least part of other steps or steps or stages in other steps.

[0193] Based on the same inventive concept, the embodiments of the present application also provide a multi-agent collaborative driving power marketing question and answer device for implementing the above-mentioned multi-agent collaborative driving power marketing question and answer method. The problem solving implementation scheme provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more multi-agent collaborative driving power marketing question and answer device embodiments provided below can refer to the limitations of the multi-agent collaborative driving power marketing question and answer method in the above text, which will not be repeated here.

[0194] In one exemplary embodiment, as shown in Figure 3 A multi-agent collaborative driving power marketing question and answer device 300 is provided, which can include:

[0195] An information acquisition module 301 is configured to acquire historical interaction information and account attribute information of a user, and construct a plurality of question sequences according to the historical interaction information; the plurality of question sequences include user questions and system replies;

[0196] A sequence extraction module 302 is configured to perform keyword extraction on the plurality of question sequences to obtain a first keyword set in the user questions and a second keyword set in the system replies;

[0197] A labeling processing module 303 is configured to perform knowledge domain category labeling processing on the first keyword set and the second keyword set according to a matching result between the first keyword set and the second keyword set and a knowledge base, to obtain a question data set;

[0198] A sequence determination module 304 is configured to determine a target question sequence from the plurality of question sequences according to the knowledge domain category labeling in the question data set;

[0199] A vector construction module 305 is configured to construct a portrait vector of the user according to the account attribute information, and construct a category labeling sequence according to the knowledge domain category labeling in the target question sequence.

[0200] The sequence input module 306 is configured to input the portrait vector and the category label sequence into the knowledge domain supplementary category prediction model to perform knowledge domain supplementary category prediction processing, and obtain knowledge domain supplementary category information.

[0201] The task allocation module 307 is configured to allocate the user's question task to an agent of a corresponding knowledge domain according to the knowledge domain category label in the target question sequence and the knowledge domain supplementary category information, and obtain a plurality of reply information.

[0202] The information fusion module 308 is configured to perform fusion processing on the plurality of reply information, and obtain a query result of the user.

[0203] In an exemplary embodiment, the labeling processing module 303 is further configured to match the keywords in the first keyword set with the keywords in the knowledge base to obtain a first matching result, perform knowledge domain category labeling processing on the first keyword set according to the first matching result, and obtain first category label information; match the keywords in the second keyword set with the keywords in the knowledge base to obtain a second matching result, perform knowledge domain category labeling processing on the second keyword set according to the second matching result, and obtain second category label information; and determine the question data set according to the first category label information and the second category label information.

[0204] In an exemplary embodiment, the sequence determination module 304 is further configured to determine a category coincidence degree corresponding to the knowledge domain category label in the question data set; and determine the target question sequence from the plurality of question sequences according to the category coincidence degree.

[0205] In an exemplary embodiment, the vector construction module 305 is further configured to extract an attribute feature field of the user from the account attribute information; perform data preprocessing on the attribute feature field to obtain a target attribute feature field; and construct the portrait vector according to the target attribute feature field.

[0206] In an exemplary embodiment, the attribute feature field includes user type, belonging region, power service type, power consumption capacity, historical power consumption data, and payment method.

[0207] In an exemplary embodiment, the vector construction module 305 is further configured to obtain a knowledge domain category label corresponding to each round of question and answer in the target question sequence; and perform permutation and combination processing on the knowledge domain category label corresponding to each round of question and answer in the target question sequence according to a time sequence of the target question sequence, to obtain a category label sequence.

[0208] The modules in the multi-agent collaborative driving power marketing question and answer device can be implemented by software, hardware, or a combination thereof. The modules can be embedded in or independent of a processor in a computer device in hardware form, or stored in a memory in the computer device in software form, so that the processor can call and execute the operations corresponding to the modules.

[0209] In an example embodiment, a computer device, which can be a terminal, has an internal structure as shown in Figure 4 The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit, and an input device. The processor, the memory, and the input / output interface are connected through a system bus, and the communication interface, the display unit, and the input device are connected to the system bus through the input / output interface. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is configured to exchange information between the processor and external devices. The communication interface of the computer device is configured to perform wired or wireless communication with external terminals. The wireless communication can be achieved through WIFI, mobile cellular network, NFC (Near Field Communication), or other technologies. The computer program is executed by the processor to implement a multi-agent collaborative driving power marketing question and answer method. The display unit of the computer device is configured to form a visually visible picture, which can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball, or touchpad arranged on the housing of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0210] Those skilled in the art can understand that Figure 4 The structure shown in the above description is only a block diagram of part of the structure related to the scheme of the present application, and does not limit the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0211] In an example embodiment, a computer device is also provided, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0212] In an exemplary embodiment, a computer readable storage medium is provided, having stored thereon a computer program which, when executed by a processor, implements the steps of any of the above method embodiments.

[0213] In an exemplary embodiment, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the steps of any of the above method embodiments.

[0214] A person of ordinary skill in the art can understand that all or part of the processes in the above method embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium and can include the processes of the above method embodiments when executed. Any reference to a memory, database, or other medium in the embodiments provided in the present application can include at least one of a non-volatile and volatile memory. The non-volatile memory can include a read-only memory (ROM), a magnetic tape, a floppy disk, a flash memory, an optical storage, a high-density embedded non-volatile memory, a resistive memory (ReRAM), a magnetoresistive random access memory (MRAM), a ferroelectric random access memory (FRAM), a phase change memory (PCM), a graphene memory, etc. The volatile memory can include a random access memory (RAM) or an external cache memory, etc. As an illustration but not as a limitation, the RAM can be in various forms such as a static random access memory (SRAM) or a dynamic random access memory (DRAM), etc. The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0215] The technical features of the above embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combinations of the technical features do not contradict each other, they should be considered as within the scope of the present disclosure.

[0216] The above-described embodiments are merely illustrative of several embodiments of the present application, which are described in more detail and in a specific manner, but should not be construed as limiting the scope of the patent of the present application. It should be noted that, for those of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for power marketing question and answer driven by multi-agent collaboration, characterized in that, The method comprises: obtaining historical interaction information and account attribute information of a user, and constructing a plurality of question sequences according to the historical interaction information; the plurality of question sequences comprise user questions and system replies; performing keyword extraction on the plurality of question sequences to obtain a first keyword set in the user questions and a second keyword set in the system replies; performing knowledge domain category labeling processing on the first keyword set and the second keyword set according to a matching result between the first keyword set, the second keyword set and a knowledge base to obtain a question data set; determining a target question sequence from the plurality of question sequences according to the knowledge domain category labeling in the question data set; constructing a portrait vector of the user according to the account attribute information, and constructing a category labeling sequence according to the knowledge domain category labeling in the target question sequence; inputting the portrait vector and the category labeling sequence into a knowledge domain supplementary category prediction model to perform knowledge domain supplementary category prediction processing to obtain knowledge domain supplementary category information; allocating a question task of the user to an agent of a corresponding knowledge domain for processing according to the knowledge domain category labeling in the target question sequence and the knowledge domain supplementary category information to obtain a plurality of reply information; performing fusion processing on the plurality of reply information to obtain a query result of the user.

2. The method of claim 1, wherein, The knowledge domain category labeling processing on the first keyword set and the second keyword set according to the matching result between the first keyword set, the second keyword set and the knowledge base to obtain the question data set comprises: matching keywords in the first keyword set with keywords in the knowledge base to obtain a first matching result, and performing knowledge domain category labeling processing on the first keyword set according to the first matching result to obtain first category labeling information; matching keywords in the second keyword set with keywords in the knowledge base to obtain a second matching result, and performing knowledge domain category labeling processing on the second keyword set according to the second matching result to obtain second category labeling information; determining the question data set according to the first category labeling information and the second category labeling information.

3. The method of claim 1, wherein, The determination of the target question sequence from the plurality of question sequences according to the knowledge domain category labeling in the question data set comprises: determining a category coincidence degree corresponding to the knowledge domain category labeling in the question data set; determining the target question sequence from the plurality of question sequences according to the category coincidence degree.

4. The method of claim 1, wherein, The construction of the portrait vector of the user according to the account attribute information comprises: extracting attribute feature fields of the user from the account attribute information; performing data preprocessing on the attribute feature fields to obtain target attribute feature fields; constructing the portrait vector according to the target attribute feature fields.

5. The method of claim 4, wherein, The attribute feature fields comprise user type, belonging region, power service type, power consumption capacity, historical power consumption data and payment method.

6. The method according to any one of claims 1 to 5, characterized in that, The constructing a category label sequence according to the knowledge field category labels in the target question sequence comprises: obtaining the knowledge field category label corresponding to each round of question and answer in the target question sequence; performing permutation and combination processing on the knowledge field category label corresponding to each round of question and answer in the target question sequence according to the time sequence of the target question sequence to obtain the category label sequence.

7. A power marketing question and answer device driven by multi-agent cooperation, characterized in that, The device comprises: an information acquisition module configured to acquire historical interaction information and account attribute information of a user, and construct a plurality of question sequences according to the historical interaction information; the plurality of question sequences comprise user questions and system replies; a sequence extraction module configured to extract keywords from the plurality of question sequences to obtain a first keyword set in the user questions and a second keyword set in the system replies; a label processing module configured to perform knowledge field category label processing on the first keyword set and the second keyword set according to a matching result between the first keyword set and the second keyword set and a knowledge base to obtain a question data set; a sequence determination module configured to determine a target question sequence from the plurality of question sequences according to the knowledge field category label in the question data set; a vector construction module configured to construct a portrait vector of the user according to the account attribute information, and construct a category label sequence according to the knowledge field category label in the target question sequence; a sequence input module configured to input the portrait vector and the category label sequence to a knowledge field supplementary category prediction model to perform knowledge field supplementary category prediction processing to obtain knowledge field supplementary category information; a task allocation module configured to allocate a question task of the user to an agent in a corresponding knowledge field according to the knowledge field category label in the target question sequence and the knowledge field supplementary category information to obtain a plurality of reply information; an information fusion module configured to perform fusion processing on the plurality of reply information to obtain a query result of the user.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to implement the steps of the method of any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 6.

Images