Intelligent question answering method and system based on deep learning and credit loan knowledge graph

Abstract

The invention discloses an intelligent question and answer method and system based on deep learning and a credit loan knowledge graph. The method comprises the following steps: constructing the credit loan knowledge graph; designing and training a double-layer generative adversarial network model; semantic analysis of user query; generating and evaluating a final answer; real-time interaction with the user is carried out by using electricity sales, insurance elimination and collection records in the credit loan knowledge graph in the loan and money return process; performing early warning on the complaint intention of the user in combination with an answer generated by the double-layer generative adversarial network model, and performing text quality inspection according to dialogue content; and analyzing the dialogue text and generating a quality inspection report. By constructing a double-layer generative adversarial network model and combining with a credit loan knowledge graph, accurate understanding of user query and knowledge reasoning in a professional field are realized, so that the system can quickly convert the user query into structured semantic information after receiving the user query, and performs efficient matching and reasoning with entities and relationships in the knowledge graph; and the accuracy and correlation of the answers are ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent question answering, and relates to an intelligent question answering method and system based on deep learning and credit loan knowledge graph. Background Art

[0002] In the field of modern financial technology, credit loans are an important financial product widely used in the financing needs of individuals and small and micro enterprises. However, with the increase in the number of users and the increase in business complexity, how to effectively manage user information, accurately assess credit risks, and provide personalized services in credit loan business has become an urgent problem to be solved. Existing intelligent question-answering systems usually rely on natural language processing technology, which can understand and respond to user queries to a certain extent. However, when dealing with the financial field, especially credit loan business, the system faces multiple challenges. First, the credit loan field involves a large number of professional terms and complex business logic. Traditional NLP (natural language processing) technology often finds it difficult to accurately understand and parse these contents, resulting in the system's inaccurate understanding of user intentions, which in turn affects the accuracy and relevance of the answer; secondly, most of the existing question-answering systems are trained based on data from general fields, lack of special optimization for credit loan business, and cannot fully respond to special needs in this field. In addition, the existing systems usually only provide standardized answers, lack the ability to deeply analyze the individual credit status of users and provide personalized services, and cannot meet the growing personalized needs of users.

[0003] In order to improve these shortcomings, some studies have begun to try to introduce knowledge graph technology into intelligent question-answering systems. Knowledge graphs organize and manage knowledge through graph structures, which can better understand the complex relationships between data. However, there are still several key problems when applying knowledge graphs to credit loan intelligent question-answering systems. First, how to effectively combine natural language processing with knowledge graphs so that the system can understand complex natural languages ​​and perform knowledge reasoning in professional fields; second, existing knowledge graphs can usually only provide static knowledge, and cannot be dynamically updated and reasoned based on users' real-time data, which limits the intelligence level of the system. Finally, the existing system lacks comprehensive utilization of user credit data and cannot achieve personalized credit assessment and risk management.

[0004] To sum up, the existing technology has the following main problems in the application of credit loan intelligent question and answer system: first, the system has comprehension deviations when dealing with credit loan professional terms and complex business logic, which affects the quality of responses; second, it lacks special optimization for credit loan business and cannot effectively respond to special needs in the field; finally, the existing system lacks the ability to conduct in-depth analysis of the user's individual credit status and provide personalized services, making it difficult to meet the personalized needs of users. The existence of the above problems limits the application effect and service level of the existing intelligent question and answer system in the field of credit loans. Summary of the invention

[0005] The purpose of the present invention is to solve the problem in the prior art that there is a misunderstanding when dealing with credit loan professional terms and complex business logic, which affects the quality of the response, and to provide an intelligent question-answering method and system based on deep learning and credit loan knowledge graph.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The intelligent question-answering method based on deep learning and credit loan knowledge graph includes the following steps:

[0008] Build a credit loan knowledge graph, which includes entities and relationships in the process of attracting customers, lending, and collecting payments related to the credit loan business;

[0009] Design and train a two-layer generative adversarial network model, which includes a first generator, a second generator, a first discriminator and a second discriminator; the first generator is used to generate a preliminary answer, and the second generator is used to generate a final answer based on the preliminary answer and combined with user historical data; the two discriminators are used to evaluate the accuracy and relevance of the generated answers respectively;

[0010] Semantic parsing of user queries: using a pre-trained natural language processing model to semantically parse the query input by the user, and passing the parsed result as input to the first generator to generate a real-time preliminary answer;

[0011] Generation and screening of preliminary answers: The first generator generates real-time preliminary answers based on the parsing results, matches the real-time preliminary answers with the credit loan knowledge graph, and evaluates their relevance through the first discriminator, screening out real-time preliminary answers that are highly relevant to entities and relationships in the knowledge graph, and then passes the screened real-time preliminary answers to the second generator;

[0012] Generation and evaluation of the final answer. The second generator generates a real-time final answer based on the screened real-time preliminary answer, combined with the user's historical credit data, repayment record, and loan application history information. The real-time final answer is evaluated by the second discriminator to ensure its accuracy and practicality.

[0013] In the process of lending and collecting payments, the telemarketing, consumer protection and collection records in the credit loan knowledge graph are used to interact with users in real time through online customer service robots. The answers generated by the two-layer generative adversarial network model are combined to warn users of their complaint intentions and conduct text quality inspection based on the content of the conversation.

[0014] Conversation text analysis and quality inspection report generation: Use a two-layer generative adversarial network model to parse the conversation text in the customer service system, extract key fields and classify them, and generate quality inspection reports related to the lending and collection processes based on the analysis results.

[0015] The construction of the credit loan knowledge graph specifically includes the following steps:

[0016] Identify and define the key entities of multi-dimensional user portraits, including user credit data entity E c , User consumption behavior entity E s , User social data entity E n 、Loan Product Entity E p , Repayment Model Entity E r and lending institution entity E l The definition of each entity is determined according to the role of the corresponding entity in the credit loan business:

[0017]

[0018] Among them, E i is the weighted comprehensive representation of entity i, α k is the weight factor, indicating the importance of the kth attribute to the entity, A ik is the kth attribute value of the i-th entity, K i is the total number of attributes involved in entity i;

[0019] The multi-dimensional information of the user's credit data, consumption behavior and social data is associated with the loan products and repayment patterns to build the relationship between entities, including the adaptability relationship between the user's credit data and the loan products. cp , the matching relationship between user consumption behavior and repayment mode R sr The influence relationship between user social data and credit data R nc :

[0020]

[0021] Among them, R ij Represents entity E i With another entity E j The strength of the relationship between and Entity E i With another entity E j The embedding representation in the mth dimension, β m is the weight factor of the mth dimension, and M is the total number of dimensions;

[0022] Assign a weight parameter W to each entity relationship r , weight parameter Wr Used to describe the importance of different data dimensions in user portraits. The weight parameters are obtained through analysis and statistical calculation of historical data:

[0023]

[0024] Among them, γ is the relationship with R i The importance coefficient of the relevant historical data, λ is the adjustment parameter to control the convergence speed of the relationship weight, and n is the total number of relationships;

[0025] In the multi-dimensional user portrait credit loan knowledge graph, each entity is connected through a relationship to construct a user portrait graph structure G(V,E), where V represents the set of entity nodes and E represents the set of relationship edges between entities. The graph structure is represented in the form of a triple as (E a ,R b ,E c ), where E a and E c are entities of different dimensions, R b The relationship between the two is:

[0026]

[0027] Where A is the number of different entity combinations in the graph, δ i is the weight factor of the ith entity combination, and × represents the Cartesian product operation, which ensures that the combination of entities and relationships forms a complete multi-dimensional user portrait credit loan knowledge graph structure, and the weight of each entity in the graph can be reflected by accumulation;

[0028] The user portraits are classified and processed through clustering algorithms, and the user's credit data, consumption behavior and social data are clustered into corresponding user groups. The multi-dimensional user portrait credit loan knowledge graph is hierarchically processed so that users can locate the corresponding user groups and corresponding loan products when they query:

[0029]

[0030] Among them, C k is the kth cluster center, is the feature vector of the i-th user in the j-th dimension, is the mean vector of the kth cluster center in the jth dimension, δ ij It is the matching degree between the user and the cluster center in the corresponding dimension;

[0031] Through data cleaning and ontology matching technology, redundant and noisy data in the multi-dimensional user portrait credit loan knowledge graph is eliminated, and the multi-dimensional user portrait credit loan knowledge graph is semantically enhanced:

[0032]

[0033] Among them, Sim(R b ) is the semantic similarity of the relationship, η m is the similarity weight factor of the mth dimension, and are the embedding representations of the i-th and j-th entities in the m-th dimension, respectively, and ||·|| is the modulus length operation of the vector;

[0034] Ultimately, a multi-dimensional user portrait credit loan knowledge graph is generated that covers user credit data, consumption behavior and social data, and is associated with loan products and repayment model entities.

[0035] The two-layer generative adversarial network model includes:

[0036] The first generator is used to generate preliminary answers related to the user's query. The first generator is trained based on the semantic information of the user's query and the entities and relationships in the credit loan knowledge graph;

[0037] The first discriminator is used to evaluate the relevance and authenticity of the preliminary answer generated by the first generator, match the preliminary answer with the credit loan knowledge graph, and determine whether it conforms to the logic and relationship in the credit loan knowledge graph;

[0038] The second generator generates a final answer based on the preliminary answer generated by the first generator and in combination with the user's historical data and needs;

[0039] The second discriminator is used to evaluate the final answer generated by the second generator and determine whether the final answer is consistent with the user's historical data and credit status.

[0040] The generation and screening of the preliminary answers specifically include the following steps:

[0041] Construct the first generator G1, which receives the semantic information q of the user query and the entities and relationships (E a ,R b ,E c ) as input, and generate a preliminary answer a1 through a multi-layer neural network:

[0042]

[0043] Among them, q is the semantic vector representation of the user query, h i is the historical context vector related to the entities and relations in the multi-dimensional user portrait credit loan knowledge graph, α i is the weight coefficient based on the attention mechanism, and are the embedding representations of entities and relations in the multi-dimensional user portrait credit loan knowledge graph, W h and W o is the weight matrix of the first generator G1, b h and b o is the bias term, represents the vector concatenation operation, || represents the connection operation, ReLU is the activation function, and softmax is the normalization function;

[0044] The first discriminator D1 is trained to evaluate the relevance and authenticity of the preliminary answer a1 generated by the first generator G1, match the preliminary answer a1 with the relationship in the multi-dimensional user portrait credit loan knowledge graph, and generate an evaluation score s1 through the following scoring mechanism:

[0045]

[0046] Among them, W d is the weight matrix of the discriminator, b d is the bias term, σ is the Sigmoid activation function, and are the embedding representations of entities and relations related to the current query in the multi-dimensional user portrait credit loan knowledge graph, γ j is the importance coefficient of the entity relationship, ⊙ represents the element-level product operation, and Z is the normalization factor.

[0047] The generation and evaluation of the final answer specifically includes the following steps:

[0048] Construct the second generator G2, which is based on the preliminary answer a1 generated by the first generator G1 and combines the user's historical data h and demand d to generate the final answer a2:

[0049]

[0050] Among them, W f is the weight matrix of the second generator, b f is the bias term, β k is the weight coefficient between the user's historical data and demand, GRU is the gated recurrent unit, which is used to fuse the user's historical data h and demand d, and tanh is the hyperbolic tangent activation function;

[0051] Design and train the second discriminator D2, which is used to evaluate the personalized matching degree of the final answer a2 generated by the second generator G2, match the final answer a2 with the user's historical data h and credit status, and generate the evaluation score s2 through the following formula:

[0052]

[0053] Among them, W g is the weight matrix of the second discriminator, b g is the bias term, δ n is the weight factor that matches the user's historical data, and BiLSTM is a bidirectional long short-term memory network used to analyze the time series correlation of the user's historical data;

[0054] By alternately training the first generator G1, the first discriminator D1, the second generator G2, and the second discriminator D2, the overall performance of the generative adversarial network is optimized so that the generated final answer a2 is optimal:

[0055]

[0056] in, is the loss function of the generative adversarial network, represents the expected value, p g1 and p g2 Probability distribution of generated answers for the first and second generators respectively.

[0057] The real-time interaction with the user through the online customer service robot is as follows:

[0058] In the process of lending and collecting payments, the telemarketing records in the credit loan knowledge graph based on multi-dimensional user portraits are tx 、Consumer Protection Records E cb and collection records cs Real-time interaction: Telemarketing records, consumer protection records, and collection records represent the user's interactive behaviors and related historical data at different stages.

[0059] Using the final answer a2 and the user’s real-time conversation content c, the probability distribution P of the user’s complaint intention is calculated through the attention mechanism. comp ;

[0060] According to the probability distribution P of complaint intention comp Set the threshold T comp , when P comp ≥T comp When alert , the early warning module generates corresponding early warning signals and sends alarm information to customer service personnel:

[0061]

[0062] Among them, A alert is the warning signal, the value is 1 for triggering the warning, and 0 for not triggering the warning. comp is the preset probability threshold of complaint intention;

[0063] The text quality inspection according to the conversation content is specifically as follows:

[0064] When performing text quality inspection, the text analysis module is used to extract key information from user conversations and compare it with entity E in the multi-dimensional user portrait credit loan knowledge graph. i and the relation R j Compare and generate quality inspection report Q rep :

[0065]

[0066] Among them, W q is the weight matrix of the quality inspection module, b q is the bias term, γ ij For entity E i With the relationship R j The matching degree in the dialogue text, ReLU is the activation function, sum is the weighted sum operation, Q rep Generated text quality inspection reports are used to evaluate the quality and compliance of user conversation content;

[0067] During the lending and collection process, the interaction records between users and customer service robots are continuously monitored, and the telemarketing records, consumer protection records and collection records in the multi-dimensional user portrait credit loan knowledge graph are dynamically updated based on the real-time conversation content.

[0068] The analysis of the dialogue text and the generation of a quality inspection report specifically include:

[0069] The first generator of the two-layer generative adversarial network model performs preliminary analysis on the conversation text in the customer service system to identify and extract user intent, question type, related entities and relationships in the conversation;

[0070] After extracting key information, the second generator of the two-layer generative adversarial network model is used to classify the information and map it to specific categories in the lending and collection process;

[0071] After the classification is completed, the classification results are evaluated by the second discriminator of the two-layer generative adversarial network model. The system evaluates the matching degree between the results and the entities and relationships in the multi-dimensional user portrait credit loan knowledge graph, verifies the accuracy and relevance of the classification, and makes the classification results consistent with the user's actual credit status, historical behavior data and related business rules;

[0072] Based on the classification evaluation results, a quality inspection report related to the lending and collection process is generated. The quality inspection report summarizes the key data and matching evaluation results involved in the customer service dialogue and is used to evaluate the quality of the entire lending and collection process;

[0073] After the quality inspection report is generated, the system will feed back the quality inspection report to the business department. At the same time, based on the latest quality inspection report and actual business conditions, it will dynamically update the data in the multi-dimensional user portrait credit loan knowledge graph.

[0074] The intelligent question-answering system based on deep learning and credit loan knowledge graph includes the following modules:

[0075] A credit loan knowledge graph construction module, which is used to construct a credit loan knowledge graph, and to construct a knowledge graph containing entities and relationships in the process of attracting, lending, and collecting payments related to the credit loan business;

[0076] A two-layer generative adversarial network model design module, which is used to design and train a two-layer generative adversarial network model. The two-layer generative adversarial network model includes a first generator, a second generator, a first discriminator, and a second discriminator. The first generator is used to generate a preliminary answer, and the second generator is used to generate a final answer based on the preliminary answer and combined with user historical data. The two discriminators are used to evaluate the accuracy and relevance of the generated answer respectively.

[0077] A user query semantic parsing module, which is used for semantic parsing of user queries, performs semantic parsing on the query input by the user using a pre-trained natural language processing model, and passes the parsed result as input to the first generator to generate a real-time preliminary answer;

[0078] A preliminary answer generation and screening module, which is used to generate and screen preliminary answers. The first generator generates a real-time preliminary answer based on the parsing result, matches the real-time preliminary answer with the credit loan knowledge graph, and evaluates its relevance through the first discriminator, and screens out the real-time preliminary answer that is highly relevant to the entities and relationships in the knowledge graph, and then passes the screened real-time preliminary answer to the second generator;

[0079] A final answer generation and evaluation module, which is used to generate and evaluate the final answer. The second generator generates a real-time final answer based on the screened real-time preliminary answer, combined with the user's historical credit data, repayment record and loan application history information. The real-time final answer is evaluated by the second discriminator to ensure its accuracy and practicality;

[0080] Real-time interaction and early warning module, which is used to use the telemarketing, consumer protection and collection records in the credit loan knowledge graph in the process of lending and collection, and interact with users in real time through online customer service robots; combined with the answers generated by the two-layer generative adversarial network model, it warns users of their complaint intentions and conducts text quality inspection based on the content of the conversation;

[0081] The dialogue text parsing and quality inspection report generation module is used for parsing dialogue text and generating quality inspection reports. It uses a two-layer generative adversarial network model to parse dialogue texts in the customer service system, extract key fields and classify them, and generate quality inspection reports related to the lending and collection processes based on the parsing results.

[0082] A device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method as described in any one of the preceding items when executing the computer program.

[0083] A computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the steps of the method described in any one of the preceding items.

[0084] Compared with the prior art, the present invention has the following beneficial effects:

[0085] The intelligent question-answering method based on deep learning and credit loan knowledge graph in the present invention realizes accurate understanding of user queries and knowledge reasoning in professional fields by constructing a two-layer generative adversarial network model, combining natural language processing technology with a multi-dimensional user portrait credit loan knowledge graph. The collaborative work of the first generator and the first discriminator enables the system to quickly convert user queries into structured semantic information after receiving them, and efficiently match and reason with entities and relationships in the knowledge graph. It not only ensures the accuracy and relevance of the preliminary answers, but also provides users with authoritative answers based on professional knowledge. The further intervention of the second generator and the second discriminator adds personalization to the preliminary answers by deeply mining multi-dimensional information such as credit records, consumption behaviors, social relationships, etc. in the user's historical data, making them closer to the user's actual needs and preferences.

[0086] The present invention combines the user's historical credit data, consumption behavior and social data with the intelligent question-answering system. The system can analyze and update the user's portrait in real time, and provide personalized credit assessment and loan recommendations based on the user's real-time conversation content and query, ensuring that it can accurately reflect the user's credit status and behavioral characteristics at any time. Through the dynamic training of the two-layer generative adversarial network, the system can timely adjust the assessment model and recommendation strategy when the user's behavioral data and external environment change, thereby improving the accuracy of credit assessment and the effectiveness of risk management.

[0087] The knowledge graph construction module in the present invention not only supports the management of static data, but also can further enhance the flexibility and adaptability of the system through the online customer service robot module and the quality inspection report generation module, realize the dynamic update of user portraits and knowledge graphs, and ensure that the system can reflect the user's current credit behavior and attitude in real time, and can reflect market changes, policy adjustments and user behavior trends in real time, and provide the system with a continuous source of the latest knowledge resources. And through the quality inspection report, the system's question-and-answer process is optimized to enhance the system's continuous learning ability and business adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0089] Figure 1 This is a flow chart of the intelligent question-answering method based on deep learning and credit loan knowledge graph of the present invention. DETAILED DESCRIPTION

[0090] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0091] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0092] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0093] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0094] See also Figure 1 , is a flow chart of the intelligent question-answering method based on deep learning and credit loan knowledge graph in the present invention, which specifically includes the following steps:

[0095] S1, build a credit loan knowledge graph, and build a knowledge graph containing entities and relationships in the process of drainage, lending, and collection related to the credit loan business.

[0096] S1.1, identify and define the key entities of multi-dimensional user portraits, including user credit data entity E c , User consumption behavior entity E s , User social data entity E n 、Loan Product Entity E p , Repayment Model Entity E r and lending institution entity E l The definition of each entity is determined according to the role of the corresponding entity in the credit loan business:

[0097]

[0098] Among them, E i is the weighted comprehensive representation of entity i, α k is the weight factor, indicating the importance of the kth attribute to the entity, A ik is the kth attribute value of the i-th entity, K i is the total number of attributes involved in entity i.

[0099] S1.2, associate the multi-dimensional information of the user's credit data, consumption behavior and social data with loan products and repayment patterns, and build the relationship between entities, including the adaptability relationship between user credit data and loan products R cp , the matching relationship between user consumption behavior and repayment mode R sr The influence relationship between user social data and credit data R nc :

[0100]

[0101] Among them, R ij Represents entity E i With another entity E j The strength of the relationship between and Entity E i With another entity E j The embedding representation in the mth dimension, β m is the weight factor of the mth dimension, and M is the total number of dimensions.

[0102] S1.3, assign weight parameter W to each entity relationship r , weight parameter W r Used to describe the importance of different data dimensions in user portraits. The weight parameters are obtained through analysis and statistical calculation of historical data:

[0103]

[0104] Among them, γ is the relationship with R i The importance coefficient of the relevant historical data, λ is the adjustment parameter that controls the convergence speed of the relationship weight, and n is the total number of relationships.

[0105] S1.4, in the multi-dimensional user portrait credit loan knowledge graph, each entity is connected through a relationship to construct a user portrait graph structure G(V,E), where V represents the set of entity nodes and E represents the set of relationship edges between entities. The graph structure is represented in the form of a triple as (E a ,R b ,E c ), where E a and E c are entities of different dimensions, R b The relationship between the two is:

[0106]

[0107] Where A is the number of different entity combinations in the graph, δ i is the weight factor of the i-th entity combination, × represents the Cartesian product operation, which ensures that the combination of entities and relationships forms a complete multi-dimensional user portrait credit loan knowledge graph structure. At the same time, the weight of each entity in the graph can be reflected in a cumulative manner.

[0108] S1.5, classify user portraits through clustering algorithms, cluster users’ credit data, consumption behavior and social data into corresponding user groups, and hierarchically process the multi-dimensional user portrait credit loan knowledge graph, so that users can locate the corresponding user groups and corresponding loan products when they query:

[0109]

[0110] Among them, C k is the kth cluster center, is the feature vector of the i-th user in the j-th dimension, is the mean vector of the kth cluster center in the jth dimension, δ ij It is the matching degree between the user and the cluster center in the corresponding dimension.

[0111] S1.6, through data cleaning and ontology matching technology, eliminate redundant and noisy data in the multi-dimensional user portrait credit loan knowledge graph, and perform semantic enhancement on the multi-dimensional user portrait credit loan knowledge graph:

[0112]

[0113] Among them, Sim(R b ) is the semantic similarity of the relationship, η m is the similarity weight factor of the mth dimension, and are the embedding representations of the i-th and j-th entities in the m-th dimension respectively, and ||·|| is the modulus operation of the vector.

[0114] S1.7, ultimately generates a multi-dimensional user portrait credit loan knowledge graph that covers user credit data, consumption behavior and social data, and is associated with loan products and repayment model entities.

[0115] S2, design and train a two-layer generative adversarial network model, which includes a first generator, a second generator, a first discriminator and a second discriminator; the first generator is used to generate a preliminary answer, and the second generator is used to generate a final answer based on the preliminary answer and combined with user historical data; the two discriminators are respectively used to evaluate the accuracy and relevance of the generated answers.

[0116] The two-layer generative adversarial network model includes:

[0117] S2.1, the first generator is used to generate preliminary answers related to user queries. The first generator is trained based on the semantic information of the user query and the entities and relationships in the credit loan knowledge graph.

[0118] S2.2, the first discriminator is used to evaluate the relevance and authenticity of the preliminary answer generated by the first generator, match the preliminary answer with the credit loan knowledge graph, and determine whether it conforms to the logic and relationship in the credit loan knowledge graph.

[0119] S2.3, the second generator generates a final answer based on the preliminary answer generated by the first generator and in combination with the user's historical data and needs.

[0120] S2.4, the second discriminator is used to evaluate the final answer generated by the second generator and determine whether the final answer is consistent with the user's historical data and credit status.

[0121] S3, semantic parsing of user queries, uses a pre-trained natural language processing model to semantically parse the query input by the user, and passes the parsed result as input to the first generator to generate a real-time preliminary answer.

[0122] S4, generation and screening of preliminary answers. The first generator generates real-time preliminary answers based on the parsing results, matches the real-time preliminary answers with the credit loan knowledge graph, and evaluates their relevance through the first discriminator, screening out real-time preliminary answers that are highly relevant to entities and relationships in the knowledge graph, and then passes the screened real-time preliminary answers to the second generator. The generation and screening of preliminary answers specifically includes the following steps:

[0123] S4.1, construct the first generator G1, the first generator G1 receives the semantic information q of the user query and the entities and relationships (E a ,R b ,E c ) as input, and generate a preliminary answer a1 through a multi-layer neural network:

[0124]

[0125] Among them, q is the semantic vector representation of the user query, h i is the historical context vector related to the entities and relations in the multi-dimensional user portrait credit loan knowledge graph, α i is the weight coefficient based on the attention mechanism, and are the embedding representations of entities and relations in the multi-dimensional user portrait credit loan knowledge graph, W h and W o is the weight matrix of the first generator G1, b h and b o is the bias term, represents the vector concatenation operation, || represents the connection operation, ReLU is the activation function, and softmax is the normalization function.

[0126] S4.2, train the first discriminator D1, which is used to evaluate the relevance and authenticity of the preliminary answer a1 generated by the first generator G1, match the preliminary answer a1 with the relationship in the multi-dimensional user portrait credit loan knowledge graph, and generate an evaluation score s1 through the following scoring mechanism:

[0127]

[0128] Among them, W d is the weight matrix of the discriminator, b d is the bias term, σ is the Sigmoid activation function, and are the embedding representations of entities and relations related to the current query in the multi-dimensional user portrait credit loan knowledge graph, γ jis the importance coefficient of the entity relationship, ⊙ represents the element-level product operation, and Z is the normalization factor.

[0129] S5, generation and evaluation of the final answer. The second generator generates a real-time final answer based on the screened real-time preliminary answer, combined with the user's historical credit data, repayment record and loan application history information. The real-time final answer is evaluated by the second discriminator to ensure its accuracy and practicality. The generation and evaluation of the final answer specifically includes the following steps:

[0130] S5.1, construct the second generator G2, which generates the final answer a2 based on the preliminary answer a1 generated by the first generator G1 and combines the user's historical data h and demand d:

[0131]

[0132] Among them, W f is the weight matrix of the second generator, b f is the bias term, β k is the weight coefficient between the user's historical data and demand, GRU is the gated recurrent unit, which is used to fuse the user's historical data h and demand d, and tanh is the hyperbolic tangent activation function.

[0133] S5.2, design and train the second discriminator D2, which is used to evaluate the personalized matching degree of the final answer a2 generated by the second generator G2, match the final answer a2 with the user's historical data h and credit status, and generate the evaluation score s2 through the following formula:

[0134]

[0135] Among them, W g is the weight matrix of the second discriminator, b g is the bias term, δ n is a weight factor that matches the user’s historical data, and BiLSTM is a bidirectional long short-term memory network that is used to analyze the time series correlation of the user’s historical data.

[0136] S5.3, by alternately training the first generator G1, the first discriminator D1, the second generator G2 and the second discriminator D2, the overall performance of the generative adversarial network is optimized so that the generated final answer a2 reaches the optimal value:

[0137]

[0138] in, is the loss function of the generative adversarial network, represents the expected value, p g1 and p g2Probability distribution of generated answers for the first and second generators respectively.

[0139] S6, in the process of lending and collecting payments, uses the telemarketing, consumer protection and collection records in the credit loan knowledge graph to interact with users in real time through an online customer service robot; combined with the answers generated by the two-layer generative adversarial network model, it warns of users' complaint intentions and performs text quality inspection based on the content of the conversation.

[0140] The real-time interaction with the user through the online customer service robot is as follows:

[0141] S6.1, in the process of lending and collecting payments, based on the telemarketing records in the credit loan knowledge graph of multi-dimensional user portraits E tx 、Consumer Protection Records E cb and collection records cs Carry out real-time interaction. Telemarketing records, consumer protection records and collection records respectively represent the user's interactive behavior and related historical data at different stages.

[0142] S6.2, using the final answer a2 combined with the user's real-time conversation content c, the probability distribution P of the user's complaint intention is calculated through the attention mechanism comp .

[0143] S6.3, according to the probability distribution P of complaint intention comp Set the threshold T comp , when P comp ≥T comp When alert , the early warning module generates corresponding early warning signals and sends alarm information to customer service personnel:

[0144]

[0145] Among them, A alert is the warning signal, the value is 1 for triggering the warning, and 0 for not triggering the warning. comp is the preset complaint intention probability threshold.

[0146] The text quality inspection according to the conversation content is specifically as follows:

[0147] S6.4, when performing text quality inspection, the key information in the user conversation is extracted through the text analysis module, and it is compared with the entity E in the multi-dimensional user portrait credit loan knowledge graph. i and the relation R j Compare and generate quality inspection report Q rep :

[0148]

[0149] Among them, W qis the weight matrix of the quality inspection module, b q is the bias term, γ ij For entity E i With the relationship R j The matching degree in the dialogue text, ReLU is the activation function, sum is the weighted sum operation, Q rep The generated text quality inspection report is used to evaluate the quality and compliance of user conversation content.

[0150] S6.5, during the lending and collection process, continuously monitor the interaction records between users and customer service robots, and dynamically update the telemarketing records, consumer protection records, and collection records in the multi-dimensional user portrait credit loan knowledge graph based on the real-time conversation content.

[0151] S7, dialogue text analysis and quality inspection report generation, uses a two-layer generative adversarial network model to analyze the dialogue text in the customer service system, extracts key fields and classifies them, and generates a quality inspection report related to the loan and collection process based on the analysis results. Specifically includes:

[0152] S7.1, through the first generator of the two-layer generative adversarial network model, the conversation text in the customer service system is preliminarily parsed to identify and extract user intentions, question types, related entities and relationships in the conversation.

[0153] S7.2, after extracting the key information, the second generator of the two-layer generative adversarial network model is used to classify the information and map it to specific categories in the lending and collection process.

[0154] S7.3, after the classification is completed, the classification results are evaluated by the second discriminator of the double-layer generative adversarial network model. The system evaluates the degree of match between the results and the entities and relationships in the multi-dimensional user portrait credit loan knowledge graph, verifies the accuracy and relevance of the classification, and ensures that the classification results are consistent with the user's actual credit status, historical behavior data and related business rules.

[0155] S7.4, based on the classification evaluation results, generates a quality inspection report related to the lending and collection process. The quality inspection report summarizes the key data and matching evaluation results involved in the customer service conversation and is used to evaluate the quality of the entire lending and collection process.

[0156] S7.5, after the quality inspection report is generated, the system will feed back the quality inspection report to the business department. At the same time, based on the latest quality inspection report and actual business situation, the system will dynamically update the data in the multi-dimensional user portrait credit loan knowledge graph.

[0157] An embodiment of the present invention is an intelligent question-answering system based on deep learning and credit loan knowledge graph, comprising the following modules:

[0158] A credit loan knowledge graph construction module, which is used to construct a credit loan knowledge graph, and to construct a knowledge graph containing entities and relationships in the process of attracting, lending, and collecting payments related to the credit loan business;

[0159] A two-layer generative adversarial network model design module, which is used to design and train a two-layer generative adversarial network model. The two-layer generative adversarial network model includes a first generator, a second generator, a first discriminator, and a second discriminator. The first generator is used to generate a preliminary answer, and the second generator is used to generate a final answer based on the preliminary answer and combined with user historical data. The two discriminators are used to evaluate the accuracy and relevance of the generated answer respectively.

[0160] A user query semantic parsing module, which is used for semantic parsing of user queries, performs semantic parsing on the query input by the user using a pre-trained natural language processing model, and passes the parsed result as input to the first generator to generate a real-time preliminary answer;

[0161] A preliminary answer generation and screening module, which is used to generate and screen preliminary answers. The first generator generates a real-time preliminary answer based on the parsing result, matches the real-time preliminary answer with the credit loan knowledge graph, and evaluates its relevance through the first discriminator, and screens out the real-time preliminary answer that is highly relevant to the entities and relationships in the knowledge graph, and then passes the screened real-time preliminary answer to the second generator;

[0162] A final answer generation and evaluation module, which is used to generate and evaluate the final answer. The second generator generates a real-time final answer based on the screened real-time preliminary answer, combined with the user's historical credit data, repayment record and loan application history information. The real-time final answer is evaluated by the second discriminator to ensure its accuracy and practicality;

[0163] Real-time interaction and early warning module, which is used to use the telemarketing, consumer protection and collection records in the credit loan knowledge graph in the process of lending and collection, and interact with users in real time through online customer service robots; combined with the answers generated by the two-layer generative adversarial network model, it warns users of their complaint intentions and conducts text quality inspection based on the content of the conversation;

[0164] The dialogue text parsing and quality inspection report generation module is used for parsing dialogue text and generating quality inspection reports. It uses a two-layer generative adversarial network model to parse dialogue texts in the customer service system, extract key fields and classify them, and generate quality inspection reports related to the lending and collection processes based on the parsing results.

[0165] The present invention provides an intelligent question-answering method based on deep learning and credit loan knowledge graph, specifically including identifying the user's credit data E c 、Consumer Behavior s 、Social Data E n And loan product E p 、Repayment Mode E r And other key entities, and generate the representation of each entity through a weighted comprehensive representation formula. Further, this application proposes a relationship strength construction formula Build the relationship strength between entities and improve the accuracy of user portraits. Therefore, first of all, the present invention conducts a comprehensive analysis of the compatibility relationship between the user's credit, consumption behavior, social data and loan products and repayment models, which can achieve efficient data matching. In addition, the protection points of the present invention include: the algorithm formula for constructing user portraits and the multi-dimensional association modeling method of the credit loan knowledge graph.

[0166] This solution proposes a two-layer recurrent generative adversarial network model to generate real-time answers related to user queries. First, the pre-trained natural language processing model is used to semantically parse the user query, and the parsing results are input into the first generator to generate a preliminary answer. The first discriminator is used to evaluate the relevance of the preliminary answer a1, using the following scoring mechanism The second generator combines the user's historical data and needs to generate the final answer a2, and evaluates the personalized matching of the answer through the second discriminator. It is assumed that through this method, the first generator generates a preliminary answer based on the user's query semantic information, and the second generator combines the user's historical data and needs to generate a personalized final answer. The first and second discriminators evaluate the relevance and matching of the generated answers. The proposed two-layer generative adversarial network model shows excellent flexibility and intelligence when dealing with complex query scenarios, and further optimizes the answer generation quality through alternating training of the generative adversarial network. The key protection point of the present invention lies in the design of the two-layer generative adversarial network model and its application in the field of credit loans. This improved scheme protects the design of the two-layer generative adversarial network model, the various algorithms in the generation process, and the related discrimination mechanism.

[0167] When the probability of complaint intention P comp Exceeds the preset threshold T comp When the warning signal A is triggered alert , and generate quality inspection reports to evaluate the quality and compliance of user interactions This part protects the real-time complaint intention analysis and early warning algorithm based on user interaction, as well as the quality inspection report generation mechanism. It is intended to build an online customer service robot system that can perform semantic analysis and complaint intention identification based on the user's real-time conversation content, and trigger the complaint early warning mechanism when necessary. In addition, the interactive customer service system proposed in the present invention is based on multi-dimensional analysis of telemarketing, consumer protection and collection records, and can dynamically update user portraits, generate quality inspection reports and provide real-time feedback to relevant business departments.

[0168] An embodiment of the present invention provides a terminal device. The terminal device of this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps in the above-mentioned method embodiments are implemented. Alternatively, when the processor executes the computer program, the functions of the modules / units in the above-mentioned device embodiments are implemented.

[0169] The computer program may be divided into one or more modules / units, and the one or more modules / units are stored in the memory and executed by the processor to accomplish the present invention.

[0170] The device / terminal equipment may be a computing device such as a desktop computer, a notebook, a palm computer, a cloud server, etc. The device / terminal equipment may include, but is not limited to, a processor and a memory.

[0171] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0172] The memory may be used to store the computer programs and / or modules, and the processor implements various functions of the apparatus / terminal device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory.

[0173] If the module / unit integrated in the device / terminal equipment is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device that can carry the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0174] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An intelligent question-answering method based on deep learning and credit loan knowledge graph, characterized in that: The following steps are involved: Build a credit loan knowledge graph, which includes entities and relationships in the process of attracting customers, lending, and collecting payments related to the credit loan business; Design and train a two-layer generative adversarial network model, which includes a first generator, a second generator, a first discriminator, and a second discriminator; the first generator is used to generate a preliminary answer, and the second generator is used to generate a final answer based on the preliminary answer and combined with user historical data; The two discriminators are used to evaluate the accuracy and relevance of the generated answers respectively; Semantic parsing of user queries: using a pre-trained natural language processing model to semantically parse the query input by the user, and passing the parsed result as input to the first generator to generate a real-time preliminary answer; Generation and screening of preliminary answers: The first generator generates real-time preliminary answers based on the parsing results, matches the real-time preliminary answers with the credit loan knowledge graph, and evaluates their relevance through the first discriminator, screening out real-time preliminary answers that are highly relevant to entities and relationships in the knowledge graph, and then passes the screened real-time preliminary answers to the second generator; Generation and evaluation of the final answer. The second generator generates a real-time final answer based on the screened real-time preliminary answer, combined with the user's historical credit data, repayment record, and loan application history information. The real-time final answer is evaluated by the second discriminator to ensure its accuracy and practicality. In the process of lending and collecting payments, the telemarketing, consumer protection and collection records in the credit loan knowledge graph are used to interact with users in real time through online customer service robots. The answers generated by the two-layer generative adversarial network model are combined to warn users of their complaint intentions and conduct text quality inspection based on the content of the conversation. Conversation text analysis and quality inspection report generation: Use a two-layer generative adversarial network model to parse the conversation text in the customer service system, extract key fields and classify them, and generate quality inspection reports related to the lending and collection processes based on the analysis results.

2. The intelligent question-answering method based on deep learning and credit loan knowledge graph according to claim 1, characterized in that: The construction of the credit loan knowledge graph specifically includes the following steps: Identify and define the key entities of multi-dimensional user portraits, including user credit data entity E c , User consumption behavior entity E s , User social data entity E n 、Loan Product Entity E p , Repayment Model Entity E r and lending institution entity E l The definition of each entity is determined according to the role of the corresponding entity in the credit loan business: Among them, E i is the weighted comprehensive representation of entity i, α k is the weight factor, indicating the importance of the kth attribute to the entity, A ik is the kth attribute value of the i-th entity, K i is the total number of attributes involved in entity i; The multi-dimensional information of the user's credit data, consumption behavior and social data is associated with the loan products and repayment patterns to build the relationship between entities, including the adaptability relationship between the user's credit data and the loan products. cp , the matching relationship between user consumption behavior and repayment mode R sr The influence relationship between user social data and credit data R nc : Among them, R ij Represents entity E i With another entity E j The strength of the relationship between and Entity E i With another entity E j The embedding representation in the mth dimension, β m is the weight factor of the mth dimension, and M is the total number of dimensions; Assign a weight parameter W to each entity relationship r , weight parameter W r Used to describe the importance of different data dimensions in user portraits. The weight parameters are obtained through analysis and statistical calculation of historical data: Among them, γ is the relationship with R i The importance coefficient of the relevant historical data, λ is the adjustment parameter to control the convergence speed of the relationship weight, and n is the total number of relationships; In the multi-dimensional user portrait credit loan knowledge graph, each entity is connected through a relationship to construct a user portrait graph structure G(V,E), where V represents the set of entity nodes and E represents the set of relationship edges between entities. The graph structure is represented in the form of a triple as (E a ,R b ,E c ), where E a and E c are entities of different dimensions, R b The relationship between the two is: Where A is the number of different entity combinations in the graph, δ i is the weight factor of the ith entity combination, and × represents the Cartesian product operation, which ensures that the combination of entities and relationships forms a complete multi-dimensional user portrait credit loan knowledge graph structure, and the weight of each entity in the graph can be reflected by accumulation; The user portraits are classified and processed through clustering algorithms, and the user's credit data, consumption behavior and social data are clustered into corresponding user groups. The multi-dimensional user portrait credit loan knowledge graph is hierarchically processed so that users can locate the corresponding user groups and corresponding loan products when they query: Among them, C k is the kth cluster center, is the feature vector of the i-th user in the j-th dimension, is the mean vector of the kth cluster center in the jth dimension, δ ij It is the matching degree between the user and the cluster center in the corresponding dimension; Through data cleaning and ontology matching technology, redundant and noisy data in the multi-dimensional user portrait credit loan knowledge graph is eliminated, and the multi-dimensional user portrait credit loan knowledge graph is semantically enhanced: Among them, Sim(R b ) is the semantic similarity of the relationship, η m is the similarity weight factor of the mth dimension, and are the embedding representations of the i-th and j-th entities in the m-th dimension, respectively, and ||·|| is the modulus length operation of the vector; Ultimately, a multi-dimensional user portrait credit loan knowledge graph is generated that covers user credit data, consumption behavior and social data, and is associated with loan products and repayment model entities.

3. The intelligent question-answering method based on deep learning and credit loan knowledge graph according to claim 1, characterized in that: The two-layer generative adversarial network model includes: The first generator is used to generate preliminary answers related to the user's query. The first generator is trained based on the semantic information of the user's query and the entities and relationships in the credit loan knowledge graph; The first discriminator is used to evaluate the relevance and authenticity of the preliminary answer generated by the first generator, match the preliminary answer with the credit loan knowledge graph, and determine whether it conforms to the logic and relationship in the credit loan knowledge graph; The second generator generates a final answer based on the preliminary answer generated by the first generator and in combination with the user's historical data and needs; The second discriminator is used to evaluate the final answer generated by the second generator and determine whether the final answer is consistent with the user's historical data and credit status.

4. The intelligent question-answering method based on deep learning and credit loan knowledge graph according to claim 1, characterized in that: The generation and screening of the preliminary answers specifically include the following steps: Construct the first generator G1, which receives the semantic information q of the user query and the entities and relationships (E a ,R b ,E c ) as input, and generate a preliminary answer a1 through a multi-layer neural network: Among them, q is the semantic vector representation of the user query, h i is the historical context vector related to the entities and relations in the multi-dimensional user portrait credit loan knowledge graph, α i is the weight coefficient based on the attention mechanism, and are the embedding representations of entities and relations in the multi-dimensional user portrait credit loan knowledge graph, W h and W o is the weight matrix of the first generator G1, b h and b o is the bias term, represents the vector concatenation operation, || represents the connection operation, ReLU is the activation function, and softmax is the normalization function; The first discriminator D1 is trained to evaluate the relevance and authenticity of the preliminary answer a1 generated by the first generator G1, match the preliminary answer a1 with the relationship in the multi-dimensional user portrait credit loan knowledge graph, and generate an evaluation score s1 through the following scoring mechanism: Among them, W d is the weight matrix of the discriminator, b d is the bias term, σ is the Sigmoid activation function, and are the embedding representations of entities and relations related to the current query in the multi-dimensional user portrait credit loan knowledge graph, γ j is the importance coefficient of the entity relationship, ⊙ represents the element-level product operation, and Z is the normalization factor.

5. The intelligent question-answering method based on deep learning and credit loan knowledge graph according to claim 1, characterized in that: The generation and evaluation of the final answer specifically includes the following steps: Construct the second generator G2, which is based on the preliminary answer a1 generated by the first generator G1 and combines the user's historical data h and demand d to generate the final answer a2: Among them, W f is the weight matrix of the second generator, b f is the bias term, β k is the weight coefficient between the user's historical data and demand, GRU is the gated recurrent unit, which is used to fuse the user's historical data h and demand d, and tanh is the hyperbolic tangent activation function; Design and train the second discriminator D2, which is used to evaluate the personalized matching degree of the final answer a2 generated by the second generator G2, match the final answer a2 with the user's historical data h and credit status, and generate the evaluation score s2 through the following formula: Among them, W g is the weight matrix of the second discriminator, b g is the bias term, δ n is the weight factor that matches the user's historical data, and BiLSTM is a bidirectional long short-term memory network used to analyze the time series correlation of the user's historical data; By alternately training the first generator G1, the first discriminator D1, the second generator G2, and the second discriminator D2, the overall performance of the generative adversarial network is optimized so that the generated final answer a2 is optimal: in, is the loss function of the generative adversarial network, represents the expected value, p g1 and p g2 Probability distribution of generated answers for the first and second generators respectively.

6. The intelligent question-answering method based on deep learning and credit loan knowledge graph according to claim 1, characterized in that: The real-time interaction with the user through the online customer service robot is as follows: In the process of lending and collecting payments, the telemarketing records in the credit loan knowledge graph based on multi-dimensional user portraits are tx 、Consumer Protection Records E cb and collection records cs Real-time interaction: Telemarketing records, consumer protection records, and collection records represent the user's interactive behaviors and related historical data at different stages. Using the final answer a2 and the user’s real-time conversation content c, the probability distribution P of the user’s complaint intention is calculated through the attention mechanism. comp ; According to the probability distribution P of complaint intention comp Set the threshold T comp , when P comp ≥T comp When alert , the early warning module generates corresponding early warning signals and sends alarm information to customer service personnel: Among them, A alert is the warning signal, the value is 1 for triggering the warning, and 0 for not triggering the warning. comp is the preset probability threshold of complaint intention; The text quality inspection according to the conversation content is specifically as follows: When performing text quality inspection, the text analysis module is used to extract key information from user conversations and compare it with entity E in the multi-dimensional user portrait credit loan knowledge graph. i and the relation R j Compare and generate quality inspection report Q rep : Among them, W q is the weight matrix of the quality inspection module, b q is the bias term, γ ij For entity E i With the relationship R j The matching degree in the dialogue text, ReLU is the activation function, sum is the weighted sum operation, Q rep Generated text quality inspection reports are used to evaluate the quality and compliance of user conversation content; During the lending and collection process, the interaction records between users and customer service robots are continuously monitored, and the telemarketing records, consumer protection records and collection records in the multi-dimensional user portrait credit loan knowledge graph are dynamically updated based on the real-time conversation content.

7. The intelligent question-answering method based on deep learning and credit loan knowledge graph according to claim 1, characterized in that: The analysis of the dialogue text and the generation of a quality inspection report specifically include: The first generator of the two-layer generative adversarial network model performs preliminary analysis on the conversation text in the customer service system to identify and extract user intent, question type, related entities and relationships in the conversation; After extracting key information, the second generator of the two-layer generative adversarial network model is used to classify the information and map it to specific categories in the lending and collection process; After the classification is completed, the classification results are evaluated by the second discriminator of the two-layer generative adversarial network model. The system evaluates the matching degree between the results and the entities and relationships in the multi-dimensional user portrait credit loan knowledge graph, verifies the accuracy and relevance of the classification, and makes the classification results consistent with the user's actual credit status, historical behavior data and related business rules; Based on the classification evaluation results, a quality inspection report related to the lending and collection process is generated. The quality inspection report summarizes the key data and matching evaluation results involved in the customer service dialogue and is used to evaluate the quality of the entire lending and collection process; After the quality inspection report is generated, the system will feed back the quality inspection report to the business department. At the same time, based on the latest quality inspection report and actual business conditions, it will dynamically update the data in the multi-dimensional user portrait credit loan knowledge graph.

8. An intelligent question-answering system based on deep learning and credit loan knowledge graph, characterized by: Includes the following modules: A credit loan knowledge graph construction module, which is used to construct a credit loan knowledge graph, and to construct a knowledge graph containing entities and relationships in the process of attracting, lending, and collecting payments related to the credit loan business; A two-layer generative adversarial network model design module, which is used to design and train a two-layer generative adversarial network model. The two-layer generative adversarial network model includes a first generator, a second generator, a first discriminator, and a second discriminator. The first generator is used to generate a preliminary answer, and the second generator is used to generate a final answer based on the preliminary answer and combined with user historical data. The two discriminators are used to evaluate the accuracy and relevance of the generated answer respectively. A user query semantic parsing module, which is used for semantic parsing of user queries, performs semantic parsing on the query input by the user using a pre-trained natural language processing model, and passes the parsed result as input to the first generator to generate a real-time preliminary answer; A preliminary answer generation and screening module, which is used to generate and screen preliminary answers. The first generator generates a real-time preliminary answer based on the parsing result, matches the real-time preliminary answer with the credit loan knowledge graph, and evaluates its relevance through the first discriminator, and screens out the real-time preliminary answer that is highly relevant to the entities and relationships in the knowledge graph, and then passes the screened real-time preliminary answer to the second generator; A final answer generation and evaluation module, which is used to generate and evaluate the final answer. The second generator generates a real-time final answer based on the screened real-time preliminary answer, combined with the user's historical credit data, repayment record and loan application history information. The real-time final answer is evaluated by the second discriminator to ensure its accuracy and practicality; Real-time interaction and early warning module, which is used to use the telemarketing, consumer protection and collection records in the credit loan knowledge graph in the process of lending and collection, and interact with users in real time through online customer service robots; combined with the answers generated by the two-layer generative adversarial network model, it warns users of their complaint intentions and conducts text quality inspection based on the content of the conversation; The dialogue text parsing and quality inspection report generation module is used for parsing dialogue text and generating quality inspection reports. It uses a two-layer generative adversarial network model to parse dialogue texts in the customer service system, extract key fields and classify them, and generate quality inspection reports related to the lending and collection processes based on the parsing results.

9. A device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images