Intelligent customer service dialogue model construction method and application

Abstract

The application discloses a kind of intelligent customer service dialogue model construction method and application, it is related to artificial intelligence technical field;Including intention recognition module, knowledge graph management module, multi-modal interaction module and reply generation optimization module;Among them:The intention recognition module judges the intention category of customer consultation;Adopt the natural language processing technology based on deep learning, in combination with word vector representation and bidirectional long short-term memory network, the text input by customer is carried out semantic analysis;Through training by labeled corpus, the characteristic expression of different business intention is learned autonomously, so that the question of customer is accurately classified into predefined intention category system.The intelligent customer service template designed in the application covers multiple personality dimensions, and the score can be dynamically adjusted according to customer feedback, and then the template most suitable for the individual needs of the customer is recommended;This can enhance the interactive experience of customer and customer service, and improve the customer's sense of identity to service.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a smart customer service dialogue model construction method and application. BACKGROUND

[0002] In the current digital era, with the continuous development and application of artificial intelligence technology, intelligent customer service systems have become one of the important means for enterprises to provide customer service. Traditional customer service mode mainly relies on manual service, but this way has problems such as low efficiency and high cost. In order to solve these problems, intelligent customer service systems have emerged, which realize automatic customer service through natural language processing, machine learning and other technologies, and improve service efficiency and customer satisfaction.

[0003] However, the existing intelligent customer service system still has some limitations. First, the accuracy of intent recognition needs to be improved. Although natural language processing technology based on deep learning has made some progress, in actual application, due to the complexity and diversity of customer inquiries, the accuracy of intent recognition is still not high. Second, the quality of reply generation needs to be improved. Most of the current intelligent customer service systems use template-based reply method, which lacks personalization and pertinence, and cannot meet the needs of different customers.

[0004] Through retrieval, the application scheme of Chinese patent application No. CN202411261151.2 discloses an implementation method and device of intelligent customer service, specifically relates to the technical field of intelligent customer service, and includes an implementation data collection module of intelligent customer service, an implementation data analysis module of intelligent customer service, an implementation intelligent evaluation module of intelligent customer service, an implementation intelligent comprehensive analysis module of intelligent customer service, and an implementation intelligent human-computer interaction module of intelligent customer service. The implementation data collection module of intelligent customer service is used for collecting intelligent customer service implementation intelligent performance data, obtaining various monitoring parameters for evaluating intelligent performance, and transmitting the various monitoring parameters to the implementation data analysis module of intelligent customer service. The scheme in the above patent has the following deficiencies: it cannot provide more suitable personalized experience according to different users, and needs to be improved. SUMMARY

[0005] The purpose of the present application is to solve the problems existing in the prior art and provide an intelligent customer service dialogue model construction method and application.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0007] An intelligent customer service dialogue model construction method includes an intent recognition module, a knowledge graph management module, a multi-modal interaction module, and a reply generation optimization module.

[0008] Among them:

[0009] The intention recognition module judges the intention category of customer consultation; adopts natural language processing technology based on deep learning, combines word vector representation and bidirectional long short-term memory network, and performs semantic analysis on the text input by the customer; the characteristics of different business intentions are learned autonomously by training through labeled corpus, so that the customer's question is accurately classified into the pre-defined intention category system;

[0010] The reply generation optimization module comprehensively considers the intention recognition result and the knowledge graph query information, adopts a hybrid generation strategy based on the combination of templates and deep learning; for different intention categories, multiple intelligent customer service templates are designed in advance, each template contains multiple personality tags and their corresponding scores, and the scores represent the degree of strength of the personality characteristics in the template; according to the customer's question content and past interaction records, the most suitable intelligent customer service template is dynamically selected for reply generation; on the basis of the template, RNN, LSTM or GRU combined with attention mechanism are used to optimize the reply content; at the same time, according to the feedback evaluation of the customer to the reply, such as satisfaction score, follow-up question situation, etc., a reinforcement learning algorithm is used to adjust and optimize the reply generation strategy, and gradually improve the accuracy and individuality of the reply.

[0011] Preferably, the knowledge graph management module extracts entity, relationship and attribute information from multiple source heterogeneous data in the enterprise, including product information, service terms, frequently asked questions, historical customer cases; uses graph database technology to store and organize these knowledge elements, forming a structured knowledge network; the model is based on the hierarchical relationship and semantic association rules between defined entities, and quickly retrieves and reasons in the knowledge system, providing information basis for subsequent reply generation;

[0012] The multi-modal interaction module adopts a speech recognition engine in voice interaction to accurately convert the customer's voice instructions into text form; at the same time, it uses speech synthesis technology to play the text reply generated by the model to the customer in a clear and natural voice, realizing a seamless voice interaction process; in the image interaction scene, the model receives and analyzes the picture or video information provided by the customer, combines the related knowledge in the knowledge graph, generates analysis text, and finally feeds back the corresponding image information to the customer.

[0013] Preferably, the intention recognition module includes:

[0014] The text preprocessing unit cleans the original text input by the customer, removes irrelevant characters, links, special symbol noise information; uses lexical analysis technology to perform word segmentation processing on the text, and splits the sentence into word units with actual semantics;

[0015] The feature extraction unit uses a word vector representation technology to map each word in the segmented text into a low-dimensional vector space, capturing the semantic information and context relationship of the word. At the same time, the syntactic structure features of the text are extracted, such as the syntactic dependency relationship of the sentence, the subject-predicate-object structure, etc., which are represented by a syntactic analysis tree.

[0016] The intent classification unit constructs an intent classification model based on a deep learning algorithm, using CNN, RNN or LSTM, GRU. The extracted word vector features and syntactic structure features are used as input to the model, which is trained to automatically determine the intent category of the customer's inquiry and output the corresponding probability distribution.

[0017] Preferably, the knowledge graph management module includes:

[0018] The knowledge extraction unit extracts knowledge entities, relationships and attribute information from various structured and unstructured data sources within the enterprise. For structured data, SQL queries or data scraping tools are used to obtain relevant data. For unstructured data, named entity recognition and relationship extraction techniques in natural language processing are used to identify entities and their relationships in the text. The extracted knowledge is cleaned and standardized to remove duplicates and errors, and the format and specifications of knowledge representation in different data sources are unified to ensure the quality and consistency of the knowledge.

[0019] The knowledge fusion unit uses entity linking technology to merge and remove duplicate entities from different data sources, establishing equivalence relationships between entities. Based on ontology and semantic rules, the knowledge is fused and integrated to build a complete knowledge network.

[0020] The knowledge storage unit stores the constructed knowledge graph based on a graph database management system. Based on the index structure in the graph database, knowledge queries are performed.

[0021] Preferably, the multi-modal interaction module includes:

[0022] The voice interaction unit uses automatic speech recognition technology to convert the customer's input voice signal into text information. The text-to-speech technology is used to convert the text reply generated by the intelligent customer service model into voice output to the customer.

[0023] The image interaction unit receives and parses the image file uploaded by the customer, extracting key features and information from the image. Based on computer vision technology and image recognition algorithms, the parsed image information is understood and analyzed, and combined with relevant knowledge in the knowledge graph, the reply and feedback information for the image content are generated.

[0024] Preferably, the reply generation optimization module adjusts the personality score formula based on customer feedback, assuming that the initial score of a certain intelligent customer service template on a certain personality dimension is S init , the impact factor of customer feedback on this dimension is ΔS, a positive value represents positive feedback, and a negative value represents negative feedback, then the adjusted score S adj is expressed as:

[0025] S adj =S init +a×ΔS

[0026] Wherein:

[0027] S adj is the adjusted score;

[0028] S init is the initial score;

[0029] a is the adjustment coefficient, used to control the degree of influence of customer feedback on the score;

[0030] ΔS is the change in score caused by customer feedback;

[0031] The total score of the template is calculated by combining the scores of each personality dimension, assuming that the intelligent customer service template has n personality dimensions, and the score of the i-th personality dimension is S i , then the comprehensive total score S total of the template is expressed as:

[0032]

[0033] Wherein:

[0034] S total is the comprehensive total score of the template;

[0035] n is the number of personality dimensions;

[0036] S i is the score of the i-th personality dimension.

[0037] Preferably, the reply generation optimization module comprises:

[0038] A template matching unit finds a matching reply template in a pre-constructed intelligent customer service template library according to the intent recognition result. These templates are classified and arranged according to different business intents and scenarios, covering standardized answers to common questions and common scenarios; at the same time, the found templates are adaptively adjusted and modified, according to the specific problems and context information of the customer, to fill or replace the variable part of the template, so that the reply is more suitable for the actual situation of the customer;

[0039] The semantic generation unit constructs a semantic generation model based on a deep learning recurrent neural network, trains the model through learning and training of artificial customer service reply corpus, and enables the model to autonomously learn the rules and patterns of text generation; when generating a reply, the model queries relevant information from the knowledge graph according to the intention recognition result and the current context, analyzes the personality preference and communication style of the customer by using a machine learning algorithm, combines personality feature vectors of various templates in the intelligent customer service template library, and recommends an intelligent customer service template that best meets the individual needs of the customer as the dominant reply style for this interaction by calculating a similarity or matching degree index, and generates a reply text.

[0040] Preferably, the reply generation optimization module further comprises:

[0041] The sentiment integration unit analyzes the sentiment tendency in the customer consultation text, judges whether the customer is happy, dissatisfied, anxious or in other emotional states, and appropriately integrates sentiment factors in the reply content according to the sentiment analysis result, so that the reply is closer to the emotional needs of the customer.

[0042] The quality evaluation and optimization unit evaluates the quality of the generated reply content, and optimizes and adjusts the reply generation model according to the quality evaluation result.

[0043] Preferably, the reply generation optimization module predicts the probability of the customer in a certain personality preference dimension according to the historical interaction record X hist , personal information X info , and current consultation question type and context environment X context , and represents the probability as y, y = 1 represents like, and y = 0 represents dislike, and the formula of the logistic regression model is as follows:

[0044]

[0045] Wherein:

[0046] p(y = 1 | X hist , X info , X context ) is the probability of the customer liking a certain personality preference under the given historical interaction, personal information and context environment;

[0047] z = β0 + β1X hist + β2X info + β3X context , wherein β0 is an intercept term, β1, β2 and β3 are regression coefficients corresponding to the features.

[0048] An application of an intelligent customer service dialogue model, which is applied in customer service of an e-commerce platform, e-commerce live broadcast, customer service center of a financial institution, smart home device, and customer service system of a telecom operator.

[0049] The present application has the following advantages:

[0050] 1. The intelligent customer service template designed in the present application covers multiple personality dimensions, and can dynamically adjust the score according to customer feedback, and then recommend the template that best meets the individual needs of the customer; this can enhance the interactive experience of customers and customer service, and improve the customer's sense of identity to the service.

[0051] 2. The template matching and semantic generation combined reply generation module of the present application can quickly respond to common problems based on templates, and can generate personalized replies with logic and coherence based on deep learning; through the emotion integration unit, the reply can also be adjusted according to the customer's emotions, improving customer satisfaction.

[0052] 3. The intent recognition module of the present application combines text preprocessing, feature extraction, and deep learning algorithms, which can effectively deal with customer's diversified expression, accurately judge the consultation intent category, and provide a solid foundation for subsequent replies; for example, in the e-commerce customer service scenario, it can quickly distinguish between product information inquiry, order processing, or after-sales complaint, etc. intent, and improve service efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 A flowchart of an intelligent customer service dialogue model is proposed in the present application. DETAILED DESCRIPTION

[0054] The technical solutions of the present application will be further described in detail below in conjunction with specific embodiments.

[0055] Example 1:

[0056] An intelligent customer service dialogue model construction method, the model mainly consists of an intent recognition module, a knowledge graph management module, a multi-modal interaction module, and a reply generation optimization module, each module cooperates with each other, and together completes efficient communication with customers;

[0057] Among them:

[0058] The intent recognition module judges the intent category of customer consultation; natural language processing technology based on deep learning is adopted, combined with word vector representation and bidirectional long short-term memory network (Bi-LSTM), to perform semantic analysis on the text input by the customer; through training of labeled corpus, the characteristics of different business intents are learned autonomously, so that the customer's question can be accurately classified into a predefined intent category system. For example, in the e-commerce customer service scenario, common intent types such as product information consultation, order inquiry, return and exchange processing, and complaint suggestions can be distinguished.

[0059] wherein the intention recognition module performs semantic analysis based on deep learning natural language processing technology combined with a bidirectional long short-term memory network, and is realized based on a CBOW word vector training model in Word2Vec, and a specific formula is as follows:

[0060]

[0061] wherein J(θ) is an objective function for maximizing the probability of predicting the central word based on the context;

[0062] θ is a parameter of the model;

[0063] N is the number of training samples;

[0064] w i is an input training word;

[0065] P(Context(w i )│w i ) is the probability of generating the context Context(w i ) of the central word w i ) given the central word w t By optimizing this objective function, the word vector representation of each word is trained, and these word vectors can capture the semantic information of the word, providing a basis for subsequent semantic analysis;

[0066] Bidirectional long short-term memory network, taking single time step calculation as an example, the formula is as follows:

[0067] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0068] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0069]

[0070] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0071] t t =o t *tanh(C t )

[0072] wherein:

[0073] f t , i t , o t respectively represent the output of the forget gate, the input gate and the output gate;

[0074] σ is an activation function;

[0075] W f , W i , W C , W o are weight matrices of the forget gate, the input gate, the candidate memory cell and the output gate respectively;

[0076] b f , b i , b C , b o are corresponding bias terms;

[0077] b t-1 is the hidden state of the previous time step;

[0078] x t is the input of the current time step;

[0079] C t is the memory cell state of the current time step;

[0080] h t is the hidden state output of the current time step.

[0081] The intent recognition module comprises:

[0082] A text preprocessing unit, which cleans the original text input by the customer, removes irrelevant characters, links, special symbol noise information. The text is processed by using morphological analysis technology, and the sentence is split into word units with actual semantics. For example, for "I want to know how the camera function of this mobile phone is", after the word segmentation processing, the word sequence "I / want / to know / this mobile phone / camera function / how" is obtained.

[0083] The feature extraction unit uses word vector representation technology to map each word in the segmented text to a low-dimensional vector space, capturing the semantic information and contextual relationships of the words. For example, the word vectors obtained by training the Word2Vec model can reflect the similarity and semantic association between words, such as the distance between "mobile phone" and "mobile phone" in the vector space is closer. At the same time, the syntactic structure features of the text are extracted, such as the syntactic dependency relationship of the sentence, the subject-predicate-object structure, etc., which are represented by the syntactic analysis tree. For example, for "I bought a piece of clothing in the mall yesterday", it can be determined that "I" is the subject, "buy" is the predicate, "clothes" is the object, "yesterday" is the time adverbial, and "in the mall" is the place adverbial.

[0084] The intent classification unit constructs an intent classification model based on deep learning algorithms, using CNN, RNN or LSTM, GRU; the extracted word vector features and syntactic structure features are used as the input of the model, and through the training of the model, the system can automatically determine the intent category of the customer's consultation and output the corresponding probability distribution. For example, for different types of problems input by the customer, such as order inquiry, product consultation, complaint suggestion, etc., the model outputs the probability value belonging to the corresponding intent category, such as the probability of order inquiry intent is 0.8, the probability of product consultation intent is 0.6, etc., and finally determines the main intent category according to the probability size.

[0085] Among them:

[0086] The knowledge graph management module extracts entity, relationship and attribute information from multiple source heterogeneous data in the enterprise, including product information, service terms, frequently asked questions, historical customer cases, etc. Use graph database technology to store and organize these knowledge elements to form a structured knowledge network. The model is based on the hierarchical relationship and semantic association rules defined between entities, and can quickly retrieve and reason in the knowledge system to provide information basis for subsequent reply generation. Taking the financial customer service field as an example, the knowledge graph covers the characteristics of various financial products, rate structure, risk level assessment and related policies and regulations, etc. knowledge, which can effectively support the customer service to answer professional questions.

[0087] The knowledge graph management module includes:

[0088] Knowledge extraction unit, from various structured data sources (such as database, data tables in business systems, etc.) and unstructured data sources (such as product documents, technical manuals, web text, etc.) inside the enterprise, extract knowledge entities, relationships and attribute information. For structured data, through writing SQL query statements or using data grabbing tools to obtain related data; for unstructured data, it is necessary to use named entity recognition (NER) and relationship extraction (RE) techniques in natural language processing technology to identify entities (such as product name, customer name, business terminology, etc.) and their relationships (such as product and customer purchase relationship, product and product accessory composition relationship, etc.) in the text. Clean and standardize the extracted knowledge, remove duplicate and incorrect information, and unify the format and specifications of knowledge representation in different data sources to ensure the quality and consistency of knowledge. For example, unify the name expressions of the same product in different documents to have a unique identifier in the knowledge graph.

[0089] Knowledge fusion unit, using entity linking technology, merging and deduplicating the same or similar entities from different data sources, establishing the equivalence relationship between entities. For example, the same product mentioned in different channels may have different numbers or name expressions, which can be identified as the same entity object through entity linking. Based on ontology and semantic rules, integrate and integrate knowledge to build a complete knowledge network. Define the domain and value domain of entity attributes and relationships to ensure the logical consistency and correctness of knowledge. For example, the attributes of product entities include name, model number, price, production date, etc., and the value range of these attributes should meet the actual business rules and data type requirements.

[0090] Knowledge storage unit, based on graph database management system (such as Neo4j, OrientDB, etc.) to store the constructed knowledge graph. Based on the index structure in the graph database, knowledge query is carried out. For example, index is established for the key attributes of knowledge entities (such as product name, customer ID, etc.) to quickly locate and retrieve related knowledge information. At the same time, the knowledge graph is stored and managed in partitions to adapt to the storage and processing needs of large-scale data.

[0091] In terms of voice interaction, the multi-modal interaction module employs a speech recognition engine to accurately convert the customer's voice instructions into text form. Meanwhile, it utilizes speech synthesis technology to play the text responses generated by the model in clear and natural voice to the customer, realizing a seamless voice interaction process. In image interaction scenarios, such as when the customer inquires about the appearance of a product or seeks operation demonstration, the model receives and analyzes the image or video information provided by the customer, combines relevant knowledge from the knowledge graph, generates analysis text, and finally feeds back the corresponding image information to the customer. In addition, through the integration of natural language processing technology with image and voice information, the model can better understand the customer's comprehensive intentions, such as more accurately determining the problem from the product fault photos and voice descriptions taken by the customer and providing targeted solutions.

[0092] The multi-modal interaction module includes:

[0093] A voice interaction unit employs automatic speech recognition (ASR) technology to convert the customer's input voice signals into text information. It utilizes text-to-speech (TTS) technology to convert the text responses generated by the intelligent customer service model into voice output for the customer. During the voice synthesis process, different voice tones, pitches, and speeds can be selected according to the customer's needs to provide personalized voice services. For example, it can provide the customer with a soft and sweet female voice or a steady and atmospheric male voice, and adjust the volume and speed of the voice to meet the customer's auditory needs in different scenarios.

[0094] An image interaction unit receives and parses the image files uploaded by the customer, extracts the key features and information in the image. Based on computer vision technology and image recognition algorithms, it understands and analyzes the parsed image information, combines relevant knowledge from the knowledge graph, and generates responses and feedback information targeting the image content. For example, if the customer uploads a picture of a math problem, the image understanding and feedback sub-unit can use optical character recognition (OCR) technology to extract the problem content and provide detailed problem-solving ideas and answer guidance for the customer based on the mathematical knowledge and problem-solving methods in the knowledge graph.

[0095] Wherein:

[0096] The reply generation optimization module comprehensively considers the intention recognition result and the knowledge graph query information, and adopts a hybrid generation strategy based on the combination of templates and deep learning. For different intention categories, multiple intelligent customer service templates are designed in advance, each template including multiple personality labels and their corresponding scores, which represent the degree of the personality characteristics in the template. For example, intelligent customer service personality template A has the personality labels of lively (60 points), humorous (20 points), and vivid (20 points); intelligent customer service personality template B has the personality labels of rational (50 points), mature (30 points), and gentle (20 points). In actual application, the most suitable intelligent customer service template is dynamically selected for reply generation according to the customer's question content and past interaction records. If the customer's question is about product usage skills and the customer group is young, personality template A may be selected for reply to provide information in a more lively and vivid manner, increasing the customer's acceptance and satisfaction; for more serious questions involving financial investment risk assessment, personality template B is preferred to convey professional knowledge and advice to the customer in a rational and mature style. On the basis of the template, RNN, LSTM or GRU combined with an attention mechanism is used to optimize the reply content. Through learning a large amount of high-quality artificial customer service reply corpus, the model can automatically learn the language expression habits and information organization methods under different intentions, generating more natural, fluent, accurate and detailed reply texts. At the same time, according to the customer's feedback evaluation of the reply, such as satisfaction score and follow-up questions, a reinforcement learning algorithm is used to adjust and optimize the reply generation strategy, gradually improving the accuracy and individualization of the reply.

[0097] If the customer shows high satisfaction and positive emotional feedback to a template reply with a lively personality label, and considers that the reply style meets his / her expectations, the lively personality score of the template will be appropriately increased; on the contrary, if the customer indicates that the reply is too lively and not professional or accurate enough, the lively personality score will be reduced accordingly, and the scores of other personality dimensions such as rationality or professionalism may be increased. This dynamic adjustment mechanism based on customer feedback enables the intelligent customer service template to continuously adapt to the needs and preferences of different customers, gradually optimizing the reply style and personality performance.

[0098] The reply generation optimization module adjusts the personality score formula based on customer feedback. Let the initial score of a certain intelligent customer service template in a certain personality dimension (such as lively) be S init , and the impact factor of the customer feedback on this dimension be ΔS (positive value indicating positive feedback, negative value indicating negative feedback), then the adjusted score S adj is represented as:

[0099] S adj =S init +a×ΔS

[0100] wherein:

[0101] S adj is the adjusted score;

[0102] S init is the initial score;

[0103] a is an adjustment coefficient for controlling the degree of influence of customer feedback on the score (the value range can be set according to specific circumstances, for example, between 0 and 1);

[0104] AS is the change in score caused by customer feedback;

[0105] The total score of the template is calculated by integrating the scores of each personality dimension. Assuming that the intelligent customer service template has n personality dimensions, and the score of the ith personality dimension is Si, then the comprehensive total score S of the template is: i total

[0106]

[0107] Wherein:

[0108] S total is the comprehensive total score of the template;

[0109] n is the number of personality dimensions;

[0110] S i is the score of the ith personality dimension.

[0111] The reply generation optimization module comprises:

[0112] A template matching unit finds a matching reply template in a pre-constructed intelligent customer service template library according to the intent recognition result. These templates are classified and arranged according to different business intents and scenarios, covering standardized answers in common questions and common scenarios. For example, for the question of customers asking about the order status, there is a corresponding order query template; for the question of customers inquiring about product functions, there is a corresponding product function introduction template. Each template contains rich text content and reply examples, which can quickly respond to common needs of customers. At the same time, the found template is adaptively adjusted and modified, and according to the specific problem and context information of the customer, the variable part in the template is filled or replaced, so that the reply is more suitable for the actual situation of the customer. For example, in the order query template, according to the order number and other information provided by the customer, the specific order status reply content is dynamically generated.

[0113] ​​The semantic generation unit constructs a semantic generation model based on a deep learning recurrent neural network, trains the model through learning and training of artificial customer service reply corpus, and makes the model autonomously learn the rules and patterns of text generation. When generating a reply, the model uses machine learning algorithms to model and analyze the customer's personality preferences and communication style based on the intent recognition results and relevant information queried from the knowledge graph, in combination with the current context. Then, by calculating the similarity or matching degree index, the most suitable intelligent customer service template for the customer's individual needs is recommended as the dominant reply style for this interaction from the personality feature vectors of each template in the intelligent customer service template library, and a coherent, accurate and logically consistent reply text is generated. For example, when a customer raises a complex question, the semantic generation unit can organize detailed and comprehensive answer content based on the semantic connotation of the question and the relevant knowledge fragments in the knowledge graph. For a customer who has shown a preference for concise, professional and efficient communication in past consultations, the model will automatically select templates with high rationality and professionalism personality scores to reply to new business questions. For another customer who prefers a relaxed and friendly atmosphere, templates with a more prominent lively and humorous style will be selected for interaction. This personalized service mode can significantly improve customer engagement and satisfaction, and enhance the stickiness and trust relationship between customers and intelligent customer service.

[0114] The emotion integration unit analyzes the emotional tendencies in the customer's consultation text and determines whether the customer is happy, dissatisfied, anxious or in other emotional states. For example, through emotion analysis technology, it is identified that the customer may have angry or dissatisfied emotions when complaining about a problem. Based on the emotion analysis results, emotional factors are appropriately integrated into the reply content to make the reply more in line with the emotional needs of the customer. For example, when the customer expresses dissatisfaction, the reply should reflect an understanding and apologetic attitude, and increase soothing words and positive solutions to improve customer satisfaction and acceptance.

[0115] The quality evaluation and optimization unit evaluates the quality of the generated reply content from multiple dimensions such as grammatical correctness, semantic accuracy, information integrity, logical coherence, etc. For example, it checks whether there are any misspelled words in the reply, whether the sentences are smooth, whether the information provided meets the customer's needs, and whether the transition between parts of the content is natural, etc. Based on the quality evaluation results, the reply generation model is optimized and adjusted. If it is found that some replies have quality problems or customer feedback is not good, the reasons are analyzed and the model is trained and improved accordingly to continuously improve the quality and effectiveness of the replies. For example, when it is found that grammatical errors often occur in a certain type of reply, the grammar checking and correction mechanism for this part of the content is strengthened; if the customer is not satisfied with a certain type of answer, the model's handling logic and answer strategy for that type of question are further optimized.

[0116] The reply generation optimization module predicts the probability of the customer on a certain personality preference dimension (such as whether to like humorous style, represented by y, y = 1 represents like, y = 0 represents not like) according to the historical interaction record (characterized as X hist ), personal information (characterized as X info ) and current consultation question type and context environment (characterized as X context ), and the formula of the logistic regression model is as follows:

[0117]

[0118] Wherein:

[0119] p(y = 1 | X hist , X info , X context ) is the probability of the customer liking a certain personality preference given the historical interaction, personal information and context environment;

[0120] z = β0 + β1X hist + β2X info + β3X context , wherein β0 is the intercept term, β1, β2, β3 are the regression coefficients of the corresponding features.

[0121] Embodiment 2:

[0122] An application of an intelligent customer service dialogue model, specifically comprising:

[0123] E-commerce field:

[0124] In the customer service of the e-commerce platform, the intelligent customer service dialogue model can handle various problems of customers about commodity information inquiry, order tracking, logistics consultation, after-sales return and exchange, etc. Through the multi-modal interaction function, customers can directly upload commodity pictures or shoot commodity videos to inquire about related problems. The model combines commodity knowledge graph and multi-modal data analysis to quickly and accurately provide customers with commodity details, use suggestions, collocation recommendations and other information, effectively improving customer shopping experience and purchase conversion rate.

[0125] In the e-commerce live broadcast scene, the host can use the intelligent customer service model to answer the audience's questions about the live broadcast goods in real time, and at the same time the model can select appropriate intelligent customer service templates according to the atmosphere of the live broadcast room and the audience interaction situation to provide humorous and interesting or professional and detailed answers, increase the interest and interactivity of the live broadcast, and promote the sales of goods.

[0126] Financial service field:

[0127] After the model is introduced into the customer service center of banks, securities and other financial institutions, it can handle various business intents such as customer account inquiries, money transfer, financial product consultation, loan application, etc. When dealing with complex issues such as financial risk assessment and investment advice, the model responds cautiously based on the financial knowledge graph and strict risk control rules to ensure the accuracy and compliance of the information. At the same time, through personalized service mode, it provides customized financial service solutions for customers with different asset sizes and risk preferences, improving customer satisfaction and loyalty.

[0128] Smart home and Internet of Things field:

[0129] Users of smart home devices can interact with the intelligent customer service model through voice or mobile app, query device status, control device operation, obtain home environment monitoring data, etc. The model realizes precise control and intelligent management of smart home devices through the call of home device knowledge graph and the understanding of user voice instructions. For example, the user can say "help me adjust the temperature of the air conditioner in the living room to 26 degrees", and the model can analyze the instruction and complete the temperature adjustment operation through the docking with the home automation system; or the user asks "how was the electricity consumption at home last night?", The model can extract relevant information from the smart home electricity meter data and generate a detailed electricity consumption report to reply to the user.

[0130] Telecommunications and communication field:

[0131] The customer service department of the telecommunications operator uses the model to handle customer problems such as package consultation, fee inquiry, traffic package change, network fault repair, etc. When customers consult package combinations, the model analyzes and compares based on the telecommunications business knowledge graph and customer historical consumption data to recommend the most favorable and most suitable package plan for the customer; when the customer reports network failure, the model can obtain fault location and estimated repair time information through data interaction with the network management system, and feedback to the customer in time, improving fault handling efficiency and customer satisfaction.

[0132] The above is only the preferred specific implementation of the present application, but the protection scope of the present application is not limited to this, any skilled person in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for constructing an intelligent customer service dialogue model, characterized in that, The system comprises an intention recognition module, a knowledge graph management module, a multi-modal interaction module, and a reply generation optimization module. The intention recognition module judges the intention category of customer consultation, adopts a natural language processing technology based on deep learning, combines word vector representation and a bidirectional long short-term memory network to perform semantic analysis on the text input by the customer, and is trained through labeled corpus to autonomously learn the feature expression of different business intentions, thereby accurately classifying the questions of the customer into a predefined intention category system. The reply generation optimization module comprehensively considers the intention recognition result and knowledge graph query information, adopts a hybrid generation strategy based on the combination of templates and deep learning, pre-designs multiple intelligent customer service templates for different intention categories, each template comprising multiple personality tags and corresponding scores representing the degree of personality characteristics in the template, dynamically selects the most suitable intelligent customer service template for reply generation according to the question content and past interaction records of the customer, optimizes the reply content by using RNN, LSTM or GRU in combination with an attention mechanism on the basis of the template, and adjusts and optimizes the reply generation strategy by using a reinforcement learning algorithm according to the feedback evaluation of the customer on the reply, such as a satisfaction score and follow-up questions, to gradually improve the accuracy and individuality of the reply. The knowledge graph management module extracts entity, relationship and attribute information from multiple source heterogeneous data in the enterprise, including product information, service terms, frequently asked questions and historical customer cases, stores and organizes these knowledge elements by using a graph database technology to form a structured knowledge network, and quickly retrieves and reasons in the knowledge system based on the hierarchical relationship and semantic association rules between the defined entities to provide information basis for subsequent reply generation. The reply generation optimization module predicts the probability of the customer on a certain personality preference dimension, denoted as y, y = 1 represents liking, and y = 0 represents disliking, according to the historical interaction record of the customer, denoted as , the personal information, denoted as , and the current consultation question type and context environment, denoted as , and the formula of the logistic regression model is as follows: wherein: is the probability that the customer likes a certain personality preference given the history of interactions, personal information, and contextual environment; wherein is an intercept term, , , are regression coefficients for the corresponding features, respectively. 2.The intelligent customer service dialogue model construction method of claim 1, wherein, The multi-modal interaction module adopts a speech recognition engine to accurately convert the voice instructions of the customer into text form in the aspect of voice interaction, and utilizes a speech synthesis technology to play the text reply generated by the model in a clear and natural voice to the customer to realize a seamless voice interaction process. In the image interaction scenario, the model receives and analyzes the picture or video information provided by the customer, generates analysis text in combination with the related knowledge in the knowledge graph, and finally feeds back the corresponding image information to the customer. The intention recognition module comprises: 3.The intelligent customer service dialogue model construction method of claim 1, wherein, A text preprocessing unit cleans the original text input by the customer to remove irrelevant characters, links, special symbol noise information, and performs word segmentation processing on the text by using morphological analysis technology to split the sentence into word units with actual semantics. A feature extraction unit adopts word vector representation technology to map each word in the segmented text into a low-dimensional vector space to capture the semantic information and context relationship of the word, and extracts the syntactic structure features of the text, such as the grammatical dependency relationship of the sentence, the subject-predicate-object structure, etc., and represents them through a syntactic analysis tree. ​ The intent classification unit constructs an intent classification model based on a deep learning algorithm, adopts CNN, RNN or LSTM, GRU, takes the extracted word vector features and syntactic structure features as inputs of the model, and learns through training of the model to automatically determine the intent category of the customer inquiry and output the corresponding probability distribution. 4.The intelligent customer service dialogue model construction method of claim 2, wherein, The knowledge graph management module includes: The knowledge extraction unit extracts knowledge entities, relationships and attribute information from various structured data sources and unstructured data sources in the enterprise; for structured data, relevant data is obtained by writing SQL query statements or using data crawling tools; for unstructured data, the named entity recognition and relationship extraction technology in natural language processing technology is used to identify entities in the text and their relationships; the extracted knowledge is cleaned and standardized to remove duplicate and incorrect information, unify the format and specifications of knowledge representation in different data sources, and ensure the quality and consistency of the knowledge; The knowledge fusion unit uses entity linking technology to merge and remove duplicate entities from different data sources, establishes equivalent relationships between entities, and integrates and integrates knowledge based on ontology and semantic rules to build a complete knowledge network; The knowledge storage unit stores the constructed knowledge graph based on a graph database management system, and queries knowledge based on the index structure in the graph database. 5.The intelligent customer service dialogue model construction method of claim 2, wherein, The multi-modal interaction module includes: The voice interaction unit converts the voice signal input by the customer into text information using automatic speech recognition technology, and converts the text reply generated by the intelligent customer service model into voice output to the customer using text-to-speech technology; The image interaction unit receives and parses the image file uploaded by the customer, extracts key features and information from the image, and understands and analyzes the parsed image information based on computer vision technology and image recognition algorithms, and generates reply and feedback information for the image content in combination with related knowledge in the knowledge graph. 6.The intelligent customer service dialogue model construction method of claim 1, wherein, The reply generation optimization module adjusts the personality score formula based on the customer feedback, supposing that the initial score of a certain intelligent customer service template on a certain personality dimension is , the influence factor of the customer feedback on the dimension is , a positive value represents positive feedback, and a negative value represents negative feedback, and the adjusted score is represented as: wherein: is the adjusted score; is the initial score; is an adjustment coefficient for controlling the degree of influence of the customer feedback on the score; is the amount of change in the score caused by the customer feedback; The template total score is calculated according to the scores of all personality dimensions, and the intelligent customer service template has n personality dimensions, the score of the i-th personality dimension is The comprehensive total score of the template is The comprehensive total score of the template is wherein: is the overall score of the template; is the number of personality dimensions; is the score of the i-th personality dimension.

7. The intelligent customer service dialogue model construction method according to claim 1, characterized in that, The reply generation optimization module includes: The template matching unit finds a matching reply template in the pre-constructed intelligent customer service template library according to the intent recognition result; these templates are classified and arranged according to different business intents and scenarios, covering standardized answers to common questions and common scenarios; at the same time, the found template is adaptively adjusted and modified, and the variable part in the template is filled or replaced according to the specific problem and context information of the customer, so that the reply is more suitable for the actual situation of the customer; The semantic generation unit constructs a semantic generation model based on a deep learning recurrent neural network, learns and trains the model through learning and training of artificial customer service reply corpus, and enables the model to autonomously learn the rules and patterns of text generation; when generating a reply, the model analyzes the personality preference and communication style of the customer by using a machine learning algorithm according to the intention recognition result and the relevant information queried from the knowledge graph, in combination with the current context; then, in combination with the personality feature vectors of various templates in the intelligent customer service template library, the similarity or matching degree index is calculated to recommend the intelligent customer service template most suitable for the individual needs of the customer as the dominant reply style for this interaction; and a reply text is generated. 8.The intelligent customer service dialogue model construction method of claim 7, wherein, The reply generation optimization module further comprises: The sentiment integration unit analyzes the sentiment tendency in the customer consultation text, judges whether the customer is happy, dissatisfied, anxious or in other emotional states, and appropriately integrates sentiment factors in the reply content according to the sentiment analysis result, so that the reply is closer to the emotional needs of the customer; The quality evaluation and optimization unit evaluates the quality of the generated reply content; and optimizes and adjusts the reply generation model according to the quality evaluation result.

Images