Customer service processing method, device and equipment based on big data, storage medium and product

Abstract

The invention provides a customer service processing method and device based on big data, equipment, a storage medium and a product, and belongs to the technical field of natural language process.The method comprises the steps that a user appeal is received, and the user appeal comprises a text or voice; the user appeal is input into a pre-trained deep learning model for semantic analysis, at least one key intention and context semantic information are extracted, and the pre-trained deep learning model is obtained based on BERT model training; matching each analysis result with a pre-constructed intention knowledge base to obtain at least one intention conforming to the user appeal; calling a corresponding agent from a pre-obtained agent group, and performing corresponding processing on the user appeal; and tracking the whole process, and feeding back a processing result to the client. According to the method, efficient customer demand identification and processing are realized, the customer service efficiency and intelligence are improved, and the user experience is improved.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of natural language processing technology, and in particular, to a customer service processing method, device, electronic device, computer-readable storage medium, and computer program product based on big data. Background Art

[0002] With the rapid development of the telecommunications industry, the complexity and demand for customer service are increasing. Traditional customer service systems often rely on manual processing of customer requests, which is not only inefficient but also prone to information omissions and errors. In recent years, customer service systems based on natural language processing have been widely used. However, there are many limitations in existing technologies:

[0003] 1) Low processing efficiency: Manual customer service needs to handle customer requests one by one. Especially when faced with highly concurrent customer demands, the processing speed is limited, resulting in long waiting times for customers and reduced service quality.

[0004] 2) Lack of intelligence: Traditional automatic response systems rely on predefined rules and templates and cannot flexibly respond to complex customer demands, especially when dealing with diverse language inputs, making it difficult to accurately understand the customer's true intentions.

[0005] 3) Lack of personalized services: The existing system lacks the ability to identify customer emotions and personalized needs, resulting in poor customer experience. In particular, when customers are anxious or angry, the response strategy cannot be adjusted in time, further exacerbating customer dissatisfaction.

[0006] 4) Insufficient system self-optimization capabilities: Most existing systems lack effective adaptive learning mechanisms and are unable to continuously optimize based on historical processing data and customer feedback.

[0007] In view of the shortcomings of existing technologies, a more optimized method for handling customer service demands is urgently needed. Summary of the invention

[0008] The purpose of the embodiments of the present disclosure is to provide a customer service processing method, device, electronic device, computer-readable storage medium and computer program product based on big data, so as to solve the above-mentioned problems existing in the prior art.

[0009] In order to achieve the above objectives, the technical solutions adopted by the embodiments of the present disclosure are as follows:

[0010] On the one hand, an embodiment of the present disclosure provides a customer service processing method, the method comprising:

[0011] Receiving a user demand, wherein the user demand includes: text or voice;

[0012] Inputting the user demand into a pre-trained deep learning model for semantic analysis to extract at least one key intent and at least one contextual semantic information, wherein the pre-trained deep learning model is based on: BERT model training;

[0013] Match the parsing results with the pre-built intent knowledge base to obtain at least one intent that meets the user's needs, fill the slots for each intent that meets the user's needs, and extract key information;

[0014] According to each intention that meets the user's demands and the corresponding extracted specific information, the corresponding agent is called from the pre-obtained agent group to process the user's demands accordingly;

[0015] Track and manage the overall process of handling user demands, and feedback the processing results to the client.

[0016] Exemplarily, the matching of each analysis result with a pre-built intent knowledge base to obtain at least one intent that meets the user's needs includes:

[0017] Calculate the similarity between the parsing result and each intent in the pre-built intent knowledge base to obtain a similarity set;

[0018] Take the top k intent samples with the highest similarity and the corresponding analysis results from the self-similarity set, input them into the pre-trained deep learning model for judgment, and select at least one intent that meets the user's needs, where k ≥ 2.

[0019] Exemplarily, the process of constructing the pre-built intention knowledge base includes:

[0020] According to historical data, various intent samples are collected, and each intent sample is vectorized to obtain an intent knowledge base, in which each intent sample has a corresponding classification.

[0021] Exemplarily, before semantically parsing the user's request, the method further includes:

[0022] If the user's request includes voice, the voice will be converted into text;

[0023] The text of user requests is preprocessed by segmenting and removing stop words.

[0024] Exemplarily, the agent group includes a plurality of agents, each of which processes a type of user intention;

[0025] Before calling the Agent method, initialize the Agent, including:

[0026] Use the reinforcement learning algorithm DQN to initialize each agent's strategy model and establish a corresponding strategy model for each agent;

[0027] The method of calling a corresponding agent from a pre-obtained agent group according to at least one intention that meets the user's demands and corresponding extracted specific information to process the user's demands accordingly includes:

[0028] Assign user demands to corresponding agents based on each intent that meets user demands and the extracted specific information;

[0029] If there are multiple candidate agents for a user intent, the best agent is selected based on historical performance data or current workload;

[0030] After calling the Agent, the method further includes:

[0031] If the agent cannot solve the user problem in the process of handling the user's request, it will determine whether it exceeds the predetermined conditions. If so, the task will be upgraded to a manual customer or a preset senior agent.

[0032] Exemplarily, the tracking and management of the overall process of handling user demands and feeding back the processing results to the client include:

[0033] Track the entire process of each user's request, including request creation, processing progress and feedback results; specifically including:

[0034] After analyzing the customer's request, a work order is automatically generated, which includes a detailed description of the customer's problem, the problem type, and the processing progress;

[0035] Track the progress of each work order in real time and provide feedback to customers;

[0036] After the work order is processed, the customer will be provided with result feedback. If the problem is solved, the work order will be closed; if not, the next step will be prompted.

[0037] Exemplarily, after feeding back the processing result to the client, the method further includes:

[0038] Collect customer interaction data;

[0039] Regularly retrain deep learning models and agent strategy models based on historical data;

[0040] Evaluate the performance of deep learning models BERT and agents to ensure performance improvements of optimized models in production environments.

[0041] Exemplarily, the method further includes:

[0042] If the user's request includes voice, convert the voice into text;

[0043] Based on the pre-trained sentiment analysis model, analyze the emotional state in the user's request text to determine whether the user is in a positive, negative or neutral emotion;

[0044] According to the analysis results, the service strategy of the intelligent agent is adjusted. For customers with strong negative emotions, their problems will be handled first and given more detailed responses.

[0045] During the entire processing process, the customer's emotional changes are continuously monitored and real-time feedback is provided to the agent or manual customer service to assist in decision-making.

[0046] Another aspect of the present disclosure is to provide a customer service processing device based on big data, the device comprising:

[0047] A receiving module, configured to receive a user demand, wherein the user demand includes: text or voice;

[0048] A parsing module, used to input the user demand into a pre-trained deep learning model for semantic parsing, and extract at least one key intent and at least one contextual semantic information, wherein the pre-trained deep learning model is trained based on a natural language understanding model including a BERT model;

[0049] The matching module is used to match the parsing results with the pre-built intent knowledge base to obtain at least one intent that meets the user's needs, fill the slots for each intent that meets the user's needs, and extract key information;

[0050] The scheduling module is used to call the corresponding agent from the pre-obtained agent group according to each intention that meets the user's demands and the corresponding extracted specific information, and process the user's demands accordingly;

[0051] The tracking module is used to track and manage the overall process of user demand processing and feed back the processing results to the client.

[0052] On the other hand, an embodiment of the present disclosure proposes an electronic device, including: a processor, and a memory communicatively connected to the processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the method steps described above.

[0053] On the other hand, an embodiment of the present disclosure provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method steps described above are implemented.

[0054] Another aspect of the embodiments of the present disclosure provides a computer program product, including a computer program, which implements the method steps described above when executed by a processor.

[0055] The beneficial effects of the embodiments of the present disclosure are:

[0056] The disclosed embodiment proposes a customer service processing method based on big data, which adopts the fusion technology of deep learning model and intelligent agent to realize efficient automatic identification, classification and automatic processing of customer demands, solves the problems of low processing efficiency and insufficient intelligence of traditional customer system, and improves customer experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flowchart of a customer service processing method based on big data proposed in an embodiment of the present disclosure;

[0058] Figure 2 It is a framework diagram of a customer service processing method based on big data proposed in an embodiment of the present disclosure;

[0059] Figure 3 It is a schematic diagram of the structure of a customer service processing engine proposed in an embodiment of the present disclosure;

[0060] Figure 4 It is a structural schematic diagram of a customer service processing device based on big data proposed in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the embodiment of the present disclosure more clear, the embodiment of the present disclosure is further described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation methods described here are only used to explain the embodiment of the present disclosure and are not used to limit the embodiment of the present disclosure.

[0062] like Figure 1 As shown, the embodiment of the present disclosure proposes a customer service processing method based on big data, and the method includes:

[0063] Step S100: receiving a user request, wherein the user request includes: text or voice.

[0064] The method of the disclosed embodiment can be used in customer service of telecommunication user demands. The user inputs the demand content on the client side, which can be in text or voice form; the receiving end receives the demand data sent by the client side, and the demand content can be complaints, consultations, services, etc.

[0065] After receiving the user's request in step S100, the user's request data may be pre-processed;

[0066] Specifically, step S110, receiving a text input or voice input request from a customer, and if the user's request is voice, converting the voice into text through voice recognition ASR;

[0067] Step S120: Use the NLP tool library BERT Tokenizer to perform pre-processing such as word segmentation and text cleaning on the text. Text cleaning includes: removing stop words.

[0068] Let's use a specific example to illustrate. The user enters the request: Hello, I want to check my bill and change my service package!

[0069] After preprocessing, I get: I want to check my bill and change my service plan.

[0070] Step S200: input the user's demands into a pre-trained deep learning model for semantic analysis, and extract at least one key intent and at least one contextual semantic information, wherein the pre-trained deep learning model is based on: BERT model training.

[0071] The preprocessed text is input into a pre-trained deep learning model for semantic analysis. The deep learning model converts the text into a word vector in vector form, and extracts at least one key intent and at least one context semantic vector. The pre-trained deep learning model is based on a natural language understanding model including a BERT model. The method of the disclosed embodiment is based on the deep learning pre-trained model BERT, which performs semantic analysis and intent recognition on the input text to obtain at least one key intent and at least one context semantic information vector, as well as a probability value for each key intent, as a parsing result. These intents can be preliminarily screened according to the probability values, and several intents with higher probabilities are selected as candidate intents.

[0072] According to the example of step S100, the key intents may include "bill inquiry" and "service change", with probability values ​​of 0.9 and 0.8 respectively. The context information may include time, location, etc. "Bill inquiry" and "service change" are candidate intents. The BERT model outputs vector format.

[0073] Step S300: Match the parsing results with the pre-built intent knowledge base to obtain at least one intent that meets the user's needs, fill the slots for each intent that meets the user's needs, and extract specific information.

[0074] In step S300 of the disclosed embodiment, user intent is identified and classified, that is, user intent and demand type need to be obtained. First, an intent knowledge base is constructed based on historical data, and vectorized matching is performed based on the semantic parsing results and the annotated intents of the intent knowledge base, and then judged by a large model, such as BERT, RoBERTa, T5, GPT, etc., to finally identify the customer's intent and question type, such as "bill inquiry", "service change" and "technical support".

[0075] Specifically, we first define a set of standard intent categories, such as "bill inquiry", "service change", "technical support", etc. We can collect various intent samples based on historical data. Each sample data should be labeled with one or more intent categories. Each intent category contains multiple examples and related contextual information. We vectorize the collected original intent samples and build an intent knowledge base. In this way, each intent sample has its own question type. We vectorize the semantic parsing results, that is, we vectorize the key intents and contextual semantic information and match them with the intent knowledge base. If there are multiple candidate intents, we match each candidate intent with the pre-built intent knowledge base.

[0076] The matching process includes: using the cosine algorithm to calculate the similarity between each parsing result and each annotated intent sample vector in the intent knowledge base, selecting the top k intent samples with the highest similarity for each parsing result, and taking the top k values, where k≥2, specifically, k=3 or 5. That is, calculating the similarity between each candidate intent and each annotated intent sample vector in the intent knowledge base, sorting the retrieved top k results from large to small, that is, taking the top k results, and submitting them to the large model for judgment. Large models such as BERT, RoBERTa, etc. select the intent that best meets the customer's semantics, or there can be multiple intents that best meet the user's demands.

[0077] According to the example of step S200, the preliminary matching results may include: bill query, similarity score: 0.95, service change, similarity score: 0.90, technical support, similarity score: 0.75.

[0078] Specifically, matching the parsing result with the pre-built intent knowledge base to obtain at least one intent that meets the user's needs includes:

[0079] Step S310: Calculate the similarity between the parsing result and each intent in the pre-built intent knowledge base to obtain a similarity set K.

[0080] In step S310, the similarity between the parsing result vector and each intent sample vector in the pre-built intent knowledge base is calculated. The specific calculation process is not repeated here.

[0081] Step S320: Take the top k results with the highest similarity from the similarity set, input them into the big model for judgment, and select at least one intention that meets the user's demands, where k≥2.

[0082] Here, the big model can be BERT or RoBERTa, or other big language models can be called. Select the intent that best meets the user's needs. The results include: intent sample, parsing result, and similarity score. There are two ways to input the results to the big model. The first way is to combine the intent samples matched from the intent knowledge base, the parsing results, and the user input into an input sequence. Specifically, the user input and the intent sample are spliced ​​together to form a new input sequence, which is input to the big model.

[0083] Method 2: The intent samples matched from the intent knowledge base, the similarity scores between each intent sample and the user input, and the user input are concatenated together to form a new input sequence, which is then input into the large model.

[0084] After the results are input into the big model, the first method is adopted. The big model outputs the probability value of each intent sample, and the intent sample with the highest probability is selected as the final user intent. There may be multiple intents that best meet the user's demands. The second method is adopted, which uses multi-indicator scoring. The big model outputs the comprehensive score of each intent sample, and the intent sample with the highest comprehensive score is selected as the final user intent. Similarly, there may be multiple intents that best meet the user's demands. Multiple indicators are introduced in the multi-indicator scoring, such as similarity score, user historical behavior, context information, etc., and weights are set for each indicator. User historical behavior and context information are used to comprehensively score each intent sample, and the score is further optimized by weighted average or machine learning models (such as random forest, XGBoost, etc.). Specifically, the multiple indicators include: similarity score, user historical behavior, context information, social information, and system information; the similarity score is the similarity between the user's demands and each intent sample in the intent knowledge base; the user's historical behavior is the user's past behavior and preference, which can be extracted from the user profile or database, and the score is calculated based on the user's past behavior and preference. Specifically, the user's past behavior frequency and preference are counted and the corresponding score is assigned. For example, the intent that the user frequently uses can be assigned a higher score. Context information, the user's current session context and other environmental information, can be obtained from session history, geolocation API, timestamp, etc., and the score is calculated based on the relevance of context information. Specifically, timestamp: different scores are assigned according to the time period when the message is sent (such as working hours, rest time), and geolocation: different scores are assigned according to the user's location information (such as home, office, outdoor). For example, messages sent by users during working hours may be more inclined to work-related intentions. Social information includes the user's social network relationships and group behaviors, which can be extracted from social platforms or community data, and the score is calculated based on the user's social network relationships and group behaviors. Social network analysis: analyze the user's social network relationships and extract the common intentions of friends. Group behavior analysis: analyze the common behaviors and preferences of the group to which the user belongs. For example, the intentions frequently used by the user's friends can be assigned higher scores. Analyze the user's social network relationships, extract the common intentions of the user's friends, and understand the user's social circle preferences. Specifically, it includes: Data collection: obtain the user's social network data from social platforms (such as Facebook, Twitter, WeChat, etc.), including friend lists and interaction records. Data preprocessing: clean the data and extract the intention history of the user's friends. Intent frequency statistics: count the common intentions and their frequencies of each friend. Weighted average: Take a weighted average of the intention frequency of friends based on the interaction frequency or intimacy of friends. Analyze the common behaviors and preferences of the group to which the user belongs. Goal: Understand the common intentions and behavior patterns of the group to which the user belongs. Specifically include: Data collection: Obtain information about the group to which the user belongs from social platforms or community data, including group members and activity records.Data preprocessing: clean the data and extract the intention history of group members. Intent frequency statistics: count the common intentions and their frequencies of each group member. Weighted average: perform a weighted average of the intention frequencies of group members based on their activity or influence.

[0085] System information includes system log analysis and system status monitoring. System log analysis: Analyze the system's operation logs and extract key information. System status monitoring: Monitor the current status of the system (such as load, resource usage), and assign different scores based on the status. For example, the system may be more inclined to handle emergency intent under high load.

[0086] The weights of multiple indicators can be respectively: the weight of the similarity score is set to 0.3, the weight of the user's historical behavior is set to 0.2, the weight of the context information is set to 0.2, the weight of the social information is set to 0.1, and the weight of the system information is set to 0.2.

[0087] After obtaining at least one intention that meets the user's needs in step S320, the method further includes:

[0088] Step S330: fill in each slot that meets the user's intent and extract specific information respectively.

[0089] In the slot filling of step S330, the sequence labeling model LSTM+CRF can be used to extract key information from the user demand input. Specific information is extracted from the user demand, such as the user ID and time range in the bill query. According to the above example, the user enters a text, such as: "I want to query my bill and change the service package. The user ID is 123456. The time range is from January 1, 2023 to February 28, 2023". The specific information is time, location, user identity ID, etc.

[0090] Step S400: According to the intention of each user's request and the corresponding extracted specific information, the corresponding agent is called from the pre-obtained agent group to process the user's request accordingly.

[0091] The intelligent agent group includes multiple agent scheduling modules, each agent is responsible for processing a type of user demand, and the processing method can be set according to the actual situation. Among them, the user intention types include: billing problems, technical support, service changes, package changes, network connection failures, and disassembly refunds, etc.

[0092] Based on the processing results of natural language NLP (Natural Language Processing), different types of customer demands are assigned to corresponding agents for processing. Each agent makes task decisions through reinforcement learning to ensure the best processing strategy.

[0093] Before calling the Agent method, initialize the Agent strategy model, including:

[0094] The reinforcement learning algorithm DQN is used to initialize the strategy model of each agent, and a corresponding strategy model is established for each agent to ensure that they can make appropriate decisions in different situations.

[0095] Different agents are responsible for handling different types of customer intent, such as billing issues, technical support, service changes, and device disassembly refunds.

[0096] like Figure 2 As shown, according to each intention that meets the user's demands and the extracted specific information, the corresponding agent is called from the pre-obtained agent group to process the user's demands accordingly; including:

[0097] Step S410: assign the user demands to corresponding Agents according to each intention that meets the user demands and the extracted specific information.

[0098] According to the intent classification and slot information provided by the NLP module, customer demands are assigned to the most suitable corresponding Agent.

[0099] Step S420: If there are multiple candidate Agents for a user's intention, the best Agent is selected according to historical performance data or current workload.

[0100] If there are multiple candidate agents for a user's intent, the best agent is selected using a load balancing algorithm based on historical performance data or current workload.

[0101] Different agents can be constructed by different fine-tuned expert models. When calling an agent, it checks whether the corresponding expert model is loaded.

[0102] Use load balancing algorithm combined with reinforcement learning model for dynamic agent allocation, and dynamically load expert large models based on the Ollama framework.

[0103] After calling the Agent, the method further includes:

[0104] If the agent cannot solve the user problem in the process of handling the user's request, it will determine whether it exceeds the predetermined conditions. If so, the task will be upgraded to a manual customer or a preset senior agent.

[0105] The system sets processing time limits and failure conditions. When the preset conditions are exceeded, the task upgrade mechanism is triggered. The preset conditions refer to processing time limits and failures. Agent dynamically loads fine-tuned models: After using the Agent, the corresponding fine-tuned model is released to reduce the system load; when a new Agent requirement is received, the model fine-tuned by the latest self-optimization mechanism is loaded and applied to ensure the Agent's ability to improve and seamlessly switch applications. An Agent can handle multiple user intentions of the same type at the same time.

[0106] Step S500: Track and manage the overall process of handling user demands, and feed back the processing results to the client.

[0107] Track and manage the entire process of each customer's request, including the creation of the request, processing progress, result feedback, etc. The method of the disclosed embodiment can automatically generate work orders and support docking with third-party systems. Third-party systems refer to external systems that are integrated with the customer service platform to expand the platform's functions and services.

[0108] The tracking and management of the overall process of handling user demands and feeding back the processing results to the client include:

[0109] Step S510: Track the entire process of each user's request, including request creation, processing progress, and feedback results; specifically including:

[0110] Step S511: After analyzing the customer's request, a work order is automatically generated, which includes a detailed description of the customer's problem, the problem type, and the processing progress.

[0111] After analyzing the customer's request, a work order is automatically generated, which includes a detailed description of the customer's problem, the type of problem, and the progress of the processing. If there is feedback, it also includes the feedback result. Specifically, the database is used to automatically record the key information of the customer's problem and generate a work order with a unique identifier.

[0112] Step S512: Track the processing progress of each work order in real time and provide feedback to the customer.

[0113] Monitor the processing status of the Agent in real time, use the message queue system (RabbitMQ) to notify progress updates, and implement progress tracking.

[0114] Step S513: After the work order is processed, provide result feedback to the customer. If the problem is solved, close the work order; if not, prompt the next step.

[0115] The processing results are fed back to the client application or SMS notification system through API or message system to achieve result feedback.

[0116] Step S600: Continuously optimize the deep learning model BERT and the intelligent agent based on the feedback data during the user demand processing.

[0117] This step S600 continuously optimizes the BERT model and the Agent strategy model based on the feedback data in the process of handling customer demands to improve the processing accuracy and efficiency of the system, which can be realized by the adaptive learning module. The feedback data includes: customer feedback results, processing result data, internal review data, log data, etc.

[0118] Step S610: Collect customer interaction data, including processing results, customer feedback, and processing time, and store them in a database.

[0119] Use a database or data warehouse system (Hadoop) to record all customer interaction data. Collect customer interaction data, including processing results, customer feedback, and processing time, and store them in the database to achieve data collection.

[0120] Step S620: Regularly retrain the deep learning model and the agent strategy model based on historical data.

[0121] Based on historical data, the deep BERT model and Agent strategy model are regularly retrained to ensure continuous optimization of each model and achieve model adjustment. The NLP model and Agent strategy model are regularly retrained using the deep learning framework (PyTorch).

[0122] Step S630: Evaluate the deep learning model and agent performance to ensure that the performance of each optimized model is improved in the production environment.

[0123] The model performance is evaluated through the automated evaluation module to ensure the performance improvement of the optimized model in the production environment. Specifically, the indicator evaluation system (Precision, Recall, F1Score) is used to evaluate the model performance to ensure the optimization effect.

[0124] The method of the disclosed embodiment also includes performing sentiment analysis on the customer, specifically analyzing the customer's emotional state, using the analysis results to adjust the Agent's response strategy and improve customer satisfaction, which can be achieved by the sentiment analysis module. Sentiment analysis can be performed in the appeal intention matching stage or in the Agent calling stage.

[0125] Specifically, the method further includes: performing sentiment analysis on the customer; performing sentiment analysis on the customer in the Agent calling stage;

[0126] If the user's request includes voice, convert the voice into text;

[0127] Step S430: Analyze the emotional state of the user's request text based on the pre-trained sentiment analysis model to determine whether the user is in a positive, negative or neutral emotion.

[0128] The sentiment classification model (BERT for Sentiment Analysis) is used to classify the sentiment tendency of customer input, that is, to classify the sentiment tendency of the received user demand text. The sentiment analysis model can also use some existing models that can perform sentiment analysis on the text to obtain the corresponding emotional state classification.

[0129] Step S440: According to the analysis results, the service strategy of the intelligent agent is adjusted. For customers with strong negative emotions, their problems will be handled first and given more detailed responses.

[0130] According to the analysis results, the agent's service strategy is adjusted. For customers with strong negative emotions, this method will give priority to their problems and give more detailed responses.

[0131] Dynamically adjust service strategies through decision trees or reinforcement learning models.

[0132] Step S450: During the entire processing process, the customer's emotional changes are continuously monitored and real-time feedback is given to the Agent or manual customer service to assist in decision-making.

[0133] During the entire processing process, the method of the disclosed embodiment continuously monitors the emotional changes of the customer and provides real-time feedback to the agent or manual customer service to assist in decision-making. A real-time emotional monitoring algorithm is set to ensure that the emotional changes of the customer can be responded to in a timely manner during the processing process.

[0134] In the customer service processing method of the disclosed embodiment, 1. The customer submits a request to the telecommunications customer service system in the form of text or voice. 2. The deep NLP module parses the customer's request and identifies the customer's needs and problem types. 3. Based on the ReAct framework, the relevance of the intent and the corresponding Agent is judged, and the request is assigned to the most suitable Agent for processing. 4. The Agent performs automated processing based on its decision-making model and the relevant knowledge base within the system, and transfers the task to manual customer service when necessary. 5. The system tracks the entire process through the request management module, and promptly feeds back the processing results to the customer. 6. The adaptive learning module will iteratively optimize the system based on the feedback to improve overall processing efficiency and customer satisfaction.

[0135] The disclosed embodiment proposes a telecommunications customer service demand management method and engine based on the fusion of deep natural language processing (NLP) and intelligent agent technology. By intelligently parsing, classifying and processing the demands expressed by customers in natural language, and combining adaptive learning and sentiment analysis mechanisms, the low processing efficiency and insufficient intelligence of the traditional customer service system are solved. The disclosed embodiment has significant advantages and effects in terms of intelligence, customer service processing efficiency, customer experience, etc.

[0136] 1) High degree of intelligence, reducing manual input: Through the combination of deep NLP and intelligent agent technology, the system can automatically handle complex customer demands and optimize itself based on actual feedback. Six months after going online, the number of front-line customer service staff was reduced by 20%.

[0137] 2) Improve processing efficiency: Compared with the traditional manual customer service system, the disclosed embodiment can significantly improve the efficiency of customer request processing. The work order processing time is increased from 2.3 orders / hour before going online to 3.2 orders / hour after going online, and the efficiency is improved by 39%.

[0138] 3) Improve customer experience: Through the application of sentiment analysis and intelligent response strategies, the system can better understand customer emotions and make appropriate responses. Customer satisfaction increased from 86.5% before the launch to 92.2% after the launch, an increase of 5.7PP.

[0139] like Figure 4 As shown, another aspect of the present disclosure embodiment provides a customer service processing device based on big data, the device comprising:

[0140] The receiving module 100 is used to receive a user demand, wherein the user demand includes: text or voice.

[0141] The parsing module 200 is used to input the user's demands into a pre-trained deep learning model for semantic parsing, and extract at least one key intent and at least one contextual semantic information, wherein the pre-trained deep learning model is based on: BERT model training.

[0142] The matching module 300 is used to match the analysis results with the pre-built intent knowledge base to obtain at least one intent that meets the user's needs, fill the slots for each intent that meets the user's needs, and extract key information.

[0143] The scheduling module 400 is used to call the corresponding agent from the pre-obtained agent group according to each intention that meets the user's demand and the corresponding extracted specific information, and process the user's demand accordingly.

[0144] The tracking module 500 is used to track and manage the overall process of handling user demands and feed back the processing results to the client.

[0145] like Figure 3 As shown, the above modules, the adaptive learning module and the sentiment analysis module together constitute the customer service management engine. Each module implements the corresponding steps in the above method. The customer service management engine also includes corresponding customer interfaces and data storage. The demand management module is also a tracking module.

[0146] The customer service management engine of the disclosed embodiment has the following advantages: 1. Linkage between intent recognition and Agent technology: Construct an intent knowledge base, identify a single or multiple intents of the user through the intent recognition module, and vectorize the identified intent with the labeled intent of the knowledge base to avoid the spread of intent and perform accurate recognition; based on the user's intent, automatically call a suitable single or multiple Agents for processing, and improve the decision-making and processing capabilities of the system. 2. Introduce sentiment analysis to improve customer experience: Traditional customer service systems rarely pay attention to the emotional state of customers. The disclosed embodiment can dynamically adjust the service strategy according to the customer's emotions through the sentiment analysis module, give priority to handling the demands of customers with strong negative emotions, and significantly improve the user experience. 3. Agent dynamically loads fine-tuning model: After using the Agent, release the corresponding fine-tuning model to reduce the system load; when receiving a new Agent demand, load and apply the model fine-tuned by the latest self-optimization mechanism to ensure the improvement of Agent's capabilities and seamless switching of applications. 4. Self-optimization mechanism: The disclosed embodiment designs an adaptive learning mechanism to enable the system to continuously self-optimize according to the actual operation data and customer feedback, and continuously improve the processing capability and response accuracy.

[0147] On the other hand, an embodiment of the present disclosure proposes an electronic device, including: a processor, and a memory communicatively connected to the processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the method described above.

[0148] On the other hand, an embodiment of the present disclosure provides a computer-readable storage medium having a computer program stored thereon, and the computer program implements the steps of the above-mentioned method when executed by a processor.

[0149] Another aspect of the embodiments of the present disclosure provides a computer program product, including a computer program, which implements the steps of the method described above when executed by a processor.

[0150] The purpose of the disclosed embodiment is to provide a telecommunications customer service request management engine based on the fusion of deep NLP and intelligent agent technology, which realizes efficient automatic identification, classification and automatic processing of customer service requests and engine self-optimization by combining natural language processing and intelligent agent technology. Specifically, it has the following advantages:

[0151] 1) Efficient request processing: With the help of deep NLP technology, the system can quickly understand the customer's voice input and accurately identify the core content of the request. Through the automatic decision-making and task allocation of intelligent agents, the system's processing efficiency is significantly improved, and the frequency and time of manual intervention are reduced.

[0152] 2) Responding to complex demands: The integration of deep NLP and intelligent agents enables the system to flexibly respond to complex and diverse demands. The NLP module can identify subtle differences and hidden intentions in customer language, while the intelligent agent can make adaptive adjustments based on specific circumstances to effectively respond to complex or uncommon questions.

[0153] 3) Personalized service: The introduction of sentiment analysis module can analyze the emotional state of customers in real time and adjust the response strategy through intelligent agent, so as to provide more humane service and improve the overall customer satisfaction. For emotionally excited customers, the system will give priority to their demands to avoid escalation of problems.

[0154] 4) Self-optimization mechanism: Through the adaptive learning module, the system can continuously optimize the NLP model and Agent decision-making strategy based on customer feedback and historical data. This mechanism enables the system to have continuous evolution capabilities in actual applications and can continuously improve the response speed and accuracy to customer needs.

[0155] The above is only a preferred implementation of the embodiment of the present disclosure. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the embodiment of the present disclosure. These improvements and modifications should also be considered as the protection scope of the embodiment of the present disclosure.

Claims

1. A customer service processing method based on big data, characterized in that: The method comprises: Receiving a user demand, wherein the user demand includes: text or voice; Inputting the user demand into a pre-trained deep learning model for semantic analysis to extract at least one key intent and at least one contextual semantic information, wherein the pre-trained deep learning model is based on: BERT model training; Match the parsing results with the pre-built intent knowledge base to obtain at least one intent that meets the user's needs, fill the slots for each intent that meets the user's needs, and extract key information; According to each intention that meets the user's demands and the corresponding extracted specific information, the corresponding agent is called from the pre-obtained agent group to process the user's demands accordingly; Track and manage the overall process of handling user demands, and feedback the processing results to the client.

2. The method according to claim 1, characterized in that: The parsing results are matched with the pre-built intent knowledge base to obtain at least one intent that meets the user's needs, including: Calculate the similarity between the parsing result and each intent in the pre-built intent knowledge base to obtain a similarity set; Take the top k intent samples with the highest similarity and the corresponding analysis results from the self-similarity set, input them into the pre-trained deep learning model for judgment, and select at least one intent that meets the user's needs, where k ≥ 2.

3. The method according to claim 1, characterized in that The agent group includes multiple agents, each agent corresponding to processing a type of user intention; Before calling the Agent method, initialize the Agent, including: Use the reinforcement learning algorithm DQN to initialize each agent's strategy model and establish a corresponding strategy model for each agent; The method of calling a corresponding agent from a pre-obtained agent group according to at least one intention that meets the user's demands and corresponding extracted specific information to process the user's demands accordingly includes: Assign user demands to corresponding agents based on each intent that meets user demands and the specific information extracted; If there are multiple candidate agents for a user intent, the best agent is selected based on historical performance data or current workload; After calling the Agent, the method further includes: If the agent cannot solve the user problem in the process of handling the user's request, it will determine whether it exceeds the predetermined conditions. If so, the task will be upgraded to a manual customer or a preset senior agent.

4. The method according to any one of claims 1 to 3, characterized in that: The tracking and management of the overall process of handling user demands and feeding back the processing results to the client include: Track the entire process of each user's request, including request creation, processing progress and feedback results; specifically including: After analyzing the customer's request, a work order is automatically generated, which includes a detailed description of the customer's problem, the problem type, and the processing progress; Track the progress of each work order in real time and provide feedback to customers; After the work order is processed, the customer will be provided with result feedback. If the problem is solved, the work order will be closed; if not, the next step will be prompted.

5. The method according to any one of claims 1 to 3, characterized in that: After feeding back the processing result to the client, the method further includes: Collect customer interaction data; Regularly retrain deep learning models and agent strategy models based on historical data; Evaluate the performance of deep learning models BERT and agents to ensure performance improvements of optimized models in production environments.

6. The method according to any one of claims 1 to 3, characterized in that: The method further comprises: If the user's request includes voice, convert the voice into text; Based on the pre-trained sentiment analysis model, analyze the emotional state in the user's request text to determine whether the user is in a positive, negative or neutral emotion; According to the analysis results, the service strategy of the intelligent agent is adjusted. For customers with strong negative emotions, their problems will be handled first and given more detailed responses. During the entire processing process, the customer's emotional changes are continuously monitored and real-time feedback is provided to the agent or manual customer service to assist in decision-making.

7. A customer service processing device based on big data, characterized in that: The device comprises: A receiving module, configured to receive a user demand, wherein the user demand includes: text or voice; A parsing module is used to input the user's demands into a pre-trained deep learning model for semantic parsing, and extract at least one key intent and at least one contextual semantic information, wherein the pre-trained deep learning model is based on: BERT model training; The matching module is used to match the parsing results with the pre-built intent knowledge base to obtain at least one intent that meets the user's needs, fill the slots for each intent that meets the user's needs, and extract key information; The scheduling module is used to call the corresponding agent from the pre-obtained agent group according to each intention that meets the user's demands and the corresponding extracted specific information, and process the user's demands accordingly; The tracking module is used to track and manage the overall process of user demand processing and feed back the processing results to the client.

8. An electronic device, characterized in that: include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the method steps according to any one of claims 1 to 6.

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

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

Images