Question answering model training method, device, equipment and storage medium

Abstract

The embodiments of the present application provide a question-answering model training method, device, equipment and storage medium. The method includes: obtaining historical customer service information; constructing a thinking chain in the customer service processing process based on the historical customer service information; based on the thinking chain, selecting multiple customer service data from the data set for annotation to obtain annotated data; training the initial model based on the annotated data to obtain a trained question-answering model. The present application extracts key information from historical customer service information, and constructs a thinking chain in the customer service processing process based on this key information, and integrates the thinking chain into the training data of the model, ensuring that the training data matches the logical structure of the thinking chain, which can effectively guide the learning process of the model, deeply understand the customer service processing business rules, and improve the accuracy of the model's response; in addition, integrating the thinking chain into the training data of the model also allows the present application to be independent of a specific knowledge base, simplifying the deployment and maintenance costs of the model.

Description

Technical Field

[0001] The present application relates to the field of big data technology, and in particular to a question-answering model training method, apparatus, device and storage medium. Background Art

[0002] In traditional customer service and operations, employees need to face a lot of boring and repetitive work, such as making customer transfer work orders. Not only does it consume a lot of employees' time and energy, but it can also easily lead to fatigue, which in turn affects work efficiency and customer satisfaction. Generative AI big models have good intent understanding and content generation capabilities. The introduction of big models has improved the intelligence level of customer service scenarios, greatly improved the efficiency of financial services, and accelerated the digital transformation of enterprises.

[0003] In the process of building applications using large pre-trained language models. Usually, the model itself is not modified extensively. Instead, the model's potential is exploited through prompt engineering, data engineering, and business logic decomposition to adapt to specific application tasks. In practice, it is found that when large models are directly adopted, there are problems such as lack of expertise, model hallucination, and risk control. When fine-tuning the model, there are problems with data diversity, noise and outliers, uneven distribution, labeling quality, privacy security, etc., and it is difficult to deeply understand the business rules. Fine-tuning may overfit the data of new tasks, resulting in a decrease in the generalization ability of the model. The quality of the output depends largely on the relevance and accuracy of the retrieved information, which may lead to a high degree of dependence on specific data and reduce the accuracy of the model's response. Summary of the invention

[0004] The embodiments of the present application provide a question-answering model training method, apparatus, device and storage medium to solve the problem in traditional technologies that the quality of model output depends largely on the relevance and accuracy of the retrieved information, which may lead to a high dependence on specific data and reduce the accuracy of the model's responses.

[0005] In a first aspect, an embodiment of the present application provides a question-answering model training method, comprising:

[0006] Get historical customer service information;

[0007] Based on the historical customer service information, a thought chain in the customer service processing process is constructed;

[0008] Based on the thought chain, multiple customer service data are selected from the data set for annotation to obtain annotated data;

[0009] The initial model is trained according to the labeled data to obtain a trained question-answering model.

[0010] In a possible implementation, constructing a thought chain in the processing process according to the historical customer service information includes:

[0011] Acquire customer input information in the historical customer service information and context information corresponding to the customer input information;

[0012] The pre-trained thought chain generation model is used to parse the customer input information and the context information to obtain the thought chain in the processing process.

[0013] In a possible implementation, the pre-trained thought chain generation model is used to parse the customer input information and the context information to obtain the thought chain in the processing process, including:

[0014] The pre-trained thought chain generation model identifies the semantic dependency between the customer input information and the context information, predicts the customer question type according to the semantic dependency, generates an optimal processing solution according to the customer question type, and obtains the thought chain in the processing process.

[0015] In a possible implementation, the pre-trained thought chain generation model identifies the semantic dependency relationship between the customer input information and the context information, including:

[0016] The pre-trained thought chain generation model performs word segmentation processing on the customer input information and the context information to obtain multiple phrases;

[0017] Encode each phrase to obtain the semantic vector representation corresponding to each phrase;

[0018] Based on the semantic vector representation, obtaining the semantic dependency relationship between each word group and other word groups;

[0019] The semantic dependency relationship corresponding to each phrase constitutes the semantic dependency relationship between the customer input information and the context information.

[0020] In a possible implementation, predicting the type of customer question according to the semantic dependency relationship includes:

[0021] The pre-trained thought chain generation model matches the semantic dependency relationship with the predefined semantic features, calculates the probability distribution of different question types according to the matching results, and selects the type corresponding to the probability distribution that meets the preset requirements as the predicted customer question type.

[0022] In a possible implementation, generating an optimal solution according to the type of customer problem includes:

[0023] The pre-trained thought chain generation model searches for solutions corresponding to the customer's problem type from a preset solution library according to predefined mapping rules;

[0024] If there are multiple solutions, sort the solutions from front to back according to the preset rules, and select the solution with the first ranking as the optimal solution.

[0025] In a possible implementation, the pre-trained thought chain generation model is used to parse the customer input information and the context information to obtain the thought chain in the processing process, and further includes:

[0026] The pre-trained thought chain generation model analyzes the customer input information to obtain the customer's emotional state;

[0027] Generate an optimal solution based on the customer's problem type and emotional state.

[0028] In a possible implementation, generating an optimal solution according to the type of the customer's problem and the emotional state includes:

[0029] Prioritize issues based on said emotional state;

[0030] Generate an optimal solution based on the customer problem type and the problem priority.

[0031] In a possible implementation manner, after generating the optimal processing solution, the method further includes:

[0032] Create a work order according to a preset template and fill in key information into the created work order; wherein the key information includes customer input information, customer problem type and optimal processing solution.

[0033] In a possible implementation, training the initial model according to the labeled data to obtain a trained question-answering model includes:

[0034] Input customer service data into the initial model to obtain recognition results;

[0035] Compare the recognition result with the labeled data corresponding to the customer service data;

[0036] According to the comparison results, the network parameters of the initial model are gradually adjusted until the recognition results of the initial model meet the training requirements, and a trained question-answering model is obtained.

[0037] In a possible implementation, the method further includes:

[0038] During the training process, customer feedback data is obtained in real time, and the feedback data is added to the data set to form a new data set;

[0039] Select multiple customer service data from the new data set for annotation to obtain new annotated data;

[0040] Train the question-answering model based on the new labeled data.

[0041] In a second aspect, an embodiment of the present application provides a question-answering model training device, comprising:

[0042] Acquisition module, used to obtain historical customer service information;

[0043] A thought chain building module, used to build a thought chain in the customer service processing process according to the historical customer service information;

[0044] A labeling module, used for selecting a plurality of customer service data from the data set for labeling based on the thought chain to obtain labeling data;

[0045] The training module is used to train the initial model according to the labeled data to obtain a trained question-answering model.

[0046] In a third aspect, an embodiment of the present application provides a computer device, including: a memory, a processor;

[0047] The memory stores computer-executable instructions;

[0048] The processor executes the computer-executable instructions stored in the memory, so that the processor executes the above first aspect and / or various possible implementations of the first aspect.

[0049] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, in which computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, they are used to implement the first aspect above and / or various possible implementations of the first aspect.

[0050] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the above first aspect and / or various possible implementation methods of the first aspect.

[0051] The question-answering model training method, device, equipment and storage medium provided in the embodiments of the present application construct a chain of thoughts in the customer service processing process based on the acquired historical customer service information; then based on the chain of thoughts, multiple customer service data are selected from the data set for annotation to obtain annotated data; finally, the initial model is trained based on the annotated data to obtain a trained question-answering model. The present application extracts key information from historical customer service information, and constructs a chain of thoughts in the customer service processing process based on this key information, and integrates the chain of thoughts into the training data of the model, ensuring that the logical structure of the training data matches the chain of thoughts, which can effectively guide the learning process of the model, deeply understand the customer service processing business rules, and improve the accuracy of the model's responses; in addition, integrating the chain of thoughts into the training data of the model also allows the present application to be independent of a specific knowledge base, simplifying the deployment and maintenance costs of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0053] Figure 1 A schematic diagram of a scenario of a question-answering model training method in one embodiment;

[0054] Figure 2 A flowchart of a question-answering model training method in one embodiment;

[0055] Figure 3 A flowchart of steps for constructing a thought chain in an embodiment;

[0056] Figure 4 Schematic diagram of the structure of a question-answering model training device in one embodiment;

[0057] Figure 5 FIG. 1 is a schematic diagram of the structure of a computer device in an embodiment.

[0058] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0059] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0060] The question-answering model training method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown, the terminal 102 communicates with the server 104 through a network. The data storage system can store data that the server 104 needs to process. The data storage system can be integrated on the server 104, or placed on the cloud or other network servers.

[0061] Among them, the customer can upload the historical customer service information to the server 104 through the terminal 102, so that the server 104 builds the thought chain in the customer service processing process according to the historical customer service information; and based on the thought chain, selects multiple customer service data from the data set for annotation to obtain the annotated data; finally, the initial model is trained according to the annotated data to obtain the trained question-answering model. In this way, after receiving the question text submitted by the customer through the terminal 102, the server 104 uses the trained question-answering model to process the question text, outputs the solution, and feeds back the solution to the terminal 102 for display to the customer.

[0062] The above-mentioned question-answering model training method constructs a chain of thoughts in the customer service processing process based on the historical customer service information obtained; then, based on the chain of thoughts, multiple customer service data are selected from the data set for annotation to obtain annotated data; finally, the initial model is trained based on the annotated data to obtain a trained question-answering model. This application extracts key information from historical customer service information, and constructs a chain of thoughts in the customer service processing process based on this key information, and integrates the chain of thoughts into the training data of the model, ensuring that the logical structure of the training data matches the chain of thoughts, which can effectively guide the learning process of the model, deeply understand the customer service processing business rules, and improve the accuracy of the model's responses; in addition, integrating the chain of thoughts into the training data of the model also allows this application to be independent of a specific knowledge base, simplifying the deployment and maintenance costs of the model.

[0063] The terminal 102 may be, but is not limited to, various personal computers, laptop computers, smart phones, tablet computers, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers.

[0064] It should be noted that the network communication between the terminal 102 and the server 104 is applicable to different network standards, for example, it can be applicable to global system of mobile communication (GSM), code division multiple access (CDMA), wideband code division multiple access (WCDMA), time division synchronous code division multiple access (TD-SCDMA), long term evolution (LTE) system and future 5G and other network standards. Optionally, the above communication system can be a system in the scenario of ultra-reliable and low latency communications (URLLC) transmission in a 5G communication system.

[0065] Therefore, optionally, the above-mentioned base station can be a base station (Base Transceiver Station, referred to as BTS) and / or a base station controller in GSM or CDMA, or a base station (NodeB, referred to as NB) and / or a radio network controller (Radio Network Controller, referred to as RNC) in WCDMA, or an evolved base station (EvolutionalNode B, referred to as eNB or eNodeB) in LTE, or a relay station or access point, or a base station (gNB) in the future 5G network, etc., and the present application is not limited here.

[0066] The terminal 102 can be a wireless terminal or a wired terminal. A wireless terminal can be a device that provides voice and / or other business data connectivity to a customer, a handheld device with a wireless connection function, or other processing devices connected to a wireless modem. A wireless terminal can communicate with one or more core network devices via a radio access network (Radio Access Network, referred to as RAN). A wireless terminal can be a mobile terminal, such as a mobile phone (or "cellular" phone) and a computer with a mobile terminal, for example, a portable, pocket-sized, handheld, computer-built-in or vehicle-mounted mobile device, which exchanges language and / or data with a radio access network. For another example, a wireless terminal can also be a personal communication service (Personal Communication Service, referred to as PCS) phone, a cordless phone, a Session Initiation Protocol (SIP) phone, a Wireless Local Loop (Wireless Local Loop, referred to as WLL) station, a Personal Digital Assistant (Personal Digital Assistant, referred to as PDA) and other devices. A wireless terminal may also be referred to as a system, a subscriber unit (Subscriber Unit), a subscriber station (Subscriber Station), a mobile station (MobileStation), a mobile station (Mobile), a remote station (Remote Station), a remote terminal (Remote Terminal), an access terminal (Access Terminal), a client terminal (User Terminal), a client agent (User Agent), a client device (User Device or User Equipment), without limitation here.

[0067] In one embodiment, a question-answering model training method is provided. This embodiment uses the method applied to the server 104 as an example. Figure 2 As shown, the question-answering model training method includes:

[0068] Step 202: Obtain historical customer service information.

[0069] Specifically, historical customer service information is customer service data generated during the customer service process collected from various channels, including customer consultation records, customer service response content, problem solving process, customer feedback, etc. Information sources include public databases, customer service systems of various service providers, product documents, customer manuals, etc.

[0070] Furthermore, after collecting the customer service data, the present application also performs preprocessing operations on the customer service data to ensure the accuracy and consistency of the data. The preprocessing steps include removing invalid, duplicate or erroneous data, and converting the data format to a preset format.

[0071] Step 204: Construct a thought chain in the customer service process based on historical customer service information.

[0072] Specifically, the thought chain represents the whole process structure from problem raising, classification, processing to solution. By analyzing historical customer service information, the key nodes in the customer service process can be identified, and the dependencies between the key nodes and the impact of these key nodes on problem solving can be learned, thereby obtaining the thought recursion of the service process in the agent record. In addition, in addition to the global process, the thought chain also has specific steps at a certain node, so that in the subsequent model training, the model can learn to pay attention to the global background and local details at the same time, and improve the ability to understand the context.

[0073] Furthermore, the key to building a thinking chain is to clarify and define the problem you are trying to solve, which involves a deep understanding of the problem, including identifying the core elements, goals, and context of the problem. For example, if the goal is to improve the intelligence level of customer service, it is necessary to determine which customer interaction links can be optimized through automation. The output of the problem definition phase should be a clear problem statement, which will guide the entire thinking chain construction process. Once the problem is clear, the next task is to design a series of logical steps that should be able to simulate the thinking process of human experts when solving similar problems. This may include identifying the key aspects of the problem, exploring possible solutions, evaluating the pros and cons of different solutions, and finally deciding on the best path of action. Logical step design requires careful planning to ensure that each step is necessary and can logically and smoothly lead to the next step.

[0074] After designing the logical steps, you also need to determine what data is needed in each step to support the decision-making process. This involves retrieving specific information from a database, extracting key data from documents, or collecting real-time input from user interactions. The design of a data retrieval strategy should take into account the accessibility, relevance, and quality of the data to ensure that the collected data can effectively support subsequent analysis and decision-making.

[0075] After acquiring the data, it is necessary to compare and analyze the data. This step includes statistical analysis, pattern recognition, trend analysis, etc., with the goal of extracting valuable insights from the data. The goal of comparative analysis is to understand the relationship between different data points and how they affect the solution of the problem. This step is crucial to generating accurate and useful answers.

[0076] After conducting in-depth data analysis, the results need to be synthesized into a coherent answer or solution. Results synthesis is more than a simple summary of data, but requires turning the analysis results into actual decisions or actions. This may involve combining different data points and analysis results to form a comprehensive perspective, as well as predicting possible outcomes and impacts.

[0077] Finally, the constructed thought chain needs to be represented in a clear and executable way, such as through a flowchart, pseudocode, decision tree or other visualization tools. In addition, the representation of the thought chain should be detailed enough so that other personnel can understand and follow it. At the same time, it should also be flexible enough to facilitate adjustment and optimization in actual applications.

[0078] Step 206: Based on the thought chain, multiple customer service data are selected from the data set for annotation to obtain annotated data.

[0079] Specifically, the data set is collected in advance through multiple channels for training and learning the question-answering model. The data set can be selected from the historical customer service information in step 204, but the data in the data set is different from the data in the historical customer service information in step 204. The data set should include enough training samples to ensure that the question-answering model can learn the distribution and regularity of the data. Generally speaking, the number of annotations can be determined according to the needs of the project. For areas with more complex professional knowledge of the project, all training samples in the data set can be annotated, so that the model can learn more features and patterns, thereby improving its accuracy and generalization ability. For projects with a smaller budget, some training samples in the data set can be annotated, which can reduce the annotation cost, but insufficient annotations may cause the model to be unable to learn enough features and performance degradation. Furthermore, some data can be extracted for annotation first, and model training is performed based on this part of the data, and then the annotated data is gradually updated according to the feedback of the model. In this way, iterative annotation and model training can improve the accuracy of the model. Specifically in this embodiment, in order to save training time and cost, multiple customer service data are selected from the data set for annotation to obtain annotated data. The multiple customer service data selected in this way and the corresponding annotated data constitute the training data of the model.

[0080] It should be noted that in the embodiments of this application, a high-performance annotation large model is used to automatically annotate the selected customer service data, with the purpose of identifying and extracting key information in the customer service data, such as customer problem descriptions, related product information, etc., and classifying and organizing them according to the logical steps of the thinking chain. In order to achieve this goal, the annotation large model needs to have strong language understanding and information extraction capabilities. There are many large models for annotating text in traditional technologies, such as hidden Markov models, conditional random fields, recurrent neural networks, etc. In actual applications, you can choose according to your needs. This application does not limit the model type.

[0081] Before the initial annotation, a detailed annotation guide needs to be created to clarify the data type and format that needs to be annotated in each thinking chain step. This guide will serve as an operating guide for the annotation model to ensure that its output data annotation meets the expected standards. In addition, the annotation model needs to be properly trained and adjusted to ensure that it can accurately identify and extract key information.

[0082] After the initial annotation is completed, the annotation results need to be quality controlled. This may involve sampling the output of the large annotation model to assess its accuracy and consistency. If problems are found, the model parameters need to be adjusted or the model needs to be retrained in time to improve its annotation accuracy.

[0083] After introducing the thinking chain into the training data, the training samples can be labeled more accurately and specifically according to the structure of the thinking chain. For example, representative samples can be selected based on the distribution of samples in different process stages to cover various typical user scenarios and problem types. In addition, the model trained with the training data introduced with the thinking chain has a modular thinking framework that can automatically identify the customer's problem type, urgency and other information, and then proceed to the next step according to the thinking chain process.

[0084] Step 208: Train the initial model based on the labeled data to obtain a trained question-answering model.

[0085] Specifically, the multiple customer service data in the data set and the labeled data corresponding to each customer service data constitute the training data, and then the training data is randomly divided into training set, test set and validation set according to the preset ratio. Among them, the training set is used for model learning and participates in the update of model parameters; the validation set is used for model tuning to help select the optimal model and prevent overfitting; the test set is used for model evaluation to measure the model's generalization ability to new data. In this way, the training data is input into the initial model, and the model parameters are gradually adjusted according to the output of the initial model, and finally a trained question-answering model is obtained.

[0086] The above-mentioned question-answering model training method constructs a chain of thoughts in the customer service processing process based on the historical customer service information obtained; then, based on the chain of thoughts, multiple customer service data are selected from the data set for annotation to obtain annotated data; finally, the initial model is trained based on the annotated data to obtain a trained question-answering model. This application extracts key information from historical customer service information, and constructs a chain of thoughts in the customer service processing process based on this key information, and integrates the chain of thoughts into the training data of the model, ensuring that the logical structure of the training data matches the chain of thoughts, which can effectively guide the learning process of the model, deeply understand the customer service processing business rules, and improve the accuracy of the model's responses; in addition, integrating the chain of thoughts into the training data of the model also allows this application to be independent of a specific knowledge base, simplifying the deployment and maintenance costs of the model.

[0087] See also Figure 3 Optionally, based on historical customer service information, a thought chain in the processing process is constructed, including:

[0088] Step 302: Obtain customer input information and context information corresponding to the customer input information in historical customer service information.

[0089] Step 304: Use the pre-trained thought chain generation model to parse the customer input information and context information to obtain the thought chain in the processing process.

[0090] Specifically, the customer input information is the customer's problem description information, which may include a text description of the customer's problem, the product model consulted, the consultation date, the solution date, etc. The context information corresponding to the customer input information includes the customer service's reply to the problem, the final solution, customer feedback, customer purchase records, etc. The above customer input information and its corresponding context information are pre-stored in the customer service system database for retrieval.

[0091] The thought chain generation model used in the embodiment of the present application is obtained by pre-training based on historical customer service information. It should be noted that the data of the historical customer service information in step 304 is different from the data in the historical customer service information in step 204, that is, the first data in the historical customer service information is first used to train the thought chain generation model, and then the trained thought chain generation model is used to analyze the second data in the historical customer service information, thereby obtaining the thought chain in the customer service process of this embodiment. Further, in order to provide a more comprehensive and in-depth perspective for the thought chain generation model and enhance its ability to solve complex problems and provide high-quality services, the embodiment of the present application also introduces data from multiple fields as training data for the thought chain generation model based on customer service information, such as logic, psychology, economics, management, etc. In this way, the thought chain generation model can learn to give reasonable solutions through logical reasoning; it can also learn to understand the emotional state of customers and give personalized suggestions that can appease customers' emotions; in addition, it can also learn to compare prices and give customers reasonable suggestions.

[0092] Optionally, a pre-trained thought chain generation model is used to parse the customer input information and context information to obtain the thought chain in the processing process, including:

[0093] The pre-trained thought chain generation model identifies the semantic dependency between customer input information and context information, predicts the type of customer question based on the semantic dependency, and generates the optimal solution based on the type of customer question to obtain the thought chain in the processing process.

[0094] Specifically, semantic dependency relations characterize the logical relationships between different components, reflecting which other components a component depends on, as well as the semantic connections between the component and other components.

[0095] Optionally, the pre-trained thought chain generation model identifies semantic dependencies between customer input information and context information, including:

[0096] The pre-trained thought chain generation model performs word segmentation on the customer input information and context information to obtain multiple phrases.

[0097] Each phrase is encoded to obtain the semantic vector representation corresponding to each phrase.

[0098] Based on the semantic vector representation, the semantic dependency relationship between each word group and other words is obtained.

[0099] The semantic dependency relationship corresponding to each phrase constitutes the semantic dependency relationship between the customer input information and the context information.

[0100] Specifically, the customer input information can be in voice form or text form. If it is in voice form, it is first converted into text form. Further, the text form of the customer input information is preprocessed, including removing special characters, punctuation marks, etc. Further, a syntactic analyzer, such as Stanford Parser or spaCy, is used to perform word segmentation on the preprocessed customer input information and its corresponding context information to obtain multiple phrases, and the part of speech of each phrase is marked. Further, the model converts the text form into a vector form, and then calculates the dependency relationship between each phrase and other phrases in turn based on the syntactic analysis and part of speech tagging results, and finally obtains the semantic dependency relationship between the customer input information and the corresponding context information.

[0101] Optionally, predict the type of customer question based on semantic dependencies, including:

[0102] The pre-trained thought chain generation model matches semantic dependencies with predefined semantic features, calculates the probability distribution of different question types based on the matching results, and selects the type corresponding to the probability distribution that meets the preset requirements as the predicted customer question type.

[0103] Specifically, semantic features are obtained by analyzing a large amount of historical customer service information in advance, including keywords, phrases, sentence structures, sentiment tendencies, etc. of the questions, aiming to capture the uniqueness and commonalities of different types of questions.

[0104] The purpose of matching the extracted semantic dependency with the predefined semantic features is to find out which predefined question type features the input text matches, wherein the matching method includes calculating semantic similarity, searching for keywords, analyzing sentence structure, etc. In this embodiment, a classifier is used to match features and calculate the probability distribution of different question types. Generally speaking, the type with the highest probability can be selected as the prediction result; in the case where there are multiple types with similar probabilities, each type with a probability greater than a preset threshold can be used as the prediction result for the customer to choose.

[0105] Optionally, generate the best solution based on the type of customer problem, including:

[0106] The pre-trained thought chain generation model searches for solutions corresponding to customer problem types from a preset solution library based on predefined mapping rules.

[0107] If there are multiple solutions, sort the solutions from front to back according to the preset rules, and select the solution with the first ranking as the optimal solution.

[0108] Specifically, the preset solution library is constructed based on historical customer service information and business logic, which includes a variety of problem types and standardized solutions corresponding to each problem type. Furthermore, the solution library of the embodiment of the present application also constructs a vector index based on vector retrieval technology, so that the corresponding solution can be quickly found according to the problem type.

[0109] Optionally, a pre-trained thought chain generation model is used to parse the customer input information and context information to obtain the thought chain in the processing process, which also includes:

[0110] The pre-trained thought chain generation model analyzes the customer input information to obtain the customer's emotional state.

[0111] Generate the best solution based on the customer's problem type and emotional state.

[0112] Specifically, the thought chain generation model extracts emotion-related features from the text and classifies the extracted features based on the machine learning algorithm. Generally speaking, the emotion classification includes positive, negative or neutral, etc. In this way, the model can determine the processing method based on the intensity of the emotional state.

[0113] For example, if a customer leaves a message saying "My phone suddenly went black and can't be turned on. This has happened many times recently. I hope this problem can be solved as soon as possible", the thought chain generation model will identify the above text and describe the problem as "My phone went black and can't be turned on". It determines that the type of problem is "hardware problem" and the corresponding solution is "Please provide your location, and display nearby repair points based on the location information". Furthermore, the thought chain generation model recognizes that the customer expressed "it has happened many times" and classifies the customer's emotion as "negative". The optimal solution provided is "The customer is emotional. Please contact the customer as soon as possible. Please provide your location, and display nearby repair points based on the location information".

[0114] Optionally, generate the best solution based on the type of customer problem and emotional state, including:

[0115] Prioritize issues based on emotional state.

[0116] Generate the best solution based on customer problem type and problem priority.

[0117] Specifically, the sentiment classification includes positive, negative or neutral, among which negative represents a more intense emotional state, followed by neutral, and positive is the lowest. Setting a higher priority for customers with more emotional emotions can effectively improve customer satisfaction. For example, the priority can be set to high, medium, and low, corresponding to the negative, neutral, and positive sentiment classifications, respectively.

[0118] Optionally, after generating the optimal processing solution, the following steps are further included:

[0119] Create a work order based on the preset template and fill in key information into the created work order; the key information includes customer input information, customer problem type and optimal processing solution.

[0120] Specifically, in order to facilitate the display of the constructed thought chain, the embodiment of the present application also presets a variety of templates to adapt to various application scenarios in customer service. The key information is the data related to the key nodes of the processing flow generated in the process of the thought chain generation model parsing the text submitted by the customer, identifying the customer's problem description, judging the problem type, and finding the corresponding solution according to the problem type. Furthermore, the key information also includes information such as the customer's emotional state, priority, and historical message records.

[0121] Optionally, the initial model is trained according to the labeled data to obtain a trained question-answering model, including:

[0122] Input the customer service data into the initial model to obtain the recognition results.

[0123] Compare the recognition result with the labeled data corresponding to the customer service data.

[0124] According to the comparison results, the network parameters of the initial model are gradually adjusted until the recognition results of the initial model meet the training requirements, and a trained question-answering model is obtained.

[0125] Specifically, the model learns to understand the questions asked by customers and gives appropriate answers by comparing the predicted results with the labeled true values. Furthermore, the question-answering model also includes a loss function, which is used to calculate the difference between the predicted results and the labeled true values. The smaller the difference, the higher the accuracy of the model's prediction.

[0126] Optionally, the question-answering model training method of the present application further includes:

[0127] During the training process, customer feedback data is obtained in real time and added to the data set to form a new data set;

[0128] Select multiple customer service data from the new data set for annotation to obtain new annotated data;

[0129] Train the question-answering model based on the new labeled data.

[0130] Specifically, feedback data, including click-through rate, dwell time, user satisfaction survey results, etc., are collected and stored by the system to form feedback data. After preprocessing, these data will be integrated into the data set, which is a key resource for model training and optimization. In this way, the question-answering model built based on this method can continuously learn from user usage and improve itself. This self-improvement ability enables the system to adapt to changing user needs and preferences, thereby providing more personalized and efficient services, forming a positive cycle.

[0131] The above-mentioned question-answering model training method constructs a chain of thoughts in the customer service processing process based on the historical customer service information obtained; then, based on the chain of thoughts, multiple customer service data are selected from the data set for annotation to obtain annotated data; finally, the initial model is trained based on the annotated data to obtain a trained question-answering model. This application extracts key information from historical customer service information, and constructs a chain of thoughts in the customer service processing process based on this key information, and integrates the chain of thoughts into the training data of the model, ensuring that the logical structure of the training data matches the chain of thoughts, which can effectively guide the learning process of the model, deeply understand the customer service processing business rules, and improve the accuracy of the model's responses; in addition, integrating the chain of thoughts into the training data of the model also allows this application to be independent of a specific knowledge base, simplifying the deployment and maintenance costs of the model.

[0132] It should be understood that, although the steps in the flowcharts involved in the above embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0133] Based on the same inventive concept, the embodiment of the present application also provides a question-answering model training device for implementing the question-answering model training method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above-mentioned question-answering model training method, so the specific limitations in one or more device embodiments provided below can refer to the limitations of the question-answering model training method above, and will not be repeated here.

[0134] See also Figure 4In one embodiment, a question-answering model training device is provided, including: an acquisition module, a thought chain construction module, a labeling module and a training module.

[0135] Acquisition module, used to obtain historical customer service information.

[0136] The thought chain building module is used to build a thought chain in the customer service processing process according to the historical customer service information.

[0137] The labeling module is used to select multiple customer service data from the data set for labeling based on the thought chain to obtain labeled data.

[0138] The training module is used to train the initial model according to the labeled data to obtain a trained question-answering model.

[0139] Optionally, the thinking chain construction module constructs a thinking chain in the processing process based on historical customer service information, including: obtaining customer input information in the historical customer service information and context information corresponding to the customer input information; using a pre-trained thinking chain generation model to parse the customer input information and context information to obtain the thinking chain in the processing process.

[0140] Optionally, the thought chain building module uses a pre-trained thought chain generation model to parse the customer input information and context information to obtain the thought chain in the processing process, including:

[0141] The pre-trained thought chain generation model identifies the semantic dependency between customer input information and context information, predicts the type of customer question based on the semantic dependency, and generates the optimal solution based on the type of customer question to obtain the thought chain in the processing process.

[0142] Optionally, the pre-trained thought chain generation model identifies the semantic dependency relationship between customer input information and context information, including: the pre-trained thought chain generation model performs word segmentation processing on the customer input information and context information to obtain multiple phrases; encodes each phrase to obtain a semantic vector representation corresponding to each phrase; based on the semantic vector representation, obtains the semantic dependency relationship between each phrase and other phrases; the semantic dependency relationship corresponding to each phrase constitutes the semantic dependency relationship between the customer input information and the context information.

[0143] Optionally, the thought chain generation model predicts the type of customer question based on semantic dependencies, including: the pre-trained thought chain generation model matches the semantic dependencies with predefined semantic features, calculates the probability distribution of different question types based on the matching results, and selects the type corresponding to the probability distribution that meets preset requirements as the predicted customer question type.

[0144] Optionally, the thought chain generation model generates the optimal solution according to the customer problem type, including: the pre-trained thought chain generation model searches for solutions corresponding to the customer problem type from a preset solution library according to predefined mapping rules; if there are multiple solutions, sort the solutions from front to back according to preset rules, and select the solution ranked first as the optimal solution.

[0145] Optionally, the thought chain construction module adopts a pre-trained thought chain generation model to parse customer input information and context information to obtain the thought chain in the processing process, and also includes: the pre-trained thought chain generation model parses customer input information to obtain the customer's emotional state; generates the optimal processing plan according to the customer's problem type and emotional state.

[0146] Optionally, the thought chain building module generates an optimal solution according to the customer's problem type and emotional state, including: determining the problem priority according to the emotional state; generating an optimal solution according to the customer's problem type and problem priority.

[0147] Optionally, after the thinking chain construction module generates the optimal processing solution, it also includes: creating a work order according to a preset template, and filling key information into the created work order; wherein the key information includes customer input information, customer problem type and optimal processing solution.

[0148] Optionally, the training module trains the initial model according to the labeled data to obtain a trained question-answering model, including: inputting the customer service data into the initial model to obtain a recognition result; comparing the recognition result with the labeled data corresponding to the customer service data; and gradually adjusting the network parameters of the initial model according to the comparison result until the recognition result of the initial model meets the training requirements, thereby obtaining a trained question-answering model.

[0149] Optionally, the training module is also used to obtain customer feedback data in real time during the training process, add the feedback data to the data set to form a new data set; select multiple customer service data from the new data set for labeling to obtain new labeled data; and train the question-answering model based on the new labeled data.

[0150] The above-mentioned question-answering model training device constructs a chain of thoughts in the customer service processing process based on the historical customer service information obtained; then based on the chain of thoughts, multiple customer service data are selected from the data set for annotation to obtain annotated data; finally, the initial model is trained based on the annotated data to obtain a trained question-answering model. This application extracts key information from historical customer service information, and constructs a chain of thoughts in the customer service processing process based on this key information, and integrates the chain of thoughts into the training data of the model, ensuring that the logical structure of the training data matches the chain of thoughts, which can effectively guide the learning process of the model, deeply understand the customer service processing business rules, and improve the accuracy of the model's responses; in addition, integrating the chain of thoughts into the training data of the model also allows this application to be independent of a specific knowledge base, simplifying the deployment and maintenance costs of the model.

[0151] Each module in the above device can be implemented in whole or in part by software, hardware or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each module above.

[0152] See also Figure 5 , Figure 5 The figure is a schematic diagram of the structure of the computer device provided by the present application. The computer device 50 provided in this embodiment includes: at least one processor 501 and a memory 502. Optionally, the computer device 50 also includes a communication component 503. The processor 501, the memory 502 and the communication component 503 are connected via a bus.

[0153] In a specific implementation process, at least one processor 501 executes the computer-executable instructions stored in the memory 502, so that at least one processor 501 executes the above method.

[0154] The specific implementation process of the processor 501 can be found in the above-mentioned question-answer model training method embodiment. Its implementation principle and technical effects are similar, and will not be repeated here in this embodiment.

[0155] In the above embodiments, it should be understood that the processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the invention may be directly implemented as being executed by a hardware processor, or may be executed by a combination of hardware and software modules in the processor.

[0156] The memory may include a high-speed memory (Random Access Memory, RAM), and may also include a non-volatile memory (Non-volatile Memory, NVM), such as at least one disk memory.

[0157] The bus may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus in the drawings of the present application is not limited to only one bus or one type of bus.

[0158] The present application also provides a computer program product, including a computer program, which implements the above-mentioned question-answering model training method when executed by a processor.

[0159] The present application also provides a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the above-mentioned question-and-answer model training method is implemented.

[0160] The above-mentioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The readable storage medium can be any available medium that can be accessed by a general or special-purpose computer.

[0161] An exemplary readable storage medium is coupled to a processor so that the processor can read information from the readable storage medium and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can be located in an application specific integrated circuit (Application Specific Integrated Circuits, referred to as: ASIC). Of course, the processor and the readable storage medium can also exist in the device as discrete components.

[0162] The division of units is only a logical function division, and there may be other divisions in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or unit, which can be electrical, mechanical or other forms.

[0163] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0164] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0165] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0166] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk and other media that can store program codes.

[0167] Finally, it should be noted that those skilled in the art will readily conceive of other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses or adaptations of the present invention, which follow the general principles of the present invention and include common knowledge or customary technical means in the art not disclosed by the present invention, are not limited to the precise structure described above and shown in the drawings, and may be modified and changed in various ways without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. A question-answering model training method, characterized in that: The method comprises: Get historical customer service information; Based on the historical customer service information, a thought chain in the customer service processing process is constructed; Based on the thought chain, multiple customer service data are selected from the data set for annotation to obtain annotated data; The initial model is trained according to the labeled data to obtain a trained question-answering model.

2. The method according to claim 1, characterized in that The constructing of a thought chain in the processing process according to the historical customer service information includes: Acquire customer input information in the historical customer service information and context information corresponding to the customer input information; The pre-trained thought chain generation model is used to parse the customer input information and the context information to obtain the thought chain in the processing process.

3. The method according to claim 2, characterized in that The pre-trained thought chain generation model is used to parse the customer input information and the context information to obtain the thought chain in the processing process, including: The pre-trained thought chain generation model identifies the semantic dependency between the customer input information and the context information, predicts the customer question type according to the semantic dependency, generates an optimal processing solution according to the customer question type, and obtains the thought chain in the processing process.

4. The method according to claim 3, characterized in that: The pre-trained thought chain generation model identifies the semantic dependency between the customer input information and the context information, including: The pre-trained thought chain generation model performs word segmentation processing on the customer input information and the context information to obtain multiple phrases; Encode each phrase to obtain the semantic vector representation corresponding to each phrase; Based on the semantic vector representation, obtaining the semantic dependency relationship between each word group and other word groups; The semantic dependency relationship corresponding to each phrase constitutes the semantic dependency relationship between the customer input information and the context information.

5. The method according to claim 3, characterized in that: The predicting the type of customer question according to the semantic dependency relationship includes: The pre-trained thought chain generation model matches the semantic dependency relationship with predefined semantic features, calculates the probability distribution of different question types according to the matching results, and selects the type corresponding to the probability distribution that meets the preset requirements as the predicted customer question type.

6. The method according to claim 3, characterized in that Generating the optimal solution according to the type of customer problem includes: The pre-trained thought chain generation model searches for solutions corresponding to the customer's problem type from a preset solution library according to predefined mapping rules; If there are multiple solutions, sort the solutions from front to back according to the preset rules, and select the solution with the first ranking as the optimal solution.

7. The method according to claim 3, characterized in that The method adopts a pre-trained thought chain generation model to parse the customer input information and the context information to obtain a thought chain in the processing process, and further includes: The pre-trained thought chain generation model analyzes the customer input information to obtain the customer's emotional state; Generate an optimal solution based on the customer's problem type and emotional state.

8. The method according to claim 7, characterized in that Generating an optimal solution according to the type of the customer's problem and the emotional state includes: Prioritize issues based on said emotional state; Generate an optimal solution based on the customer problem type and the problem priority.

9. The method according to claim 7, characterized in that: After generating the optimal treatment plan, it also includes: Create a work order according to a preset template and fill in key information into the created work order; wherein the key information includes customer input information, customer problem type and optimal processing solution.

10. The method according to claim 1, characterized in that The initial model is trained according to the labeled data to obtain a trained question-answering model, including: Input customer service data into the initial model to obtain recognition results; Compare the recognition result with the labeled data corresponding to the customer service data; According to the comparison results, the network parameters of the initial model are gradually adjusted until the recognition results of the initial model meet the training requirements, and a trained question-answering model is obtained.

11. The method according to claim 10, characterized in that The method further comprises: During the training process, customer feedback data is obtained in real time, and the feedback data is added to the data set to form a new data set; Select multiple customer service data from the new data set for annotation to obtain new annotated data; Train the question-answering model based on the new labeled data.

12. A question-answering model training device, characterized in that: include: Acquisition module, used to obtain historical customer service information; A thought chain building module, used to build a thought chain in the customer service processing process according to the historical customer service information; A labeling module, used for selecting a plurality of customer service data from the data set for labeling based on the thought chain to obtain labeling data; The training module is used to train the initial model according to the labeled data to obtain a trained question-answering model.

13. A computer device, characterized in that: include: Memory, processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory, so that the processor performs the method according to any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 11 when executed by a processor.

15. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 11 when being executed by a processor.

Images