Problem integration method and device, storage medium and computer equipment

Abstract

According to the problem integration method and device, the storage medium and the computer equipment provided by the invention, a to-be-integrated problem is acquired, key feedback information in the to-be-integrated problem is identified, and a target service node is determined according to the key feedback information; according to the target service node, obtaining a target prompt template corresponding to the to-be-integrated question, and combining the key feedback information with the target prompt template to generate a prompt text corresponding to the to-be-integrated question; inputting prompt texts corresponding to the to-be-integrated questions into the question classification large language model to obtain target question types of the to-be-integrated questions; determining a high-dimensional vector of the to-be-integrated problem, and determining a high-dimensional vector of each problem library in the target problem type; and calculating the vector similarity between the high-dimensional vector of the to-be-integrated problem and the high-dimensional vector of each problem library, and determining a target problem library of the to-be-integrated problem in each problem library according to each vector similarity. Therefore, manual participation can be remarkably reduced, the possibility of human errors is reduced, and the problem processing efficiency is improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a problem integration method, device, storage medium and computer equipment. Background Art

[0002] With the rapid development of e-commerce, online shopping has become an indispensable part of consumers' daily lives. However, the number of customer feedback issues is also increasing, covering many aspects from product quality, logistics and distribution to after-sales service. E-commerce platforms receive a large amount of customer feedback every day. These feedbacks are not only of various types, but also of different urgency and handling methods. Therefore, how to efficiently classify and summarize these feedback issues has become the key to improving customer satisfaction and operational efficiency for e-commerce platforms.

[0003] Currently, many e-commerce platforms still rely on manual methods to handle customer feedback issues. Customer service staff usually manually classify customer feedback according to preset categories and group similar issues together. However, as the number of customers and the amount of feedback continue to increase, this traditional processing method has gradually exposed the problem of inefficiency. Summary of the invention

[0004] The purpose of this application is to solve at least one of the above-mentioned technical deficiencies, especially the technical deficiencies of low efficiency in the prior art.

[0005] In a first aspect, the present application provides a problem integration method, the method comprising:

[0006] Obtain the issues to be integrated, identify key feedback information in the issues to be integrated, and determine the target business nodes corresponding to the issues to be integrated based on the key feedback information;

[0007] According to the target business node corresponding to the problem to be integrated, a target prompt template corresponding to the problem to be integrated is obtained, and the key feedback information of the problem to be integrated and the target prompt template are combined to generate a prompt text corresponding to the problem to be integrated;

[0008] Input the prompt text corresponding to the question to be integrated into the pre-trained question classification language model to obtain the target question type of the question to be integrated;

[0009] Determine a high-dimensional vector of the problems to be integrated, and determine a high-dimensional vector of each problem bank in the target problem type;

[0010] The vector similarity between the high-dimensional vector of the problem to be integrated and the high-dimensional vector of each question bank is calculated, and the target question bank of the problem to be integrated is determined in each question bank according to the vector similarities.

[0011] In one embodiment, the step of identifying key feedback information in the problem to be integrated includes:

[0012] A self-supervised learning model is used to determine the attention weight and semantic information of each word in the question to be integrated;

[0013] According to the attention weight of each word in the problem to be integrated, each word in the problem to be integrated is screened;

[0014] According to the semantic information corresponding to each filtered word, key feedback information is extracted from the questions to be integrated.

[0015] In one embodiment, the step of determining the target service node corresponding to the problem to be integrated according to the key feedback information includes:

[0016] Obtain feedback information corresponding to multiple business nodes;

[0017] Match the key feedback information with the feedback information corresponding to each business node, and use the business nodes with high matching degree as the target business nodes for the problems to be integrated.

[0018] In one embodiment, the step of obtaining a target prompt template corresponding to the problem to be integrated according to the target service node corresponding to the problem to be integrated includes:

[0019] According to the target business node of the problem to be integrated, querying the prompt template to be selected corresponding to the target business node from the pre-established prompt template library;

[0020] If there are multiple prompt templates to be selected, the matching degree between the problem to be integrated and each prompt template to be selected is calculated, and the target prompt template is selected from each prompt template to be selected according to each matching degree.

[0021] In one embodiment, the training process of the large language model for question classification includes:

[0022] obtaining a plurality of sample questions and determining a standard question type for each sample question;

[0023] Each sample question is used to train the large language model, and the standard question type of each sample question is used to evaluate the accuracy of the trained large language template until the training round reaches the preset threshold to obtain the question classification large language model.

[0024] In one embodiment, the step of determining a target question library of questions to be integrated in each question library according to the similarities of each vector includes:

[0025] The question library with the highest vector similarity is used as the target question library for the questions to be integrated.

[0026] In one embodiment, the high-dimensional vector of each question library in the target question type is stored in Milvus' efficient vector database.

[0027] In a second aspect, the present application provides a question integration device, the device comprising:

[0028] A target business node determination module is used to obtain the problem to be integrated, identify key feedback information in the problem to be integrated, and determine the target business node corresponding to the problem to be integrated according to the key feedback information;

[0029] A prompt text generation module is used to obtain a target prompt template corresponding to the problem to be integrated according to the target business node corresponding to the problem to be integrated, and combine the key feedback information of the problem to be integrated with the target prompt template to generate a prompt text corresponding to the problem to be integrated;

[0030] A target question type determination module is used to input the prompt text corresponding to the question to be integrated into a pre-trained question classification language model to obtain the target question type of the question to be integrated;

[0031] A high-dimensional vector determination module is used to determine the high-dimensional vector of the problem to be integrated and the high-dimensional vector of each problem library in the target problem type;

[0032] The target question library determination module is used to calculate the vector similarity between the high-dimensional vector of the problem to be integrated and the high-dimensional vector of each question library, and determine the target question library of the problem to be integrated in each question library based on the vector similarities.

[0033] In a third aspect, the present application provides a storage medium: the storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the problem integration method in any of the above embodiments.

[0034] In a fourth aspect, the present application provides a computer device, comprising: one or more processors, and a memory;

[0035] The memory stores computer-readable instructions, and when the computer-readable instructions are executed by one or more processors, the steps of the problem integration method in any of the above embodiments are performed.

[0036] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0037] In the problem integration method, device, storage medium and computer equipment provided by the present application, by identifying the key feedback information in the problem and generating prompt text based on the predefined business node template, this structured processing method makes the problem classification more accurate and consistent, avoiding classification errors caused by manual classification; by inputting the prompt text into the pre-trained problem classification large language model, the target type of the problem can be automatically identified by the learning ability of the model, and the large language model can identify potential and complex feedback information. Even if the feedback content is vague, the problem type can be inferred by the model; first determine the problem type and then determine the target problem library, which effectively reduces the amount of calculation and improves the processing efficiency; use high-dimensional vectors to express the characteristics of the problem, and identify the target problem library to which the problem belongs by calculating the vector similarity. It can accurately find similar problems by measuring the similarity between different problems, and can also dynamically adapt to multi-dimensional feature changes, enhancing the accuracy and extensiveness of problem matching. In this way, by introducing an automated key feedback identification and classification mechanism, manual participation can be significantly reduced, the possibility of human error can be reduced, and the efficiency of problem handling can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0039] Figure 1 A flowchart of a problem integration method provided in an embodiment of the present application;

[0040] Figure 2 A schematic diagram of the structure of a problem integration device provided in an embodiment of the present application;

[0041] Figure 3 A schematic diagram of the internal structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0043] The present application provides a method for problem integration. The following embodiments are described by taking the method applied to a computer device as an example. It can be understood that the computer device can be any device with data processing functions, including but not limited to a single server, a server cluster, a personal laptop, a desktop computer, etc. Figure 1 As shown, the method may include the following steps:

[0044] S101: Acquire the problem to be integrated, identify key feedback information in the problem to be integrated, and determine the target business node corresponding to the problem to be integrated according to the key feedback information.

[0045] Among them, the issues to be integrated refer to customer feedback or issues that need to be processed, classified, or integrated. They can be questions, complaints, or suggestions submitted by customers through various channels. Key feedback information refers to the core content or important information extracted from customer feedback. The target business node refers to the specific business module or process related to the feedback issue. Each business node usually corresponds to a specific business category, such as orders, payments, logistics, etc.

[0046] In this step, feedback questions submitted by customers, that is, questions to be integrated, can be automatically collected from various input channels through API interfaces, database queries, or message queues. Then, algorithms such as natural language processing can be used to analyze the text content of the questions to be integrated. For example, first, the feedback on the questions to be integrated is decomposed into words or phrases through techniques such as word segmentation and part-of-speech tagging; then the keyword extraction algorithm is used to identify important information in the questions to be integrated. Once the key feedback information is extracted, the information can be mapped to a target business node by matching predefined business rules or using a classifier model. For example, this can be done through a machine learning model, which automatically classifies the questions and assigns them to the appropriate business modules by analyzing the relationship between key feedback information and different business nodes.

[0047] It is understandable that customer feedback usually contains redundant information. Directly processing raw data will increase computational complexity and affect the accuracy of problem classification. By identifying and extracting key feedback information, the most meaningful content can be extracted from a large amount of text, reducing the processing scope, making subsequent business node matching and problem classification more accurate, and reducing the risk of misclassification caused by too much information.

[0048] S102: According to the target business node corresponding to the problem to be integrated, a target prompt template corresponding to the problem to be integrated is obtained, and key feedback information of the problem to be integrated and the target prompt template are combined to generate a prompt text corresponding to the problem to be integrated.

[0049] The target prompt template refers to a predefined template for a specific business node, which is used to generate prompt text for a question.

[0050] In this step, first, the corresponding prompt template database or configuration file is queried through the determined target business node. Each business node corresponds to one or more pre-defined templates, and the matching template can be quickly located by using the business node identifier as a query condition. Afterwards, the extracted key feedback information is inserted into the target prompt template using a template engine or string replacement technology. For example, first, placeholders or variable identifiers are found in the template, and then the key feedback information is inserted into these placeholders to form a complete prompt text.

[0051] It is understandable that each business node corresponds to different types of feedback questions, so it is necessary to predefine a suitable prompt template for each type. Template design can ensure the consistency of the expression of the question and reduce the ambiguity or format confusion caused by free expression. The combination of templates and feedback information is to simplify the expression of complex questions and format the questions uniformly. Key feedback information may be scattered or lengthy. By combining it with a template, the core content of customer feedback can be extracted and the information transmission can be ensured to have a consistent structure.

[0052] S103: Input the prompt text corresponding to the question to be integrated into a pre-trained large language model for question classification to obtain a target question type for the question to be integrated.

[0053] Among them, the pre-trained question classification large language model is a large model based on natural language processing technology. It is learned through a large amount of training data and can classify questions based on the input text and determine which question category the feedback belongs to. The target question type refers to the classification result output by the large language model based on the prompt text, that is, the specific category of the feedback question. This category usually corresponds to the preset business classification, such as technical issues, payment issues, logistics issues, etc.

[0054] In this step, you first prepare the generated prompt text after processing, and then you can input the text into the pre-trained large language model through API calls or directly embedded running of the question classification large language model. For example, you can use RESTful API, gRPC, or embedded reasoning service to pass text data. The question classification large language model will process the input prompt text, extract features and perform reasoning. The input data can then undergo necessary preprocessing, such as word segmentation, removal of invalid characters, and conversion to the format required by the model.

[0055] After model processing, the inference output of the model is obtained, which can be one or more classification labels. The large language model analyzes the semantic structure in the text and combines the classification knowledge learned during training to output the classification result that the question may correspond to. In one example, the output target question type can be a classification label with confidence. The final target question type is selected based on the highest confidence, or a multi-label strategy is used for processing.

[0056] It can be understood that the large language model has powerful text comprehension capabilities. It can automatically complete question classification by analyzing the semantic relationships in the prompt text, reduce the need for manual classification, and is efficient when processing massive amounts of data.

[0057] S104: Determine the high-dimensional vector of the problem to be integrated, and determine the high-dimensional vector of each question library in the target question type.

[0058] Among them, high-dimensional vector refers to the process of converting text into a mathematical representation. This vector captures the semantic information of the text in a multi-dimensional space, and each dimension corresponds to a certain feature of the text. The question library refers to a stored, classified set of questions. Each question library corresponds to a question type and contains multiple questions related to that type.

[0059] In this step, the questions to be integrated first need to be converted from natural language text into processable numerical representations. In one example, a pre-trained language model or vectorization tool can be used to convert the question text into a high-dimensional vector. Specifically, by calling these models, the questions to be integrated are input into the embedding layer of the model, and the model outputs the corresponding high-dimensional vector. The questions in each question bank can be processed by similar vectorization to convert each question in the question bank into a high-dimensional vector. For each question bank corresponding to the target question type, the vectorization model is called to convert these questions one by one into high-dimensional vectors and store them as a vector set.

[0060] It is understandable that natural language has a lot of ambiguity and ambiguity, and directly processing text data can easily lead to inaccurate matching. By converting text into high-dimensional vectors, the device can capture richer semantic information. By representing problems with high-dimensional vectors, it can automatically identify and process a large amount of complex feedback, reducing the need for manual intervention.

[0061] S105: Calculate the vector similarity between the high-dimensional vector of the problem to be integrated and the high-dimensional vector of each question library, and determine the target question library of the problem to be integrated in each question library according to the vector similarities.

[0062] Among them, vector similarity refers to calculating the similarity between two high-dimensional vectors.

[0063] In this step, we first need to obtain the high-dimensional vectors of the questions to be integrated and the high-dimensional vectors of each question bank. We can use vector similarity calculation methods to measure the similarity between these vectors, such as cosine similarity, Euclidean distance, etc. After that, we can compare the obtained similarity value with the preset threshold or sorting standard to determine the target question bank that best matches the questions to be integrated.

[0064] In the above embodiment, by identifying the key feedback information in the problem and generating prompt text based on the predefined business node template, this structured processing method makes the problem classification more accurate and consistent, avoiding classification errors caused by manual classification; by inputting the prompt text into the pre-trained problem classification large language model, the target type of the problem can be automatically identified by the learning ability of the model. The large language model can identify potential and complex feedback information, and the problem type can be inferred through the model even if the feedback content is vague; first determine the problem type and then determine the target problem library, which effectively reduces the amount of calculation and improves the processing efficiency; use high-dimensional vectors to express the characteristics of the problem, and identify the target problem library to which the problem belongs through vector similarity calculation, which can accurately find similar problems by measuring the similarity between different problems, and can also dynamically adapt to multi-dimensional feature changes, enhancing the accuracy and extensiveness of problem matching. In this way, by introducing an automated key feedback identification and classification mechanism, manual participation can be significantly reduced, the possibility of human error can be reduced, and the efficiency of problem handling can be improved.

[0065] In one embodiment, the step of identifying key feedback information in the problem to be integrated includes:

[0066] A self-supervised learning model is used to determine the attention weight and semantic information of each word in the question to be integrated;

[0067] According to the attention weight of each word in the problem to be integrated, each word in the problem to be integrated is screened;

[0068] According to the semantic information corresponding to each filtered word, key feedback information is extracted from the questions to be integrated.

[0069] Among them, the self-supervised learning model is a machine learning method that trains the model by predicting part of the information of the input data, without the need for manual labels. In the natural language processing model, the attention mechanism is used to calculate the importance of each word when generating representations. The attention weight indicates the contribution of each word to the model output and can highlight the key information in the text. Semantic information refers to the meaning and content expressed in the text. The semantic information of each word indicates the specific meaning of the word in the context and is used to understand the meaning of the entire sentence.

[0070] Specifically, a pre-trained self-supervised learning model, such as BERT, GPT, etc., is used to input the text of the problem to be integrated into the model. The model calculates the attention weight of each word through the self-attention mechanism, that is, the degree of attention each word pays to other words when generating representations. In one example, the text of the problem to be integrated is converted into the input format required by the model; the model calculates the weight of each word to other words through a multi-layer self-attention mechanism. These weights reflect the importance of each word in the context; the model calculates the contextual representation of each word through a deep neural network and generates a word vector containing semantic information.

[0071] Afterwards, the attention weights can be used to sort or filter the words in the problem to be integrated. In one example, the attention weights of all words are sorted from high to low to determine which words are more important in the context; based on a set threshold or the weights of the top N words, important words are selected for further processing.

[0072] Finally, the semantic information of the filtered important words is used to extract key feedback. In one example, the semantic vector is used to find the specific semantic representation of the filtered words; based on the semantic information of these important words, the key feedback information of the problem to be integrated is constructed.

[0073] In this embodiment, the self-supervised learning model can capture the deep semantic relationship between words in context. By calculating the attention weight and semantic information, it can more accurately understand the role and importance of each word in the sentence, which helps to improve the quality and accuracy of text processing. By filtering out high-weight words, you can focus on the most critical information, reduce the interference of irrelevant information, and improve processing efficiency. Screening out important words can more accurately extract core feedback information and ensure that key issues are effectively handled.

[0074] In one embodiment, the step of determining the target service node corresponding to the problem to be integrated according to the key feedback information includes:

[0075] Obtain feedback information corresponding to multiple business nodes;

[0076] Match the key feedback information with the feedback information corresponding to each business node, and use the business nodes with high matching degree as the target business nodes for the problems to be integrated.

[0077] Specifically, feedback information of each business node can be retrieved from the system. For example, a database or data source storing feedback information of business nodes is accessed; relevant feedback information is extracted from each business node according to preset query conditions, which may include historical feedback records, problem solving cases, customer evaluations, etc.; the extracted feedback information is sorted and classified to facilitate subsequent matching operations.

[0078] Afterwards, the key feedback information can be compared with the feedback information of each business node. For example, the key feedback information and the business node feedback information are preprocessed, including word segmentation, removal of stop words, etc.; the text similarity calculation method is used to calculate the matching degree between the key feedback information and the feedback information of each business node; the calculated matching degrees are sorted to determine which business node has the highest matching degree with the key feedback information.

[0079] Finally, the most appropriate business node can be selected based on the matching degree. For example, the business node with the highest matching degree is selected as the target node, and the problem to be integrated is assigned to the determined target business node to ensure that the problem can be handled correctly.

[0080] In this embodiment, by obtaining feedback information from each service node, we can fully understand the problem handling capabilities and history of each node, which helps to determine which node is most suitable for handling the current problem to be integrated. By matching key feedback information with feedback information from service nodes, we can ensure that the problem is assigned to the node that can best solve the problem. Selecting a service node with a high degree of matching can ensure that the problem is handled more quickly and accurately.

[0081] In one embodiment, the step of obtaining a target prompt template corresponding to the problem to be integrated according to the target service node corresponding to the problem to be integrated includes:

[0082] According to the target business node of the problem to be integrated, querying the prompt template to be selected corresponding to the target business node from the pre-established prompt template library;

[0083] If there are multiple prompt templates to be selected, the matching degree between the problem to be integrated and each prompt template to be selected is calculated, and the target prompt template is selected from each prompt template to be selected according to each matching degree.

[0084] Specifically, first, the database storing the prompt templates can be accessed, and the characteristics of the target business node can be used as query conditions to filter out the prompt templates related to the target business node from the template library. If there are multiple prompt templates to be selected, the matching degree between the problem to be integrated and each prompt template to be selected is calculated. For example, the problem to be integrated and each prompt template to be selected are preprocessed with text, such as word segmentation, stop word removal, etc., and the text similarity calculation method is used to calculate the matching degree between the problem to be integrated and each prompt template to be selected, and the calculated matching value is recorded to form a matching degree list of the prompt templates to be selected. Finally, according to each matching degree, the target prompt template is selected from each prompt template to be selected. For example, the prompt templates to be selected are sorted according to the calculated matching degree, and the templates with high matching degree are ranked in front, and the template with the highest matching degree is selected as the target prompt template.

[0085] In this embodiment, by associating the query condition with the target business node, it is possible to ensure that the template extracted from the prompt template library matches the business requirements of the problem to be integrated, thereby improving the efficiency and accuracy of problem processing. Calculating the matching degree can determine which of the selected prompt templates is most suitable for processing the current problem to be integrated. Selecting the template with the highest matching degree as the target prompt template can ensure that the response or solution provided subsequently best meets the requirements of the actual problem and enhances the processing effect.

[0086] In one embodiment, the training process of the large language model for question classification includes:

[0087] obtaining a plurality of sample questions and determining a standard question type for each sample question;

[0088] Each sample question is used to train the large language model, and the standard question type of each sample question is used to evaluate the accuracy of the trained large language template until the training round reaches the preset threshold to obtain the question classification large language model.

[0089] Sample questions are instance questions used to train the model, which can be examples collected from real data to help the model learn how to deal with similar problems. Standard question types are predefined categories or labels used to identify the types of sample questions, such as technical questions, service questions, return questions, etc.

[0090] Specifically, first, multiple sample questions can be extracted from a database or data source. These questions should cover a wide range of scenarios and types. A standard question type is assigned to each sample question. The sample questions and their standard question types are organized into a data format suitable for model training, such as a comparison table or training set. The preprocessed sample questions and their standard question types are input into the large language model to start the training process. These data are used to adjust the parameters of the model so that it can better predict the question type. In each training round, the parameters of the model are updated according to the training data and the target type to gradually improve the accuracy.

[0091] In each round, obtain the prediction results of the model, compare the prediction results of the model with the actual standard question type, calculate accuracy indicators such as precision, recall and F1 value, and according to the accuracy evaluation results, it may be necessary to adjust the model's hyperparameters or further train to improve model performance. After each round of training, check whether the current training round has reached the preset threshold. When the training round reaches the preset threshold, stop the training process and finally get a fully trained large language model.

[0092] In this embodiment, through training, the model can learn the mapping relationship from sample questions to standard question types, and gradually improve the classification accuracy. Accuracy evaluation can verify whether the model has achieved the expected performance level and ensure the reliability of the model in practical applications. Setting a threshold for training rounds can prevent overtraining, avoid overfitting the model on the training data, and ensure that its performance on unseen data is also good enough.

[0093] In one embodiment, the step of determining a target question library of questions to be integrated in each question library according to the similarities of each vector includes:

[0094] The question library with the highest vector similarity is used as the target question library for the questions to be integrated.

[0095] Specifically, based on the calculated similarity values, all question banks are sorted from high to low in terms of similarity. The question bank with the highest similarity is selected as the target question bank, and the questions to be integrated are assigned to the selected target question bank.

[0096] In this embodiment, selecting the question library with the highest similarity means allocating the questions to be integrated to the most appropriate library, thereby improving the efficiency and accuracy of question processing.

[0097] In one embodiment, the high-dimensional vectors of each question bank in the target question type are stored in Milvus' efficient vector database.

[0098] In this embodiment, Milvus is an open source vector database for efficient storage and retrieval of high-dimensional vector data. Milvus optimizes the similarity search of high-dimensional vectors and can quickly find the target question library that is most similar to the question vector to be integrated, thereby improving classification accuracy and optimizing the problem processing process.

[0099] The following describes the problem integration device provided in the embodiment of the present application. The problem integration device described below and the problem integration method described above can be referred to in correspondence with each other. Figure 2 As shown, the present application provides a problem integration device, which includes:

[0100] The target business node determination module 201 is used to obtain the problem to be integrated, identify the key feedback information in the problem to be integrated, and determine the target business node corresponding to the problem to be integrated according to the key feedback information;

[0101] The prompt text generation module 202 is used to obtain a target prompt template corresponding to the problem to be integrated according to the target business node corresponding to the problem to be integrated, and combine the key feedback information of the problem to be integrated with the target prompt template to generate a prompt text corresponding to the problem to be integrated;

[0102] A target question type determination module 203 is used to input the prompt text corresponding to the question to be integrated into a pre-trained question classification language model to obtain a target question type for the question to be integrated;

[0103] A high-dimensional vector determination module 204 is used to determine the high-dimensional vector of the problem to be integrated, and to determine the high-dimensional vector of each problem library in the target problem type;

[0104] The target question library determination module 205 is used to calculate the vector similarity between the high-dimensional vector of the problem to be integrated and the high-dimensional vector of each question library, and determine the target question library of the problem to be integrated in each question library according to each vector similarity.

[0105] In one embodiment, the target service node determination module 201 includes:

[0106] An attention weight determination unit, used to determine the attention weight and semantic information of each word in the problem to be integrated using a self-supervised learning model;

[0107] A word screening unit is used to screen each word in the problem to be integrated according to the attention weight of each word in the problem to be integrated;

[0108] The key feedback information extraction unit is used to extract key feedback information from the questions to be integrated according to the semantic information corresponding to each word after screening.

[0109] In one embodiment, the target service node determination module 201 includes:

[0110] A feedback information acquisition unit, used to acquire feedback information corresponding to multiple service nodes;

[0111] The target business node determination unit is used to match the key feedback information with the feedback information corresponding to each business node, and use the business node with a high matching degree as the target business node of the problem to be integrated.

[0112] In one embodiment, the prompt text generation module 202 includes:

[0113] A candidate prompt target query unit is used to query a candidate prompt template corresponding to the target business node from a pre-established prompt template library according to the target business node of the problem to be integrated;

[0114] The target prompt target determination unit is used to calculate the matching degree between the problem to be integrated and each of the prompt templates to be selected if there are multiple prompt templates to be selected, and select the target prompt template from each of the prompt templates to be selected according to each matching degree.

[0115] In one embodiment, the target question type determination module 203 includes:

[0116] A sample question acquisition unit, used for acquiring a plurality of sample questions and determining a standard question type for each sample question;

[0117] The question classification large language model determination unit is used to train the large language model using each sample question, and use the standard question type of each sample question to evaluate the accuracy of the trained large language template until the training round reaches a preset threshold to obtain the question classification large language model.

[0118] In one embodiment, the target question library determination module 205 includes:

[0119] The target question library determination unit is used to use the question library with the highest vector similarity as the target question library for the questions to be integrated.

[0120] In one embodiment, the high-dimensional vectors of each question bank in the target question type are stored in Milvus' efficient vector database.

[0121] In one embodiment, the present application also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the problem integration method as described in any of the above embodiments.

[0122] In one embodiment, the present application also provides a computer device having computer-readable instructions stored therein. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the problem integration method as described in any one of the above embodiments.

[0123] Indicatively, Figure 3 As shown, Figure 3 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of the present application. The computer device 300 may be provided as a server. Figure 3 The computer device 300 includes a processing component 302, which further includes one or more processors, and a memory resource represented by a memory 301 for storing instructions executable by the processing component 302, such as an application. The application stored in the memory 301 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 302 is configured to execute the instructions to perform the problem integration method of any of the above embodiments.

[0124] The computer device 300 may further include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 may operate based on an operating system stored in the memory 301, such as Windows Server TM, Mac OS X TM, Unix TM, Linux TM, Free BSD TM, or the like.

[0125] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0126] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the term "include", "comprise" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not clearly listed, or also includes elements inherent to such process, method, article or equipment. In the absence of more restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the process, method, article or equipment including the elements. Herein, "one", "one", "said", "the" and "it" may also include plural forms, unless the context clearly indicates another way. A plurality refers to at least two cases, such as 2, 3, 5 or 8, etc. "And / or" includes any and all combinations of the relevant listed items.

[0127] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can refer to each other.

[0128] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A problem integration method, characterized in that: The method comprises: Acquire the problem to be integrated, identify key feedback information in the problem to be integrated, and determine the target business node corresponding to the problem to be integrated according to the key feedback information; According to the target service node corresponding to the problem to be integrated, a target prompt template corresponding to the problem to be integrated is obtained, and key feedback information of the problem to be integrated and the target prompt template are combined to generate a prompt text corresponding to the problem to be integrated; Inputting the prompt text corresponding to the problem to be integrated into a pre-trained large language model for problem classification to obtain a target problem type of the problem to be integrated; Determine the high-dimensional vector of the problem to be integrated, and determine the high-dimensional vector of each question library in the target question type; The vector similarity between the high-dimensional vector of the problem to be integrated and the high-dimensional vector of each of the question libraries is calculated, and according to each of the vector similarities, a target question library of the problem to be integrated is determined in each of the question libraries.

2. The problem integration method according to claim 1, characterized in that: The step of identifying key feedback information in the problem to be integrated includes: Using a self-supervised learning model to determine the attention weight and semantic information of each word in the problem to be integrated; Screening each word in the problem to be integrated according to the attention weight of each word in the problem to be integrated; The key feedback information is extracted from the question to be integrated according to the semantic information corresponding to each filtered word.

3. The problem integration method according to claim 1, characterized in that: The step of determining the target service node corresponding to the problem to be integrated according to the key feedback information includes: Obtain feedback information corresponding to multiple business nodes; The key feedback information is matched with the feedback information corresponding to each of the service nodes, and the service nodes with high matching degree are used as the target service nodes of the problem to be integrated.

4. The problem integration method according to claim 1, characterized in that: The step of acquiring a target prompt template corresponding to the problem to be integrated according to the target service node corresponding to the problem to be integrated includes: According to the target service node of the problem to be integrated, querying a prompt template to be selected corresponding to the target service node from a pre-established prompt template library; If there are multiple prompt templates to be selected, the matching degree between the problem to be integrated and each prompt template to be selected is calculated, and the target prompt template is selected from each prompt template to be selected according to each matching degree.

5. The problem integration method according to claim 1, characterized in that: The training process of the large language model for question classification includes: obtaining a plurality of sample questions and determining a standard question type for each of the sample questions; The large language model is trained using each of the sample questions, and the standard question type of each of the sample questions is used to evaluate the accuracy of the trained large language template until the training round reaches a preset threshold, thereby obtaining the large language model for question classification.

6. The problem integration method according to claim 1, characterized in that: The step of determining the target question library of the questions to be integrated in each of the question libraries according to the vector similarities comprises: The question library with the highest vector similarity is used as the target question library for the questions to be integrated.

7. The problem integration method according to any one of claims 1 to 6, characterized in that: The high-dimensional vectors of each question library in the target question type are stored in Milvus's efficient vector database.

8. A problem integration device, characterized in that: The device comprises: A target business node determination module is used to obtain the problem to be integrated, identify key feedback information in the problem to be integrated, and determine the target business node corresponding to the problem to be integrated according to the key feedback information; A prompt text generation module, used to obtain a target prompt template corresponding to the problem to be integrated according to the target business node corresponding to the problem to be integrated, and combine the key feedback information of the problem to be integrated with the target prompt template to generate a prompt text corresponding to the problem to be integrated; A target question type determination module, used for inputting the prompt text corresponding to the question to be integrated into a pre-trained question classification language model to obtain the target question type of the question to be integrated; A high-dimensional vector determination module, used to determine the high-dimensional vector of the problem to be integrated, and to determine the high-dimensional vector of each question library in the target question type; The target question library determination module is used to calculate the vector similarity between the high-dimensional vector of the problem to be integrated and the high-dimensional vector of each of the question libraries, and determine the target question library of the problem to be integrated in each of the question libraries based on the vector similarities.

9. A storage medium, characterized in that: The storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the problem integration method as described in any one of claims 1 to 7.

10. A computer device, characterized in that: include: one or more processors, and memory; The memory stores computer-readable instructions, and when the computer-readable instructions are executed by the one or more processors, the steps of the problem integration method according to any one of claims 1 to 7 are performed.

Images