Customer complaint problem replying method and device and computer equipment

Abstract

The invention relates to a reply method and device for a customer complaint problem and computer equipment, and relates to the technical field of artificial intelligence. The method comprises the following steps: acquiring a customer complaint question; the customer complaint problem is input into a problem processing model, a prediction output result output by the problem processing model is obtained, and the problem processing model comprises a converter-based bidirectional encoder representation BERT model and a deep pyramid convolutional neural network DPCNN model; the prediction output result is used for indicating probability distribution of each candidate output result; the BERT model is used for capturing context information of customer complaint questions, and the DPCNN model is used for capturing multi-level semantic feature information; generating reply content of the customer complaint question based on the prediction output result; through the method, semantic analysis and feature extraction can be performed on the customer complaint problem by integrating the capabilities of the BERT model and the DPCNN model, so that the reply accuracy of the customer complaint problem is improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of artificial intelligence technology, and in particular, to a method, device and computer device for replying to customer complaint problems. Background Art

[0002] Currently, there are some problems in the intelligent customer complaint system when processing user problems, such as insufficient data volume, overfitting, and poor generalization ability. In order to improve the problem matching ability, a common method is to use traditional text classification models with relatively simple structures to classify problems, such as Bayesian models, decision tree models, K-nearest neighbor models, etc.

[0003] However, traditional text classification models have certain limitations in capturing sentence semantic information, resulting in poor accuracy of the generated reply content corresponding to customer complaint problems, and thus poor service effects of the intelligent customer complaint system. Summary of the Invention

[0004] The embodiments of the present application provide a method, device and computer device for replying to customer complaint problems, which can comprehensively use the semantic understanding ability of the BERT model and the feature extraction ability of the DPCNN model to perform semantic parsing and feature extraction on customer complaint problems, improve the accuracy of the predicted output results, improve the accuracy of replying to customer complaint problems, and improve the service effects of the intelligent customer complaint system. The technical solution is as follows.

[0005] On the one hand, a training method for a recommendation model is provided, and the method includes: Obtain customer complaint problems; Input the customer complaint problems into a problem processing model to obtain a predicted output result output by the problem processing model. The problem processing model includes a Bidirectional Encoder Representations from Transformers (BERT) model and a Deep Pyramid Convolutional Neural Network (DPCNN) model. The BERT model is used to capture the context information of the customer complaint problems, and the DPCNN model is used to capture multi-level semantic feature information. The predicted output result is used to indicate the probability distribution of each candidate output result; Generate a reply content for the customer complaint problems based on the predicted output result.

[0006] On the other hand, a training device for a recommendation model is provided, and the device includes: A problem acquisition module for obtaining customer complaint problems; A prediction module, configured to input the customer complaint problem into a problem processing model to obtain a predicted output result output by the problem processing model. The problem processing model includes a Bidirectional Encoder Representations from Transformers (BERT) model and a Deep Pyramid Convolutional Neural Network (DPCNN) model. The BERT model is used to capture the context information of the customer complaint problem, and the DPCNN model is used to capture multi-level semantic feature information. The predicted output result is used to indicate the probability distribution of each candidate output result. A generation module, configured to generate a response content for the customer complaint problem based on the predicted output result.

[0007] In a possible implementation manner, when the training samples of the problem processing model are sample problems and the corresponding category labels of the sample problems, the problem processing model is used to predict the candidate problem categories of the input customer complaint problem, and the predicted output result is used to indicate the probability distribution of the input customer complaint problem belonging to each candidate problem category.

[0008] In a possible implementation manner, the generation module includes: A first determination sub-module, configured to determine the problem category of the customer complaint problem based on the predicted output result. The problem category is the candidate problem category with the highest probability value in the predicted output result. A text acquisition sub-module, configured to obtain a response text for the customer complaint problem based on the problem category of the customer complaint problem. A first generation sub-module, configured to generate a response content for the customer complaint problem based on the response text of the customer complaint problem.

[0009] In a possible implementation manner, the text acquisition sub-module is configured to, When the customer complaint problem is a known problem, perform a response text retrieval based on the problem category of the customer complaint problem to obtain a response text for the customer complaint problem. When the customer complaint problem is an unknown problem, perform a customer complaint problem assignment based on the problem category of the customer complaint problem to obtain a response text for the feedback customer complaint problem.

[0010] In a possible implementation manner, when the training samples of the problem processing model are sample question-and-answer pairs, the problem processing model is used to predict the candidate response texts of the input customer complaint problem, and the predicted output result is used to indicate the probability distribution of each candidate response text corresponding to the input customer complaint problem.

[0011] In a possible implementation manner, the generation module includes: A second determination sub-module, configured to determine a response text for the customer complaint problem based on the prediction output result, where the response text is the candidate response text with the highest probability value in the prediction output result; A second generation sub-module, configured to generate a response content for the customer complaint problem based on the response text of the customer complaint problem.

[0012] In a possible implementation, the prediction module is configured to, Input the customer complaint problem into the BERT model of the problem processing model to extract context information, and obtain a context word embedding vector of the customer complaint problem; Input the context word embedding vector into a pooling layer to perform pooling processing on the features of each context word embedding vector, and obtain a pooling result; Input the pooling result into the DPCNN model to perform multi-level semantic feature information extraction, and obtain text classification features; Input the text classification features into a classification layer to perform classification prediction, and obtain a probability distribution of each candidate output result.

[0013] In a possible implementation, the number of stacked layers in the BERT model for context information extraction is less than a target threshold.

[0014] On the other hand, a computer device is provided, where the computer device includes a processor and a memory, and the memory stores at least one computer program, and the at least one computer program is loaded and executed by the processor to implement the above-mentioned method for replying to customer complaint problems.

[0015] On the other hand, a computer-readable storage medium is provided, where at least one computer program is stored in the computer-readable storage medium, and the computer program is loaded and executed by a processor to implement the above-mentioned method for replying to customer complaint problems.

[0016] On the other hand, a computer program product is provided, where the computer program product includes at least one computer program, and the computer program is loaded and executed by a processor to implement the method for replying to customer complaint problems provided in the above various optional implementations.

[0017] The technical solution provided by this application may include the following beneficial effects: The method for replying to customer complaint problems provided by the embodiments of the present application inputs the obtained customer complaint problems into a problem processing model including a BERT model and a DPCNN model to obtain a predicted output result of the customer complaint problems, and the predicted output result is used to indicate the probability distribution of each candidate output result; based on the predicted output result, a reply content for the customer complaint problems is generated; through the above method, the semantic understanding ability of the BERT model to capture the context information of the complaint problems and the feature extraction ability of the DPCNN model to capture multi-level language feature information can be comprehensively used to perform semantic parsing and feature extraction on the customer complaint problems, improve the accuracy of feature extraction, thereby improve the accuracy of the predicted output result, improve the accuracy of the reply to the customer complaint problems, and further improve the service effect of the intelligent customer complaint system.

[0018] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present application. Brief Description of the Drawings

[0019] The drawings here are incorporated into the specification and constitute a part of this specification, showing embodiments consistent with the present application, and are used together with the specification to explain the principles of the present application.

[0020] Figure 1 Shows a flowchart of the method for replying to customer complaint problems provided by an exemplary embodiment of the present application; Figure 2 Shows a flowchart of the method for replying to customer complaint problems provided by another exemplary embodiment of the present application; Figure 3 Shows a schematic diagram of a problem processing model provided by an exemplary embodiment of the present application; Figure 4 Shows a schematic diagram of a BERT model provided by an exemplary embodiment of the present application; Figure 5 Shows a schematic diagram of a deep pyramid model provided by an exemplary embodiment of the present application; Figure 6 Shows a schematic diagram of an intelligent customer complaint system provided by an exemplary embodiment of the present application; Figure 7 Shows a block diagram of a device for replying to customer complaint problems provided by an exemplary embodiment of the present application; Figure 8 Shows a structural block diagram of a computer device shown in an exemplary embodiment of the present application; Figure 9 Shows a structural block diagram of a computer device shown in an exemplary embodiment of the present application. Detailed Description of the Embodiments

[0021] Exemplary embodiments will be described in detail herein, and examples thereof are shown in the accompanying drawings. When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0022] An embodiment of the present application provides a method for replying to customer complaint problems, which can make full use of the advantages of a deep learning model to improve the accuracy of replying to customer complaint problems, thereby improving the service effect of the intelligent customer complaint system. Figure 1 The flowchart of the method for replying to customer complaint problems provided by an exemplary embodiment of the present application is shown. This method can be executed by a computer device, which can be implemented as a server or a terminal, such as Figure 1 As shown, the method for replying to customer complaint problems may include the following steps.

[0023] Step 110, obtain customer complaint problems.

[0024] In the embodiment of the present application, the computer device can obtain customer complaint problems through multiple problem entry portals, including but not limited to the online customer service form on the website, the feedback entry built into the mobile application, the customer service hotline call recording, and the customer complaint feedback recorded and submitted by relevant offline service personnel. Among them, the customer complaint problem is the complaint problem uploaded by the user.

[0025] Since the customer complaint information from different channels may have different formats, for the convenience of subsequent processing by the computer device, in a possible implementation, the computer device can preprocess the obtained customer complaint information. Among them, the preprocessing can include format unification, text cleaning, word segmentation processing, and so on. Format unification can include transcribing voice information into text format, etc. Text cleaning can include removing special symbols in the text (such as redundant punctuation, emoticons, etc. that interfere with model processing), extra spaces, garbled characters, and incorrect characters, etc., and standardizing the text, such as unifying case, standardizing abbreviations, etc., so that the text expression of the obtained customer complaint problem is more standardized and convenient for subsequent models to perform semantic understanding. Word segmentation processing can indicate that through a natural language processing word segmentation tool, the customer complaint problem text is segmented according to semantic units, and continuous sentences are split into multiple words, which is convenient for subsequent models to understand and encode.

[0026] Step 120: Input the customer complaint problem into the problem handling model to obtain the predicted output result output by the problem handling model. The problem handling model includes a Bidirectional Encoder Representations from Transformers (BERT) model and a Deep Pyramid Convolutional Neural Network (DPCNN) model. The BERT model is used to capture the context information of the customer complaint problem, and the DPCNN model is used to capture multi-level semantic feature information. The predicted output result is used to indicate the probability distribution of each candidate output result.

[0027] The problem handling model in the embodiment of this application is constructed by combining a BERT (Bidirectional Encoder Representations from Transformers) model and a DPCNN (Deep Pyramid Convolutional Neural Network) model. After inputting the customer complaint problem into the problem handling model, it is processed by the BERT model and the DPCNN model in sequence, and the predicted output result corresponding to the customer complaint problem can be obtained based on the model processing result. Among them, the BERT model can, based on its own multi-layer stacked layer (or transformer) architecture, deeply encode the position information and context semantic information of each word in the received customer complaint text, capture the complex semantic relationships between words, and generate a vector representation integrating rich semantic features. The vector representation output by the BERT model is used as the input of the DPCNN model. The DPCNN model, based on the deep pyramid convolution structure, performs multi-layer convolution operations on the input feature vector, gradually extracts higher-level and more abstract semantic features, and then obtains the probability distribution of each candidate output result corresponding to the customer complaint problem based on this semantic feature.

[0028] Step 130: Generate a reply content for the customer complaint problem based on the predicted output result.

[0029] The computer device analyzes the probability distribution of each candidate output result output by the model, determines the most likely solution or reply direction corresponding to the customer complaint problem, and generates a reply content for the customer complaint problem accordingly. Among them, the candidate output result can be the problem classification of the customer complaint problem, or it can also be the reply text corresponding to the customer complaint problem.

[0030] In a possible implementation manner, when generating the content of the customer complaint problem, the computer device can call a reply template library, which contains text templates for customer complaint problems of different problem types and solutions. When generating the reply content for the customer complaint problem, the reply text corresponding to the customer complaint problem can be filled into the corresponding text template to obtain the reply content for the customer complaint problem.

[0031] Further, after generating the response content for the customer complaint problem, the response content is fed back to the user through the interaction platform with the user.

[0032] In a possible implementation manner, the process by which the computer device generates the response content for the customer complaint problem can be implemented as follows: Based on the candidate output result with the highest probability value in the prediction output result, generate the response content for the customer complaint problem.

[0033] That is to say, the computer device determines the candidate output result with the highest probability value in the prediction output result as the final prediction result of the customer complaint problem, and obtains or generates the response content based on this.

[0034] In a possible implementation manner, a probability threshold is set in the computer device. If the highest probability value in the probability distribution of each candidate output result indicated by the prediction output result does not exceed the probability threshold, the computer device can comprehensively analyze the top N candidate output results with relatively high probability values to generate the response content for the customer complaint problem, where N is a positive integer; or, if the highest probability value in the probability distribution of each candidate output result indicated by the prediction output result does not exceed the probability threshold, the computer device can mark the customer complaint problem as a difficult problem and feedback it to the relevant personnel so that the relevant personnel can generate and feedback the response text for the customer complaint problem.

[0035] In summary, for the response method for customer complaint problems provided in the embodiments of the present application, by inputting the obtained customer complaint problem into the problem processing model including the BERT model and the DPCNN model, the prediction output result of the customer complaint problem is obtained, and the prediction output result is used to indicate the probability distribution of each candidate output result; based on the prediction output result, the response content for the customer complaint problem is generated; through the above method, the semantic understanding ability of the BERT model to capture the context information of the complaint problem and the feature extraction ability of the DPCNN model to capture multi-level language feature information can be comprehensively used to perform semantic parsing and feature extraction on the customer complaint problem, improve the accuracy of feature extraction, thereby improve the accuracy of the prediction output result, improve the accuracy of the response to the customer complaint problem, and further improve the service effect of the intelligent customer complaint system.

[0036] In the embodiments of the present application, the type of the predicted output result output by the problem processing model can be determined based on the training samples used in the training process of the problem processing model. When the training samples of the problem processing model are sample problems and the class labels corresponding to the sample problems, the predicted output result is used to indicate the probability distribution of the input customer complaint problem belonging to each problem category, and the problem processing model is used to predict the candidate problem categories of the input customer complaint problem; when the training samples of the problem processing model are sample question-and-answer pairs, the predicted output result is used to indicate the probability distribution of each candidate response text corresponding to the input customer complaint problem, and the problem processing model is used to predict the candidate response text of the input customer complaint problem; Figure 2 FIG. shows a flowchart of a method for replying to a customer complaint problem provided by another exemplary embodiment of the present application. This method can be executed by a computer device, and the computer device can be implemented as a server or a terminal, such as Figure 2 shown, the method for replying to the customer complaint problem may include the following steps.

[0037] Step 210, obtain a customer complaint problem.

[0038] Step 220, input the customer complaint problem into the problem processing model to obtain the predicted output result output by the problem processing model. The problem processing model includes a Bidirectional Encoder Representations from Transformers (BERT) model based on a transformer and a Deep Pyramid Convolutional Neural Network (DPCNN) model. The BERT model is used to capture the context information of the customer complaint problem, and the DPCNN model is used to capture multi-level semantic feature information; the predicted output result is used to indicate the probability distribution of each candidate output result.

[0039] Optionally, the process of the problem processing model for processing the customer complaint problem is as follows: Input the customer complaint problem into the BERT model in the problem processing model to extract context information, and obtain the context word embedding vector of the customer complaint problem; Input the context word embedding vector into the pooling layer to perform pooling processing on the features of each context word embedding vector, and obtain a pooling result; Input the pooling result into the model to extract multi-level semantic feature information, and obtain text classification features; Input the text classification features into the classification layer to perform classification prediction, and obtain the probability distribution of each candidate output result.

[0040] Figure 3 FIG. shows a schematic diagram of a problem processing model provided by an exemplary embodiment of the present application, such as Figure 3As shown, the problem processing model 300 may include an input layer 310, a BERT layer 320, a pooling layer 330, a convolutional layer 340, a classification layer 350, and an output layer 360. Among them, the BERT layer corresponds to the BERT model, and the convolutional layer corresponds to the DPCNN model. The functions of each network layer are as follows: The input layer is used to receive customer complaints and input the customer complaints into the BERT model.

[0041] In the BERT layer, the BERT model performs semantic extraction on the customer complaints. Figure 4 The figure shows a schematic diagram of the BERT model provided by an exemplary embodiment of the present application. As Figure 4 shown, after receiving the customer complaints input by the input layer, the BERT model converts each word in the text into a one-dimensional vector, that is, a word embedding vector (Token Embedding), as the model input. In addition, the input of the BERT model also includes a segment embedding vector (Segment Embedding) and a position embedding vector (Position Embedding); among them, the segment embedding vector is used to indicate the continuation situation of the next sentence and the previous sentence, with values of 0 and 1; the position embedding vector is used to distinguish words / tokens at different positions; the sum of the word embedding vector, the segment embedding vector, and the position embedding vector is input into the multi-layer transformer structure of the BERT model, and by combining the context information of the text, each word vector is further encoded to generate a word embedding vector containing rich context semantics. Since the BERT model in the embodiments of the present application is combined with the DPCNN model, the vector obtained after the last addition and normalization layer is determined as the context word embedding vector; in the above process of semantic extraction by the BERT model, the BERT model can, through the attention mechanism, enable each word to pay attention to the information of other relevant words in the sentence during encoding, fully capture complex language phenomena such as semantic relationships and referential relationships in the text, so as to provide a high-quality semantic representation basis for subsequent feature processing.

[0042] The pooling layer is used to perform a pooling operation on the features of each context word embedding vector output by the BERT layer; the pooling method of this pooling layer can be one of the following: max pooling, average pooling, and sum pooling; Among them, max pooling refers to extracting the maximum value on each feature dimension of the context word embedding vectors output by the BERT model. Through max pooling, the pooled result can highlight the most significant semantic features in the text; mean pooling refers to extracting the average value on each feature dimension of the context word embedding vectors output by the BERT model. Through mean pooling, the pooled result can reflect the overall average level of the text semantic information; sum pooling refers to summing up on each dimension of the context word embedding vectors output by the BERT model. Through sum pooling, the pooled result can emphasize the total sum information of the features, such as counting the total intensity or total occurrence times of certain semantic features in the entire text, thereby providing a quantitative index for the total amount of relationship features for the model.

[0043] Based on different actual application requirements, the pooling method of the pooling layer can be set accordingly, and this application does not limit this.

[0044] In the convolutional layer, the DPCNN model further extracts features from the pooling result output by the pooling layer. Among them, the DPCNN model consists of two parts: Region embedding and Deep Pyramid, and text classification features are extracted through these two parts. Among them, region embedding is used to initially integrate and extract the pooled result, encode the features of the local region, and prepare for subsequent deep feature extraction; deep pyramid is used to build a deep pyramid structure by continuously stacking convolutional layers and pooling layers on the basis of region embedding, gradually expanding the receptive field, and extracting higher-level and more comprehensive semantic features to achieve accurate classification of the text.

[0045] Among them, the convolutional operations of each part in the DPCNN model can be divided into the following three categories: Assume that the length of the input sequence is n, the size of the convolutional kernel is m, the stride is s, and p zeros (zero padding) are filled at both ends of the input sequence. Then the output sequence of this convolutional layer is (n - m + 2p) / s + 1.

[0046] Narrow convolution: stride s = 1, no zero padding at both ends, that is, p = 0, and the output length after convolution is n - m + 1.

[0047] Wide convolution: stride s = 1, zero padding p = m - 1 at both ends, and the output length after convolution is n + m - 1.

[0048] Equal-width convolution: stride s = 1, zero padding p = (m - 1) / 2 at both ends, and the output length after convolution is n.

[0049] Schematically, in the regional embedding part, the above different types of convolution operations can be selected to process the pooling result. For example, if narrow convolution is selected with a stride of 1 and no zero-padding at both ends, it will slide the convolution kernel on the sequence of the pooling result, and through operations such as multiplying the corresponding elements of the convolution kernel and the sequence and summing them, extract the features of the local region, and the output length is n - m + 1. This extraction of local features can capture the correlation and combined information between adjacent elements in the pooling result, and initially excavate some local semantic features in the text. In the deep pyramid part, Figure 5 shows a schematic diagram of the deep pyramid model provided by an exemplary embodiment of the present application, as Figure 5 shown, in the deep pyramid model, multiple convolutional layers are stacked in sequence, and each convolutional layer uses the output of the previous layer as input to continue feature extraction. As the convolutional layers are stacked, the model can learn more and more complex semantic feature combinations. In the above process, different types of convolution operations can be flexibly selected according to actual needs to adapt to different text features and task requirements; in addition, a pooling layer is also deployed at an appropriate position in the deep pyramid model to perform dimensionality reduction processing on the convolved features, reduce the amount of calculation and model parameters, and at the same time further expand the receptive field, so that the model can pay attention to more macroscopic text features and obtain text classification features.

[0050] The classification layer is used to receive the text classification features output by the convolutional layer (i.e., the DPCNN model), and uses the softmax function to convert continuous numerical values into probabilities for predicting different candidate output results.

[0051] The output layer is used to output the predicted output result, and the predicted output result is the probability distribution of the input data on each candidate output result.

[0052] In a possible implementation manner, when constructing the problem processing model, the number of stacked layers in the BERT model can be appropriately reduced, that is, the number of stacked layers in the BERT model used for extracting context information is less than the target threshold, and the target threshold can be set based on the initial number of stacked layers of the BERT model. Schematically, the number of stacked layers in the BERT model is 50% of the initial number of stacked layers, etc. It should be noted that based on different actual needs, the target threshold can be set differently, and the present application does not limit this.

[0053] Step 230, when the problem processing model is used to predict the candidate problem categories of the input customer complaint problems, determine the problem category of the customer complaint problem based on the predicted output result, and the problem category is the candidate problem category with the highest probability value in the predicted output result.

[0054] When the training samples of the problem handling model are sample problems and the corresponding category labels of the sample problems, the problem handling model is used to predict the candidate problem categories of the input customer complaint problems, and the prediction output result is used to indicate the probability distribution of the input customer complaint problems belonging to each candidate problem category; the computer device can determine the problem category of the customer complaint problem as the candidate problem category with the highest probability value in the prediction output result.

[0055] Step 240, based on the problem category of the customer complaint problem, obtain the response text of the customer complaint problem.

[0056] In a possible implementation manner, a problem library can be maintained in the computer device, and each received customer complaint problem can be recorded in the problem library. When obtaining the response text of the customer complaint problem, the computer device can retrieve the customer complaint problem in the problem library. If retrieved, it is determined that the customer complaint problem is a known problem, and the computer device can retrieve the response text based on the problem category and the customer complaint problem to obtain the response text of the customer complaint problem; if not retrieved, it is determined that the customer complaint problem is an unknown problem, and the computer device can assign the customer complaint problem to relevant personnel for response based on the problem category of the customer complaint problem, so as to obtain the response text of the customer complaint problem; that is to say, this process can be implemented as: When the customer complaint problem is a known problem, retrieve the response content based on the problem category of the customer complaint problem to obtain the response text of the customer complaint problem; When the customer complaint problem is an unknown problem, assign the customer complaint problem based on the problem category of the customer complaint problem to obtain the response text of the feedback customer complaint problem.

[0057] Furthermore, for the case where the customer complaint problem is a known problem, when the response content generated based on the retrieved response text is fed back to the user, if positive feedback from the user is received, the computer device can increase the answer weight of the response text, so that when the customer complaint problem is received again, the response text with a higher weight can be fed back to the user, thereby improving the response accuracy.

[0058] For the case where the customer complaint problem is an unknown problem, when the response text of the customer complaint problem is received, add the response text to the response text library and mark the customer complaint problem as a known problem, so that when the customer complaint problem is received next time, the response text of the customer complaint problem can be retrieved through response text retrieval.

[0059] Step 250, based on the response text of the customer complaint problem, generate the response content of the customer complaint problem.

[0060] After obtaining the response text for the customer complaint problem based on retrieval or feedback, the computer device can fill the response text into the text template corresponding to the customer complaint problem to obtain the response content for the customer complaint problem. Among them, the text templates corresponding to each customer complaint problem can be pre-configured based on actual requirements, and an association relationship is established between the customer complaint problems and the text templates.

[0061] Step 260, when the problem processing model is used to predict the candidate response text for the input customer complaint problem, determine the response text for the customer complaint problem based on the prediction output result, where the response text is the candidate response text with the highest probability value in the prediction output result.

[0062] When the training samples of the problem processing model are sample question-and-answer pairs, the problem processing model is used to predict the candidate response text for the input customer complaint problem, and the prediction output result is used to indicate the probability distribution of each candidate response text corresponding to the input customer complaint problem. The computer device can determine the candidate response text with the highest probability value in the prediction output result as the response text for the customer complaint problem.

[0063] Step 270, generate the response content for the customer complaint problem based on the response text for the customer complaint problem.

[0064] After obtaining the response text for the customer complaint problem, the computer device can fill the response text into the text template corresponding to the customer complaint problem to obtain the response content for the customer complaint problem.

[0065] Before model training, a problem corpus can be constructed. The problem corpus contains training samples. Depending on the training requirements of the problem processing model, the training samples can be sample questions and corresponding category labels, or the training samples can also be sample question-and-answer pairs. In a possible implementation, to improve the generalization ability and accuracy of the model, the scale and diversity of the problem corpus can be expanded for data augmentation. For example, high-quality data with manual annotation and automatically generated training samples can be introduced to improve the richness and reliability of the sample data. Among them, when automatically generating training samples, the computer device can input the sample customer complaint text into a large language model and input the instruction text to obtain the training samples output by the large language model. By enriching the problem corpus, the problem processing model's recognition ability for different types of problems can be increased, and the phenomenon of overfitting of the model to specific problems can be avoided, thereby improving the generalization ability of the model.

[0066] In summary, for the method for replying to customer complaint problems provided in the embodiments of the present application, by inputting the obtained customer complaint problems into a problem processing model including a BERT model and a DPCNN model, a predicted output result of the customer complaint problems is obtained, and the predicted output result is used to indicate the probability distribution of each candidate output result; based on the predicted output result, a reply content for the customer complaint problems is generated; through the above method, the semantic understanding ability of the BERT model to capture the context information of the complaint problems and the feature extraction ability of the DPCNN model to capture multi-level language feature information can be comprehensively used to perform semantic parsing and feature extraction on the customer complaint problems, improve the accuracy of feature extraction, thereby improve the accuracy of the predicted output result, improve the accuracy of the reply to the customer complaint problems, and further improve the service effect of the intelligent customer complaint system.

[0067] Taking the problem processing model provided in the embodiments of the present application for determining the problem type of customer complaint problems as an example, Figure 6 The schematic diagram of the intelligent customer complaint system provided in an exemplary embodiment of the present application is shown, as Figure 6 shown, the intelligent customer complaint system can provide the following services: problem entry, problem distribution, problem viewing and searching, answering questions, and viewing problem reports; among them, the problem entry service supports multiple problem entry ports. After receiving the customer complaint problems through the entry port and performing preprocessing, it enters the intelligent retrieval process. The intelligent retrieval process is executed by the problem processing model in the embodiment shown in Figure 1 or Figure 2 shown. After predicting the problem type of the customer complaint problems through the problem processing model, if it is determined in the problem library that the customer complaint problems are known problems, then based on the problem type of the customer complaint problems and the customer complaint problems, a reply text is retrieved in the reply text library, a reply content is generated and fed back to the user. If positive feedback from the user (that is, the reply answer is correct) is received, the weight of the reply text is increased. If negative feedback from the user (that is, the reply answer is wrong) is received, the result status is recorded; further, the customer complaint problems are assigned based on the problem type for customer complaint problem assignment and enter the problem answering process; if the customer complaint problems are unknown problems, then the customer complaint problems can be assigned based on the problem type for customer complaint problem assignment and enter the problem answering process; after obtaining the reply text of the customer complaint problems through the problem answering process, a reply content is generated based on the reply text and fed back to the user. In addition, the reply text is added to the reply text library, and the customer complaint problems are marked as known problems.

[0068] Optionally, in the problem answering process, relevant personnel can answer the customer complaint problems, or the customer complaint problems can be input into a large language model pre-trained based on customer complaint logs to generate a reply text.

[0069] Figure 7The block diagram of a device for replying to customer complaint problems provided by an exemplary embodiment of the present application is shown. The device can execute all or part of the steps of the embodiments shown as Figure 1 or Figure 2 shown. As shown in Figure 7 , the device includes: A problem acquisition module 710, configured to acquire customer complaint problems; A prediction module 720, configured to input the customer complaint problem into a problem processing model to obtain a predicted output result output by the problem processing model. The problem processing model includes a Bidirectional Encoder Representations from Transformers (BERT) model and a Deep Pyramid Convolutional Neural Network (DPCNN) model. The BERT model is used to capture the context information of the customer complaint problem, and the DPCNN model is used to capture multi-level semantic feature information. The predicted output result is used to indicate the probability distribution of each candidate output result; A generation module 730, configured to generate a reply content for the customer complaint problem based on the predicted output result.

[0070] In a possible implementation manner, when the training samples of the problem processing model are sample problems and the corresponding category labels of the sample problems, the problem processing model is used to predict the candidate problem categories of the input customer complaint problems, and the predicted output result is used to indicate the probability distribution of the input customer complaint problems belonging to each candidate problem category.

[0071] In a possible implementation manner, the generation module 730 includes: A first determination sub-module, configured to determine the problem category of the customer complaint problem based on the predicted output result. The problem category is the candidate problem category with the highest probability value in the predicted output result; A text acquisition sub-module, configured to acquire a reply text for the customer complaint problem based on the problem category of the customer complaint problem; A first generation sub-module, configured to generate a reply content for the customer complaint problem based on the reply text of the customer complaint problem.

[0072] In a possible implementation manner, the text acquisition sub-module is configured to, when the customer complaint problem is a known problem, perform a reply text retrieval based on the problem category of the customer complaint problem to obtain a reply text for the customer complaint problem; when the customer complaint problem is an unknown problem, perform a customer complaint problem assignment based on the problem category of the customer complaint problem to obtain a reply text for the feedback customer complaint problem.

[0073] In a possible implementation, when the training samples of the problem processing model are sample question-and-answer pairs, the problem processing model is used to predict the candidate response text of the input customer complaint problem, and the predicted output result is used to indicate the probability distribution of each candidate response text corresponding to the input customer complaint problem.

[0074] In a possible implementation, the generating module 730 includes: A second determining sub-module, configured to determine the response text of the customer complaint problem based on the predicted output result, where the response text is the candidate response text with the highest probability value in the predicted output result; A second generating sub-module, configured to generate the response content of the customer complaint problem based on the response text of the customer complaint problem.

[0075] In a possible implementation, the predicting module 720 is configured to, Input the customer complaint problem into the BERT model in the problem processing model to extract context information, and obtain the context word embedding vector of the customer complaint problem; Input the context word embedding vector into the pooling layer to perform pooling processing on the features of each context word embedding vector, and obtain a pooling result; Input the pooling result into the DPCNN model to perform multi-level semantic feature information extraction, and obtain text classification features; Input the text classification features into the classification layer to perform classification prediction, and obtain the probability distribution of each candidate output result.

[0076] In a possible implementation, the number of stacked layers in the BERT model for context information extraction is less than the target threshold.

[0077] In summary, the response device for customer complaint problems provided by the embodiments of the present application inputs the obtained customer complaint problem into a problem processing model including a BERT model and a DPCNN model to obtain a predicted output result of the customer complaint problem, and the predicted output result is used to indicate the probability distribution of each candidate output result; based on the predicted output result, generate the response content of the customer complaint problem; through the above device, the semantic understanding ability of the BERT model to capture the context information of the customer complaint problem and the feature extraction ability of the DPCNN model to capture multi-level language feature information can be combined to perform semantic parsing and feature extraction on the customer complaint problem, improve the accuracy of feature extraction, thereby improving the accuracy of the predicted output result, improving the accuracy of the response to the customer complaint problem, and further improving the service effect of the intelligent customer complaint system.

[0078] Figure 8The block diagram of a computer device 800 shown in an exemplary embodiment of the present application is presented. This computer device can be implemented as the server in the above-mentioned solution of the present application. The computer device 800 includes a Central Processing Unit (CPU) 801, a system memory 804 including a Random Access Memory (RAM) 802 and a Read-Only Memory (ROM) 803, and a system bus 805 connecting the system memory 804 and the central processing unit 801. The computer device 800 further includes a mass storage device 806 for storing an operating system 808, application programs 810, and other program modules 811. The above-mentioned system memory 804 and mass storage device 806 can be collectively referred to as the memory.

[0079] According to various embodiments of the present application, the computer device 800 can also run on a remote computer on the network, such as the Internet. That is, the computer device 800 can be connected to a network 809 through a network interface unit 807 connected to the system bus 805. Or rather, the network interface unit 807 can also be used to connect to other types of networks or remote computer systems (not shown).

[0080] The memory further includes at least one instruction, at least one program, a code set, or an instruction set. The at least one instruction, at least one program, code set, or instruction set is stored in the memory, and the central processing unit 801 implements all or part of the steps in the method for replying to customer complaint problems shown in the above various embodiments by executing the at least one instruction, at least one program, code set, or instruction set.

[0081] Figure 9 The block diagram of a computer device 900 shown in an exemplary embodiment of the present application is presented. This computer device 900 can be implemented as the above-mentioned terminal.

[0082] Generally, the computer device 900 includes a processor 901 and a memory 902.

[0083] In some embodiments, the computer device 900 may also optionally include a peripheral device interface 903 and at least one peripheral device. The processor 901, the memory 902, and the peripheral device interface 903 can be connected through a bus or signal lines. Each peripheral device can be connected to the peripheral device interface 903 through a bus, signal lines, or a circuit board. Specifically, the peripheral device includes at least one of a radio frequency circuit 904, a display screen 905, a camera assembly 906, an audio circuit 907, and a power supply 908.

[0084] In some embodiments, the computer device 900 further includes one or more sensors 909. The one or more sensors 909 include but are not limited to: an acceleration sensor 910, a gyroscope sensor 911, a pressure sensor 912, an optical sensor 913, and a proximity sensor 914.

[0085] Those skilled in the art can understand that Figure 9 the structure shown in does not constitute a limitation on the computer device 900, and it may include more or fewer components than shown in the figure, combine certain components, or adopt different component arrangements.

[0086] In an exemplary embodiment, a computer-readable storage medium is further provided. At least one computer program is stored in the computer-readable storage medium, and the computer program is loaded and executed by a processor to implement all or part of the steps in the above method for replying to customer complaint problems. For example, the computer-readable storage medium may be a Read-Only Memory (ROM), a Random Access Memory (RAM), a Compact Disc Read-Only Memory (CD-ROM), a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0087] In an exemplary embodiment, a computer program product is further provided. The computer program product includes at least one computer program, and the computer program is loaded and executed by a processor to implement all or part of the steps in the Figure 1 or Figure 2 reply method for customer complaint problems shown in any of the above embodiments.

[0088] After considering the specification and practicing the invention disclosed herein, those skilled in the art will readily conceive of other embodiments of the present application. The present application is intended to cover any variations, uses, or adaptations of the present application, which follow the general principles of the present application and include common general knowledge or conventional technical means in the technical field not disclosed in the present application. The specification and embodiments are only regarded as exemplary, and the true scope and spirit of the present application are pointed out by the following claims.

[0089] It should be understood that the present application is not limited to the exact structure already described and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present application is only limited by the appended claims.

Claims

1. A method for responding to customer complaints, characterized in that: The method comprises: Obtain customer complaints; Input the customer complaint into the problem processing model to obtain the predicted output result of the problem processing model, wherein the problem processing model includes a transformer-based bidirectional encoder representation BERT model and a deep pyramid convolutional neural network DPCNN model, wherein the BERT model is used to capture the context information of the customer complaint, and the DPCNN model is used to capture multi-level semantic feature information; the predicted output result is used to indicate the probability distribution of each candidate output result; Based on the predicted output result, a response content to the customer complaint is generated.

2. The method according to claim 1, characterized in that When the training samples of the problem processing model are sample problems and category labels corresponding to the sample problems, the problem processing model is used to predict the candidate problem categories of the input customer complaint problems, and the prediction output result is used to indicate the probability distribution of the input customer complaint problems belonging to each candidate problem category.

3. The method according to claim 2, characterized in that Generating a response to the customer complaint based on the prediction output result includes: Determine the problem category of the customer complaint based on the prediction output result, the problem category being the candidate problem category with the highest probability value in the prediction output result; Based on the problem category of the customer complaint, obtaining a reply text of the customer complaint; Based on the reply text of the customer complaint, the reply content of the customer complaint is generated.

4. The method according to claim 2, characterized in that: The step of obtaining a reply text to the customer complaint based on the category of the customer complaint includes: In the case where the customer complaint is a known problem, a reply text search is performed based on the problem category of the customer complaint to obtain a reply text of the customer complaint; In the case that the customer complaint issue is an unknown issue, the customer complaint issue is assigned based on the issue category of the customer complaint issue to obtain a reply text of the customer complaint issue that is fed back.

5. The method according to claim 1, characterized in that When the training samples of the question processing model are sample question-answer pairs, the question processing model is used to predict candidate response texts for the input customer complaint question, and the prediction output result is used to indicate the probability distribution of each candidate response text corresponding to the input customer complaint question.

6. The method according to claim 5, characterized in that Generating a response to the customer complaint based on the prediction output result includes: Determine a reply text for the customer complaint based on the prediction output result, the reply text being the candidate reply text with the highest probability value in the prediction output result; Based on the reply text of the customer complaint, the reply content of the customer complaint is generated.

7. The method according to claim 1, characterized in that The step of inputting the customer complaint into the problem processing model to obtain a predicted output result output by the problem processing model includes: Input the customer complaint into the BERT model of the problem processing model to extract context information, and obtain a context word embedding vector of the customer complaint; Input the context word embedding vector into the pooling layer to perform pooling processing on the features of each context word embedding vector to obtain a pooling result; Inputting the pooling result into the DPCNN model to extract multi-level semantic feature information to obtain text classification features; The text classification features are input into the classification layer for classification prediction to obtain the probability distribution of each candidate output result.

8. The method according to claim 7, characterized in that The number of stacked layers used for context information extraction in the BERT model is less than the target threshold.

9. A device for responding to customer complaints, characterized in that: The device comprises: Problem acquisition module, used to obtain customer complaints; A prediction module, used for inputting the customer complaint into a problem processing model to obtain a prediction output result output by the problem processing model, wherein the problem processing model includes a transformer-based bidirectional encoder representation BERT model and a deep pyramid convolutional neural network DPCNN model, wherein the BERT model is used to capture the context information of the customer complaint, and the DPCNN model is used to capture multi-level semantic feature information; the prediction output result is used to indicate the probability distribution of each candidate output result; A generation module is used to generate a response content for the customer complaint based on the predicted output result.

10. A computer device, characterized in that: The computer device includes a processor and a memory, wherein the memory stores at least one computer program, and the at least one computer program is loaded and executed by the processor to implement the method for responding to customer complaints as described in any one of claims 1 to 8.

Images