Customer complaint probability prediction method and device, equipment and storage medium

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses a customer complaint probability prediction method and device, equipment and a storage medium, and the method comprises the steps: extracting the features of historical call data through a multi-task classification model, and obtaining target global features, outputting scene probability distribution, emotion probability distribution and appeal probability distribution based on the target global feature, integrating with customer behavior data to obtain a compound feature vector, processing the compound feature vector to obtain a complaint occurrence probability value, and judging whether the customer is a customer with high complaint tendency or not; when the customer is a customer with high complaint tendency, labeling a complaint reason through a label system to obtain a target complaint reason; formulating a pacifying strategy according to the target complaint reason, and feeding back the target complaint reason and the pacifying strategy to the seat personnel; the method can be applied to implementation scenes of the financial field, the accuracy of customer complaint prediction is improved, and then the customer conversion rate is improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to a method, apparatus, device, and storage medium for predicting the probability of customer complaints. Background Art

[0002] In today's booming insurance industry, direct channels, thanks to their convenience and efficiency, have captured a significant portion of the insurance business. These channels typically integrate multiple communication and interaction technologies, such as telephone, internet, and mobile applications, to build a service network covering a broad customer base, enabling customers to access insurance product information, conduct transactions, and consult related issues anytime, anywhere. However, as direct channels' business volume has skyrocketed, so too has the number of customer complaints.

[0003] At present, the existing insurance customer service system has significant technical defects in complaint warning and cause analysis. At the complaint warning level, it lacks an intelligent and precise warning system. It is unable to predict and identify potential complaint risks in advance based on multi-dimensional information such as customers' historical behavior data, real-time interaction data, and business rules. As a result, agents are often in a passive response state after customer complaints occur, and it is difficult to take effective measures to intervene in the embryonic stage of complaints. This not only increases service costs but may also lead to customer churn. In addition, the existing system makes it difficult to conduct in-depth mining and analysis of massive complaint data. Agents are unable to quickly and accurately identify the root causes of customer complaints, resulting in inefficient complaint handling and difficulty in improving customer satisfaction. Summary of the Invention

[0004] The purpose of the embodiments of the present application is to propose a method, device, equipment and storage medium for predicting the probability of customer complaints. Its main purpose is to improve the accuracy of complaint probability prediction and then formulate corresponding appeasement strategies to improve customer conversion rate.

[0005] First, to solve the above technical problems, the present application provides a method for predicting the probability of customer complaints, which adopts the following technical solution:

[0006] Obtain customers' historical call data and customer behavior data;

[0007] Extract features from the historical call data using a multi-task classification model to obtain target global features, and calculate the target global features using the classification layer of the multi-task classification model to output a scene probability distribution, an emotion probability distribution, and a demand probability distribution;

[0008] Integrate the scenario probability distribution, the emotion probability distribution, the demand probability distribution, and the customer behavior data to obtain a composite feature vector, and process the composite feature vector using a binary classification prediction model to obtain a complaint occurrence probability value;

[0009] Determine whether the customer is a high-complaint-tendency customer based on the complaint occurrence probability value. If the customer is a high-complaint-tendency customer, label the complaint reason using a pre-built labeling system to obtain a target complaint reason.

[0010] A soothing strategy is formulated based on the target complaint cause, and the target complaint cause and the soothing strategy are fed back to the agent.

[0011] Secondly, in order to solve the above technical problems, the embodiment of the present application further provides a device for predicting the probability of customer complaints, which adopts the following technical solution:

[0012] Data acquisition module, used to obtain customers' historical call data and customer behavior data;

[0013] a multidimensional distribution calculation module, configured to extract features from the historical call data using a multi-task classification model to obtain target global features, calculate the target global features using the classification layer of the multi-task classification model, and output a scene probability distribution, an emotion probability distribution, and an appeal probability distribution;

[0014] A complaint occurrence probability prediction module is used to integrate the scenario probability distribution, the emotion probability distribution, the appeal probability distribution, and the customer behavior data to obtain a composite feature vector, and process the composite feature vector using a binary classification prediction model to obtain a complaint occurrence probability value;

[0015] A complaint reason determination module is used to determine whether the customer is a high-complaint-prone customer based on the complaint occurrence probability value. If the customer is a high-complaint-prone customer, the complaint reason is labeled using a pre-built labeling system to obtain a target complaint reason;

[0016] The appeasement strategy formulation module is used to formulate an appeasement strategy based on the target complaint cause and feed back the target complaint cause and the appeasement strategy to the seat staff.

[0017] On the third aspect, in order to solve the above-mentioned technical problems, an embodiment of the present application also provides a computer device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the customer complaint probability prediction method as described above.

[0018] Fourthly, in order to solve the above technical problems, an embodiment of the present application further provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method for predicting the probability of customer complaints as described above.

[0019] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0020] By acquiring historical customer call data and customer behavior data, we provide a basis for analyzing the probability of customer complaints. This not only improves the overall efficiency of data acquisition, but also enables the formulation of appeasement strategies based on the analysis results, thereby increasing customer conversion rates.

[0021] By acquiring historical customer call data and customer behavior data, we provide a basis for analyzing the probability of customer complaints. This not only improves the overall efficiency of data acquisition, but also enables the formulation of appeasement strategies based on the analysis results, thereby increasing customer conversion rates.

[0022] By integrating multi-dimensional information and performing subsequent complaint probability calculations based on composite feature vectors, the comprehensiveness of the predictions is improved, thereby increasing the accuracy of complaint probability predictions. By processing composite feature vectors using binary classification prediction models (such as logistic regression and XGBoost), the model can integrate multi-dimensional information and more accurately capture the underlying patterns of complaint occurrence, improving the accuracy of complaint probability values ​​and enabling the pre-emptive formulation of appeasement strategies to improve customer conversion rates.

[0023] By calculating the probability of a complaint and comparing it with a preset threshold, we can more accurately identify customers with a high tendency to complain, providing a basis for formulating subsequent appeasement strategies and thus improving customer conversion rates. Using a pre-built labeling system, we can automatically label the causes of complaints, reducing the workload and time cost of manual labeling, improving the efficiency of identifying complaint causes, and thus quickly formulating appeasement strategies, thereby improving customer conversion rates.

[0024] By formulating soothing strategies based on the target complaint causes, agents can more accurately solve customer problems and provide personalized soothing measures, thereby improving customer conversion rates. By feeding back the target complaint causes and soothing strategies to agents, agents can continuously optimize their handling methods based on the feedback information, improve the efficiency and quality of complaint handling, and thereby improve customer conversion rates. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the solutions in this application, a brief introduction will be given below to the drawings required for use in the description of the embodiments of this application. Obviously, the drawings described below are some embodiments of this application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0026] Figure 1 is an exemplary system architecture diagram to which the present application may be applied;

[0027] Figure 2 A flowchart of an embodiment of a method for predicting the probability of customer complaints according to the present application;

[0028] Figure 3 This is a schematic structural diagram of an embodiment of a device for predicting the probability of customer complaints according to the present application;

[0029] Figure 4 It is a structural diagram of an embodiment of a device according to the present application. DETAILED DESCRIPTION

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of the application are for the purpose of describing specific embodiments only and are not intended to limit this application. The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0031] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0032] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0033] like Figure 1As shown, system architecture 100 may include a terminal device 101, a network 102, and a server 103. Terminal device 101 may be a laptop computer 1011, a tablet computer 1012, or a mobile phone 1013. Network 102 is a medium for providing a communication link between terminal device 101 and server 103. Network 102 may include various connection types, such as wired or wireless communication links or fiber optic cables.

[0034] The user can use the terminal device 101 to interact with the server 103 via the network 102 to receive or send messages, etc. Various communication client applications can be installed on the terminal device 101, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0035] The terminal device 101 can be various electronic devices with a display screen and supporting web browsing. In addition to the laptop computer 1011, tablet computer 1012 or mobile phone 1013, the terminal device 101 can also be an e-book reader, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 (Moving Picture Experts Group Audio Layer IV) player, a laptop computer and a desktop computer, etc.

[0036] The server 103 may be a server that provides various services, such as a background server that provides support for web pages displayed on the terminal device 101 .

[0037] It should be noted that the method for predicting the customer complaint probability provided in the embodiment of the present application is generally executed by a server / terminal device. Accordingly, the device for predicting the customer complaint probability is generally set in the server / terminal device.

[0038] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is merely illustrative. Any number of terminal devices, networks and servers may be provided as required.

[0039] Continue to refer Figure 2, shows a flowchart of an embodiment of a method for predicting the probability of customer complaints according to the present application. Depending on different needs, the order of the steps in the flowchart can be changed, and some steps can be omitted. The method for predicting the probability of customer complaints provided by the embodiment of the present application can be applied to any scenario where product recommendation scripts need to be generated, and the method for predicting the probability of customer complaints can be applied to products in these scenarios. The method for predicting the probability of customer complaints includes the following steps:

[0040] Step S201: Acquire the customer's historical call data and customer behavior data.

[0041] In this embodiment, the method for predicting the probability of customer complaints is executed on the electronic device (eg Figure 1 The server / terminal device shown in the figure) can obtain the customer's historical call data and customer behavior data through a wired connection or a wireless connection. It should be noted that the above-mentioned wireless connection methods may include but are not limited to 3G / 4G / 5G connections, WiFi connections, Bluetooth connections, WiMAX connections, Zigbee connections, UWB (ultra wideband) connections, and other wireless connection methods currently known or to be developed in the future.

[0042] In this embodiment, a connection is established with a customer relationship management database, and the customer's historical call data and customer behavior data are obtained from the customer relationship management database through a data interface. The historical call data includes but is not limited to call records between the customer and customer service; the customer behavior data includes but is not limited to risk level, historical complaint records, call duration, speaking speed, speech volume, etc.

[0043] In this embodiment, in order to prevent the waste of computing resources due to excessive data, historical call data is acquired by an incremental data capture method. For new customers (i.e., customers who have not been analyzed), the system acquires all their historical call records; for existing customers (i.e., customers who have been analyzed), the system only acquires call records added after the last analysis to improve data processing efficiency. After acquiring the historical call data, the acquired historical call data is preprocessed, including removing invalid data, standardizing processing, and converting data formats to ensure the accuracy of subsequent analysis; customer behavior data is acquired through multi-channel integration, obtaining the customer's risk level and historical complaint records from each business system, and extracting the customer's call duration, speaking speed, and speech volume from the call history data as customer behavior data.

[0044] In this embodiment, by obtaining the customer's historical call data and customer behavior data, an analysis basis is provided for the subsequent probability value of customer complaints. This not only improves the overall efficiency of data acquisition, but also allows the formulation of appeasement strategies based on the analysis results, thereby improving customer conversion rates.

[0045] In one embodiment, after obtaining the customer's historical call data and customer behavior data, the method further includes:

[0046] Extracting timbre features from the historical call data to obtain timbre features to be identified;

[0047] Matching the to-be-identified timbre feature with a pre-stored timbre feature library to obtain the customer timbre feature and the customer service timbre feature;

[0048] Performing a speech-to-text operation on the historical call data using a preset speech conversion technology to obtain historical call text;

[0049] The historical call text is annotated according to the customer's voice characteristics and the customer service's voice characteristics to obtain annotated historical call text.

[0050] In this embodiment, the historical call data is preprocessed, and the preprocessing operation includes normalization, denoising, etc. Normalization refers to converting the historical call data into a standard format. For example, when the historical call data exists in a non-standard format, the historical call data is converted into a standard format. Denoising refers to using a noise reduction algorithm to perform denoising on the historical call data to remove the environmental noise existing during the call process and improve the accuracy of subsequent feature extraction. After the historical call data is preprocessed, the timbre feature extraction algorithm is used to extract the timbre feature of the historical call data to obtain the timbre feature to be identified, wherein the timbre feature to be identified includes the customer timbre feature and the customer service timbre feature. The timbre feature extraction algorithm includes but is not limited to Mel-frequency cepstral coefficients (MFCC) or linear prediction cepstral coefficients (LPCC), etc. The multiple timbre features extracted by the timbre feature extraction algorithm are combined to obtain a multi-dimensional feature vector as the timbre feature to be identified. Pre-store A timbre feature library is used to collect the timbre features of all customer service staff and store them in the timbre feature library. The timbre features to be identified are matched with the timbre features stored in the timbre feature library by similarity, wherein the similarity matching includes a cosine similarity algorithm, etc. When the similarity value is greater than a preset similarity threshold, the timbre feature to be identified is classified as a customer service timbre feature. When the similarity value is less than or equal to the similarity threshold, the timbre feature to be identified is classified as a customer service timbre feature, and finally the classified customer timbre features and customer service timbre features are obtained; the historical call data is converted into text form by a preset speech-to-text technology, such as ASR technology, etc. Since the timbre feature extraction and speech-to-text operations are based on different processing flows, it is necessary to ensure that the customer timbre features, customer service timbre features and historical call texts are aligned in time, and the customer timbre features and customer service timbre features are annotated on the historical call text in chronological order to obtain an annotated historical call text.

[0051] In this embodiment, customer voice features and customer service voice features are classified from historical call data, and the converted historical call text is annotated with the extracted customer voice features and customer service voice features to obtain annotated historical call text, which can be subsequently analyzed based on the historical call data with customer annotated information, thereby improving the accuracy of complaint prediction.

[0052] Step S202: extracting features from the historical call data using a multi-task classification model to obtain target global features, and calculating the target global features using the classification layer of the multi-task classification model to output a scene probability distribution, an emotion probability distribution, and a demand probability distribution;

[0053] In this embodiment, the historical call data obtained in step S201 is input into a pre-trained multi-task classification model. The multi-task classification model adopts a deep learning architecture, including an output layer, a shared feature extraction layer, a cross-attention layer, three independent classification layers and an output layer, which are respectively used for scene classification, emotion classification and appeal classification; the feature extraction layer adopts a BERT-based pre-trained language model structure, which can effectively capture the semantic information in the historical call data. The feature extraction layer extracts features from the historical call data and converts the text of the historical call data into a high-dimensional vector (such as 768 dimensions). The classification layer of the multi-task classification model outputs a scene probability distribution, an emotion probability distribution and an appeal probability distribution, wherein the scene categories include multiple predefined scenarios such as billing category, business processing, product consultation, and service complaint; the emotion categories include multiple emotion categories such as satisfaction, general, dissatisfaction, anger, and confusion; the appeal categories include multiple appeal categories such as refund request, service improvement, information consultation, and complaint escalation. The three classification layers of the multi-task classification model output the scene probability distribution, emotion probability distribution and appeal probability distribution respectively.

[0054] In this embodiment, feature extraction is performed on historical call data through a multi-task classification model. The model can mine deep semantic information in call texts, thereby improving the accuracy of model feature extraction. The classification layer of the multi-task classification model is used to perform parallel calculations on target global features, while outputting scene probability distribution, emotion probability distribution, and appeal probability distribution. This not only improves the computational efficiency of the model, but also improves the accuracy of multi-dimensional information presentation, thereby improving the accuracy and efficiency of subsequent complaint predictions.

[0055] In one embodiment, before extracting features from the historical call data using the multi-task classification model, the method further includes:

[0056] Pre-build an initial multi-task classification model;

[0057] Obtaining labeled training data and unlabeled training data, performing label prediction on the unlabeled data to obtain a confidence value, filtering the unlabeled data according to the confidence value to obtain pseudo-labeled data, mixing the labeled data and the unlabeled data to obtain a dataset, and dividing the dataset into a training set, a test set, and a validation set;

[0058] The initial multi-task classification model is trained using the training set, global training features of the training set are extracted through a feature extraction layer, attention weights between classification tasks in each global training feature are calculated through a cross attention layer, and weighted fusion is performed based on the attention weights. The probability distribution of the weighted fused global training features is output through a classification layer to obtain a trained initial multi-task classification model;

[0059] Validating the trained initial multi-task classification model using the validation set to obtain an evaluation result, and optimizing the trained initial multi-task classification model based on the evaluation result by calling a corresponding optimization method;

[0060] The optimized initial multi-task classification model is tested using the test set to obtain a trained multi-task classification model.

[0061] In this embodiment, an initial multi-task classification model is pre-constructed. The model architecture of the initial multi-task classification model has been mentioned above and will not be elaborated here. Labeled data and unlabeled data are obtained, wherein the labeled data includes text, emotion labels, scene labels, and appeal labels; for the unlabeled data set, a pre-trained model (such as BERT) is used to predict the unlabeled data to generate prediction results of emotions, scenes, and appeals, and samples with prediction confidence values ​​higher than the confidence threshold (such as 0.9) are screened from the unlabeled data, and the prediction results are used as pseudo-labels to obtain pseudo-labeled data, and the labeled data and pseudo-labeled data are mixed to obtain a data set, and the data set is divided into a training set, a validation set, and a test set for model training, validation, and testing; the initial multi-task classification model is trained with the training set, and the deep semantic features of the training set are extracted through the feature extraction layer. In which, the feature extraction layer captures the contextual information in the text through its multi-layer Transformer structure, and calculates the attention weight of any classification task and the other two classification tasks through the attention weight calculation formula on the cross attention layer. The features are fused based on the attention weight, and three probability distributions are output through three independent classification layers. The cross entropy loss is calculated for each classification task separately, the gradient is calculated according to the loss function, and the model parameters are updated through back propagation to obtain the trained initial multi-task classification model. The trained initial multi-classification task is verified by the validation set to obtain the evaluation result. According to the evaluation result, the model parameters are optimized using the Adam optimizer to obtain the optimized initial multi-classification task model. The optimized initial multi-task classification model is tested using the test set to obtain the trained multi-task classification model.

[0062] In this example, by mixing labeled and pseudo-labeled data, we leverage both high-quality labeled data and the potential of unlabeled data, improving overall data utilization efficiency. The introduction of a cross-attention layer enables feature interaction between different classification tasks. This mechanism enables the model to dynamically focus on the features most relevant to the current task, enhancing the semantic connections between tasks, helping to improve classification accuracy and the accuracy of model predictions, thereby increasing the accuracy of subsequent complaint probabilities.

[0063] In one embodiment, the step of extracting features from the historical call data using a multi-task classification model to obtain target global features includes:

[0064] Using a word segmentation tool to perform a word segmentation operation on the historical call data to obtain a word segmentation sequence, and performing an encoding operation on the word segmentation sequence to obtain a word segmentation encoding sequence;

[0065] Extracting features from the word segmentation coding sequence using a multi-head self-attention mechanism and a feedforward neural network in the feature extraction layer of the multi-task classification model to obtain global features;

[0066] Constructing a classification task set, and calculating, based on each classification task in the classification task set, an attention score of the classification task and other classification tasks using the global feature and a preset attention calculation formula;

[0067] The global feature is adjusted based on the attention score to obtain the target global feature.

[0068] In this embodiment, a word segmentation tool (such as BERT's WordPiece or Jieba) is used to segment historical call data to obtain a word segmentation sequence. This word segmentation sequence is then encoded to obtain an initial word segmentation code sequence. Since the subsequent self-attention mechanism cannot identify the position of the word segmentation sequence, a position encoding vector needs to be inserted into the word segmentation sequence to obtain a word segmentation code sequence. The word segmentation code sequence is input into a trained multi-task classification model. The multi-head self-attention mechanism and feed-forward neural network (FFN) in the feature extraction layer of the multi-task classification model extract context-related features of the word segmentation code sequence to obtain a global feature. A classification task set is constructed based on business needs, including a scene classification task, a sentiment classification task, and a demand classification task. A query vector is initialized for each classification task. For each classification task, the dimension of the query vector is consistent with the dimension of the global vector. The scaled dot product attention formula is used to calculate the similarity between the query vector and the global feature for each classification task to obtain an attention score. The attention score is used as a weight to perform a weighted summation on the global features to generate an adjusted global feature, thereby obtaining a target global feature.

[0069] In this embodiment, by introducing the attention mechanism, voice information can be shared and transmitted between subsequent different classification tasks, thereby enhancing the model's ability to understand complex semantics, thereby improving the effect of subsequent model predictions, that is, improving the accuracy of the probability of subsequent complaints.

[0070] In one embodiment, the step of calculating the target global features through the classification layer of the multi-task classification model and outputting the scene probability distribution, the emotion probability distribution, and the appeal probability distribution includes:

[0071] Inputting the target global feature into the classification layer of the multi-task classification model, and computing the target global feature in parallel through each sub-classifier of the classification layer;

[0072] Performing a linear transformation on the target global feature using the weight matrix and bias term of each sub-classifier to obtain the original score of each category;

[0073] According to the original scores of each category, a preset conversion function is used to perform probability distribution conversion on the original scores to obtain the scene probability distribution, the emotion probability distribution and the appeal probability distribution.

[0074] In this embodiment, the classification layer of the multi-task classification model is composed of multiple parallel sub-classifier layers, each sub-classifier is responsible for a specific task (such as the sub-classifier for scene classification, the sub-classifier for emotion classification, and the sub-classifier for appeal classification are all responsible for corresponding specific tasks). The target global feature is input into the classification layer of the multi-task classification model, and the target global feature is calculated in parallel by each sub-classifier of the classification layer. The target global feature is linearly transformed using the weight matrix and bias term of each sub-classifier to obtain the original score of each category. Specifically, each sub-classifier calculates the original score using the following formula:

[0075] score i =W i *feature+b i

[0076] Among them, W i refers to the weight matrix of the i-th sub-classifier, b i Refers to the bias term of the i-th subclassifier, feature is the target global feature vector, score i is the original score vector output by the i-th sub-classifier.

[0077] According to the original scores of each category, a preset conversion function (such as a softmax function) is used to convert the original scores into probability distribution to obtain scene probability distribution, emotion probability distribution and appeal probability distribution, where each probability value is between 0 and 1, and the sum of the probabilities of all categories is 1.

[0078] In this embodiment, the target global features are calculated in parallel by the sub-classifiers of the classification layer, and the model can process multiple related tasks at the same time, thereby improving the computational efficiency of the model; the target global features are linearly transformed using the weight matrix and bias term of the sub-classifier, and the original features are mapped to the category space of different tasks, thereby improving the expressive power of the model; the distribution probability of each task is calculated based on the original score, which is beneficial to the calculation of subsequent complaint predictions and improves the accuracy of the probability of complaint occurrence.

[0079] Step S203: Integrate the scenario probability distribution, the emotion probability distribution, the demand probability distribution, and the customer behavior data to obtain a composite feature vector, and process the composite feature vector using a binary classification prediction model to obtain a complaint occurrence probability value;

[0080] In this embodiment, the scenario probability distribution, emotion probability distribution, and appeal probability distribution obtained in step S202 are directly spliced ​​to obtain a call feature vector, and the customer behavior data is subjected to feature engineering processing. Specifically, feature extraction is performed on the customer behavior data to obtain customer behavior features, and the call feature vector and the customer behavior features are integrated to obtain a composite feature vector. The composite feature vector is input into a pre-trained binary classification prediction model. The binary classification prediction model adopts a gradient boosting decision tree (GBDT) algorithm and is trained with historical customer complaint data. It can effectively predict the customer's complaint tendency. Specifically, the composite feature vector is input into the binary classification prediction model, and a complaint occurrence probability value is output, which indicates the probability that the customer will file a complaint.

[0081] In this embodiment, by integrating multi-dimensional information and performing subsequent complaint probability calculations based on composite feature vectors, the comprehensiveness of the prediction is improved, thereby improving the accuracy of the complaint probability prediction; by processing the composite feature vectors through a binary prediction model (such as logistic regression, XGBoost), the model can integrate multi-dimensional information, more accurately capture the potential patterns of complaint occurrence, improve the accuracy of the complaint probability value, and thus formulate a soothing strategy in advance to improve customer conversion rate.

[0082] In one embodiment, the composite feature vector is processed by a binary classification prediction model to obtain a complaint occurrence probability value, including:

[0083] Inputting the composite feature vector into the binary classification prediction model, and outputting a hierarchical prediction score according to each leaf node in the decision tree of the binary classification prediction model;

[0084] Based on the weight corresponding to each leaf node, the corresponding hierarchical prediction score is weighted and summed according to the weight to obtain the prediction score of the decision tree;

[0085] The predicted score is converted using a preset probability conversion function to obtain the complaint occurrence probability value.

[0086] In this embodiment, it is mentioned above that the composite feature vector usually contains information of multiple dimensions, such as scenario probability distribution, emotion probability distribution, appeal probability distribution and customer behavior data; the composite feature vector is input into a pre-trained binary classification prediction model, and the corresponding hierarchical prediction score is output through each leaf node in the decision tree, wherein the construction of the decision tree is calculated by minimizing the loss calculation function, and the optimal splitting feature and splitting point are recursively selected, and each splitting point is used as a leaf node so that the leaf node outputs the corresponding hierarchical prediction score, and a corresponding weight is assigned to each leaf node according to the learning rate. Based on the weight corresponding to each leaf node and the hierarchical prediction score output by each leaf node, all hierarchical prediction scores are weightedly summed to obtain the prediction score of the decision tree, and the prediction score of the decision tree is converted into a probability value through a predicted probability conversion function (such as a sigmoid function) to obtain the probability value of complaint occurrence;

[0087] The weighted summation formula is:

[0088]

[0089] Wherein, S refers to the prediction score of the decision number, T refers to the number of leaf nodes of the decision tree, and a t Refers to the weight of the tth leaf node, f t (X) refers to the hierarchical prediction score of the t-th leaf node;

[0090] The formula for using the preset conversion function is:

[0091]

[0092] Wherein, P refers to the probability value of complaint occurrence, e refers to the natural logarithm, and S refers to the prediction score of the decision tree.

[0093] In this embodiment, a hierarchical prediction score is output according to each leaf node in the decision tree of the binary prediction model; based on the weight corresponding to each leaf node, the corresponding hierarchical prediction score is weightedly summed according to the weight to obtain the prediction score of the decision tree; the prediction score is converted by a preset probability conversion function to obtain the probability value of the complaint occurrence, which not only realizes the effective processing of the composite feature vector, but also improves the prediction accuracy of the complaint occurrence probability through the decision tree and weighted summation mechanism, thereby improving the conversion rate of subsequent customers.

[0094] S204: Determine whether the customer is a high-complaint-prone customer based on the complaint occurrence probability value. If the customer is a high-complaint-prone customer, label the complaint reason using a pre-built labeling system to obtain a target complaint reason.

[0095] In this embodiment, the complaint occurrence probability value obtained in step S203 is compared with a pre-set complaint probability threshold to determine whether the customer is a high complaint tendency customer, that is, when the complaint occurrence probability value is greater than the complaint probability threshold (such as 0.7), the customer is determined to be a high complaint tendency customer, and when the complaint occurrence probability value is less than or equal to the complaint probability threshold, the customer is determined to be a low complaint tendency customer; when the customer is a high complaint tendency customer, possible complaint reasons are extracted from historical call data, and historical complaint reasons are extracted and summarized to obtain complaint reasons to be labeled, and the complaint reasons are labeled using a pre-built labeling system to obtain target complaint reasons, wherein the pre-built labeling system adopts a combination of rule-based and machine learning methods to label the complaint reasons to be labeled, the rule-based method can quickly label known complaint reasons, and the machine learning method can be used to process complex text data to improve the degree of automation and accuracy of labeling.

[0096] In this embodiment, by calculating the probability value of complaint occurrence and comparing it with the preset threshold, customers with a high tendency to complain can be identified more accurately, which provides a basis for the subsequent formulation of appeasement strategies, thereby improving customer conversion rates; using a pre-built labeling system, the causes of complaints can be automatically labeled, reducing the workload and time cost of manual labeling, improving the efficiency of identifying the causes of complaints, and then quickly formulating appeasement strategies to improve customer conversion rates.

[0097] In one embodiment, when the customer is a customer with a high tendency to complain, the step of labeling the complaint reason using a pre-built labeling system to obtain the target complaint reason includes:

[0098] When the customer is a customer with a high complaint tendency, obtaining the customer's historical call data and historical complaint reasons;

[0099] Identifying the reasons for the pending complaint from the historical call data, summarizing the reasons for the pending complaint and the historical complaint reasons to obtain the reasons for the complaint to be marked;

[0100] The complaint reason to be labeled is labeled using a rule-based labeling method in the pre-built labeling system to obtain a first labeled complaint reason;

[0101] Based on the complaint reasons to be labeled that cannot be labeled by the rule-based labeling method, labeling is performed using the pre-built machine learning labeling method in the labeling system to obtain a second labeled complaint reason;

[0102] The target complaint reason is filtered out from the first labeled complaint reason and the second labeled complaint reason according to the complaint category and timestamp of the first labeled complaint reason and the second labeled complaint reason.

[0103] In this embodiment, when the customer is a customer with a high tendency to complain, the customer's historical call data and historical complaint reasons (if any) are obtained, and based on the historical call data, the historical call data is converted into text by the ASR speech-to-text method to obtain the historical call text, and the keywords of the historical call text are extracted. Based on the keywords, the natural language processing (NLP) tool or custom rules are used to identify the reasons for the complaints to be filed, and the identified reasons for the complaints to be filed are summarized with the historical complaint reasons to obtain the reasons for the complaints to be labeled; the rule labeling method in the pre-built labeling system is used to label the reasons for the complaints to be labeled, and the rule labeling method is used to label the reasons for the complaints to be labeled. Specifically, According to the rules defined by business knowledge and expert experience, complaint reasons of known patterns are quickly processed to obtain the first labeled complaint reasons. For complaint reasons to be labeled that cannot be labeled by the rule-based labeling method, a pre-trained machine learning model is used for labeling. Specifically, a trained classification model (such as SVM, random forest, deep learning model, etc.) is used to automatically identify and label complaint reasons. The results of rule labeling and machine learning labeling are summarized to obtain the first labeled complaint reasons and the second labeled complaint reasons. According to the complaint category and timestamp, the target complaint reasons are screened out from the first labeled complaint reasons and the second labeled complaint reasons, and the screened out target complaint reasons are sampled and verified to ensure the accuracy of the labeling.

[0104] In this embodiment, by obtaining the customer's historical call data and historical complaint reasons, and identifying the reasons for the complaint from the historical call data, the comprehensiveness of the complaint reasons is improved. By labeling the complaint reasons through a pre-built labeling system, it is possible to improve the labeling efficiency while ensuring accuracy, and filter out the most likely complaint reasons based on timestamps and complaint categories, thereby improving the accuracy of subsequent appeasement strategy formulation and thereby improving customer conversion rate.

[0105] Step S205: formulate a soothing strategy according to the target complaint cause, and feed back the target complaint cause and the soothing strategy to the agent.

[0106] In this embodiment, according to the target complaint cause determined in step S204, a suitable appeasement strategy is selected from a preset appeasement strategy library, wherein the appeasement strategy library is a structured knowledge base containing processing solutions, communication techniques, compensation plans, and other contents for different complaint causes.

[0107] The formulation of the aforementioned appeasement strategies can follow predefined principles, such as: 1. Targetedness: Appeasement strategies should address the specific cause of the complaint and provide targeted solutions; 2. Gradual Approach: Appeasement measures at different levels should be implemented based on the severity of the complaint and the customer's value; 3. Prevention: Appeasement strategies should not only address the current problem but also prevent similar problems from recurring; 4. Customer Experience: Appeasement strategies should focus on improving the customer experience and emphasize communication methods and attitudes;

[0108] In this embodiment, based on the target complaint cause, the most matching appeasement strategy template is retrieved from the appeasement strategy library, and the strategy template is personalized according to the specific situation of the customer (such as customer level, etc.) to form a final appeasement strategy. Specifically, the content of the appeasement strategy includes: helping the agent to accurately understand and confirm the problems that the customer may face; providing appropriate appeasement and empathy based on the customer's emotional state; providing specific solution steps and methods for the target complaint cause; providing appropriate compensation suggestions based on customer value and problems; and formulating a follow-up plan after the problem is resolved. The target complaint cause and the formulated appeasement strategy are fed back to the agent through the customer service workbench interface. The feedback content can be presented in a structured form, including complaint risk warnings, possible complaint causes, recommended appeasement strategies and other information. Based on the information provided by the system, the agent can carry out targeted preventive appeasement during the communication process with the customer, reduce the possibility of customer complaints, and thus improve customer conversion rate.

[0109] In this embodiment, by formulating a soothing strategy based on the target complaint cause, the agent can more accurately solve the customer's problem and provide personalized soothing measures, thereby improving the customer conversion rate; the target complaint cause and soothing strategy are fed back to the agent, and the agent can continuously optimize his or her handling method based on the feedback information, improve the efficiency and quality of complaint handling, and thereby improve the customer conversion rate.

[0110] It should be emphasized that in order to further ensure the privacy and security of the above-mentioned customers' historical call data and customer behavior data, the above-mentioned customers' historical call data and customer behavior data can also be stored in a blockchain node.

[0111] The blockchain referred to in this application is a new application model for computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Blockchain is essentially a decentralized database, a series of data blocks generated using cryptographic methods. Each data block contains information about a batch of network transactions, which is used to verify the validity of this information (to prevent counterfeiting) and generate the next block. Blockchain can include the underlying blockchain platform, the platform product service layer, and the application service layer.

[0112] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Artificial Intelligence (AI) is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to achieve optimal results.

[0113] Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0114] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware via computer-readable instructions. The computer-readable instructions can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes in the above-described method embodiments. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0115] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0116] Further references Figure 3 , as a response to the above Figure 2 The present application provides an embodiment of a device for predicting the probability of customer complaints. Figure 2 Corresponding to the method embodiment shown, the apparatus can be specifically applied to various computer devices.

[0117] like Figure 3As shown, the customer complaint probability prediction device 300 described in this embodiment includes: a data acquisition module 301, a multidimensional distribution calculation module 302, a complaint probability prediction module 303, a complaint cause determination module 304, and a soothing strategy formulation module 305. Among them:

[0118] Data acquisition module 301, used to obtain the customer's historical call data and customer behavior data;

[0119] In one embodiment, the apparatus further comprises:

[0120] A timbre extraction module is used to extract timbre features from the historical call data to obtain timbre features to be identified;

[0121] A timbre matching module is used to match the timbre feature to be identified with a pre-stored timbre feature library to obtain the customer timbre feature and the customer service timbre feature;

[0122] A text conversion module, configured to convert the historical call data into text using a preset voice conversion technology to obtain the historical call text;

[0123] The timbre marking module is used to mark the historical call text according to the customer's timbre characteristics and the customer service's timbre characteristics to obtain the marked historical call text.

[0124] A multi-dimensional distribution calculation module 302 is configured to extract features from the historical call data using a multi-task classification model to obtain target global features, calculate the target global features using the classification layer of the multi-task classification model, and output a scene probability distribution, an emotion probability distribution, and a demand probability distribution;

[0125] In one embodiment, the apparatus further comprises:

[0126] Initial model building module, used to pre-build the initial multi-task classification model;

[0127] a training data acquisition module, configured to acquire labeled training data and unlabeled training data, perform label prediction on the unlabeled data to obtain a confidence value, filter the unlabeled data based on the confidence value to obtain pseudo-labeled data, mix the labeled data with the unlabeled data to obtain a dataset, and divide the dataset into a training set, a test set, and a validation set;

[0128] A model training module is used to train the initial multi-task classification model using the training set, extract global training features of the training set through a feature extraction layer, calculate attention weights between classification tasks in each global training feature through a cross attention layer, perform weighted fusion based on the attention weights, and output probability distribution of the weighted fused global training features through a classification layer to obtain a trained initial multi-task classification model;

[0129] A model verification module is used to verify the trained initial multi-task classification model using the verification set to obtain an evaluation result, and to optimize the trained initial multi-task classification model based on the evaluation result by calling a corresponding optimization method;

[0130] The model testing module is used to test the optimized initial multi-task classification model using the test set to obtain a trained multi-task classification model.

[0131] In one embodiment, the multidimensional distribution calculation module includes:

[0132] A word segmentation submodule, configured to perform a word segmentation operation on the historical call data using a word segmentation tool to obtain a word segmentation sequence, and perform an encoding operation on the word segmentation sequence to obtain a word segmentation encoding sequence;

[0133] A global feature extraction submodule, configured to extract features from the word segmentation encoding sequence using a multi-head self-attention mechanism and a feedforward neural network of the feature extraction layer of the multi-task classification model to obtain global features;

[0134] An attention score calculation submodule is used to construct a classification task set, and based on each classification task in the classification task set, calculate the attention score of the classification task and other classification tasks using the global feature and a preset attention calculation formula;

[0135] A feature adjustment submodule is used to adjust the global feature based on the attention score to obtain the target global feature.

[0136] In one embodiment, the multidimensional distribution calculation module includes:

[0137] A parallel computing submodule, configured to input the target global feature into the classification layer of the multi-task classification model, and to parallelly compute the target global feature through each sub-classifier of the classification layer;

[0138] A linear transformation submodule, configured to perform a linear transformation on the target global feature using the weight matrix and bias term of each subclassifier to obtain an original score for each category;

[0139] The probability conversion submodule is used to convert the original scores of each category into probability distributions using a preset conversion function to obtain the scene probability distribution, the emotion probability distribution and the appeal probability distribution.

[0140] The complaint occurrence probability prediction module 303 is used to integrate the scenario probability distribution, the emotion probability distribution, the appeal probability distribution, and the customer behavior data to obtain a composite feature vector, and process the composite feature vector using a binary classification prediction model to obtain a complaint occurrence probability value;

[0141] In one embodiment, the complaint occurrence probability prediction module includes:

[0142] A hierarchical prediction score calculation submodule is used to input the composite feature vector into the binary classification prediction model and output a hierarchical prediction score according to each leaf node in the decision tree of the binary classification prediction model;

[0143] A prediction score submodule, configured to perform weighted summation of the corresponding hierarchical prediction scores according to the weights corresponding to the leaf nodes to obtain a prediction score of the decision tree;

[0144] The score conversion submodule is used to convert the predicted score through a preset probability conversion function to obtain the complaint occurrence probability value.

[0145] The complaint reason determination module 304 is configured to determine whether the customer is a high-complaint-prone customer based on the complaint occurrence probability value. If the customer is a high-complaint-prone customer, the complaint reason is labeled using a pre-built labeling system to obtain a target complaint reason.

[0146] In one embodiment, the complaint cause determination module includes:

[0147] A complaint reason acquisition submodule is used to acquire the customer's historical call data and historical complaint reasons when the customer is a high-complaint-tendency customer;

[0148] A complaint reason summary submodule, configured to identify the reasons for the pending complaint from the historical call data, summarize the reasons for the pending complaint and the historical complaint reasons, and obtain the reasons for the complaint to be marked;

[0149] A first complaint reason labeling submodule is used to label the complaint reason to be labeled using a rule labeling method in the pre-built labeling system to obtain a first labeled complaint reason;

[0150] A second complaint reason labeling submodule is configured to label the complaint reasons to be labeled that cannot be labeled by the rule labeling method by using the pre-built machine learning labeling method in the labeling system to obtain a second labeled complaint reason;

[0151] The target complaint reason determination submodule is used to filter out the target complaint reason from the first labeled complaint reason and the second labeled complaint reason based on the complaint category and timestamp of the first labeled complaint reason and the second labeled complaint reason.

[0152] The appeasement strategy formulation module 305 is used to formulate an appeasement strategy according to the target complaint reason, and feed back the target complaint reason and the appeasement strategy to the seat staff.

[0153] In this embodiment, by obtaining the customer's historical call data and customer behavior data, a basis for subsequent analysis of the probability of customer complaints is provided. This not only improves the overall efficiency of data acquisition, but also enables the formulation of appeasement strategies based on the analysis results, thereby improving customer conversion rates.

[0154] By acquiring historical customer call data and customer behavior data, we provide a basis for analyzing the probability of customer complaints. This not only improves the overall efficiency of data acquisition, but also enables the formulation of appeasement strategies based on the analysis results, thereby increasing customer conversion rates.

[0155] By integrating multi-dimensional information and performing subsequent complaint probability calculations based on composite feature vectors, the comprehensiveness of the predictions is improved, thereby increasing the accuracy of complaint probability predictions. By processing composite feature vectors using binary classification prediction models (such as logistic regression and XGBoost), the model can integrate multi-dimensional information and more accurately capture the underlying patterns of complaint occurrence, improving the accuracy of complaint probability values ​​and enabling the pre-emptive formulation of appeasement strategies to improve customer conversion rates.

[0156] By calculating the probability of a complaint and comparing it with a preset threshold, we can more accurately identify customers with a high tendency to complain, providing a basis for formulating subsequent appeasement strategies and thus improving customer conversion rates. Using a pre-built labeling system, we can automatically label the causes of complaints, reducing the workload and time cost of manual labeling, improving the efficiency of identifying complaint causes, and thus quickly formulating appeasement strategies, thereby improving customer conversion rates.

[0157] By formulating soothing strategies based on the target complaint causes, agents can more accurately solve customer problems and provide personalized soothing measures, thereby improving customer conversion rates. By feeding back the target complaint causes and soothing strategies to agents, agents can continuously optimize their handling methods based on the feedback information, improve the efficiency and quality of complaint handling, and thereby improve customer conversion rates.

[0158] In order to solve the above technical problems, the embodiment of the present application also provides a device (computer device). Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0159] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected through a system bus. It should be noted that the figure only shows a computer device 4 with a memory 41, a processor 42, and a network interface 43, but it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to a microprocessor, an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a digital signal processor (DSP), an embedded device, etc.

[0160] The computer device may be a desktop computer, notebook computer, PDA, cloud server, etc. The computer device may interact with the user via a keyboard, mouse, remote control, touchpad, or voice control device.

[0161] The memory 41 includes at least one type of readable storage medium, including flash memory, a hard disk, a multimedia card, a card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, a magnetic disk, an optical disk, etc. In some embodiments, the memory 41 may be an internal storage unit of the computer device 4, such as the hard disk or memory of the computer device 4. In other embodiments, the memory 41 may also be an external storage device of the computer device 4, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash memory card, etc. equipped on the computer device 4. Of course, the memory 41 may also include both the internal storage unit of the computer device 4 and its external storage device. In this embodiment, the memory 41 is generally used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions for the customer complaint probability prediction method. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or are to be output.

[0162] In some embodiments, the processor 42 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor 42 is generally used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to execute computer-readable instructions stored in the memory 41 or process data, such as computer-readable instructions for executing the method for predicting the probability of customer complaints.

[0163] The network interface 43 may include a wireless network interface or a wired network interface. The network interface 43 is generally used to establish a communication connection between the computer device 4 and other electronic devices.

[0164] During implementation, the electronic device of this application obtains historical customer call data and customer behavior data, providing a basis for analyzing the probability of subsequent customer complaints. This not only improves the overall efficiency of data acquisition, but also enables the formulation of appeasement strategies based on the analysis results, thereby improving customer conversion rates.

[0165] By acquiring historical customer call data and customer behavior data, we provide a basis for analyzing the probability of customer complaints. This not only improves the overall efficiency of data acquisition, but also enables the formulation of appeasement strategies based on the analysis results, thereby increasing customer conversion rates.

[0166] By integrating multi-dimensional information and performing subsequent complaint probability calculations based on composite feature vectors, the comprehensiveness of the predictions is improved, thereby increasing the accuracy of complaint probability predictions. By processing composite feature vectors using binary classification prediction models (such as logistic regression and XGBoost), the model can integrate multi-dimensional information and more accurately capture the underlying patterns of complaint occurrence, improving the accuracy of complaint probability values ​​and enabling the pre-emptive formulation of appeasement strategies to improve customer conversion rates.

[0167] By calculating the probability of a complaint and comparing it with a preset threshold, we can more accurately identify customers with a high tendency to complain, providing a basis for formulating subsequent appeasement strategies and thus improving customer conversion rates. Using a pre-built labeling system, we can automatically label the causes of complaints, reducing the workload and time cost of manual labeling, improving the efficiency of identifying complaint causes, and thus quickly formulating appeasement strategies, thereby improving customer conversion rates.

[0168] By formulating soothing strategies based on the target complaint causes, agents can more accurately solve customer problems and provide personalized soothing measures, thereby improving customer conversion rates. By feeding back the target complaint causes and soothing strategies to agents, agents can continuously optimize their handling methods based on the feedback information, improve the efficiency and quality of complaint handling, and thereby improve customer conversion rates.

[0169] The present application also provides another embodiment, namely, providing a storage medium (computer-readable storage medium), wherein the computer-readable storage medium stores computer-readable instructions, and the computer-readable instructions can be executed by at least one processor to enable the at least one processor to perform the steps of the method for predicting the probability of customer complaints as described above.

[0170] During implementation, the computer-readable storage medium of the present application obtains historical customer call data and customer behavior data, providing a basis for analyzing the probability of subsequent customer complaints. This not only improves the overall efficiency of data acquisition, but also enables the formulation of appeasement strategies based on the analysis results, thereby improving customer conversion rates.

[0171] By acquiring historical customer call data and customer behavior data, we provide a basis for analyzing the probability of customer complaints. This not only improves the overall efficiency of data acquisition, but also enables the formulation of appeasement strategies based on the analysis results, thereby increasing customer conversion rates.

[0172] By integrating multi-dimensional information and performing subsequent complaint probability calculations based on composite feature vectors, the comprehensiveness of the predictions is improved, thereby increasing the accuracy of complaint probability predictions. By processing composite feature vectors using binary classification prediction models (such as logistic regression and XGBoost), the model can integrate multi-dimensional information and more accurately capture the underlying patterns of complaint occurrence, improving the accuracy of complaint probability values ​​and enabling the pre-emptive formulation of appeasement strategies to improve customer conversion rates.

[0173] By calculating the probability of a complaint and comparing it with a preset threshold, we can more accurately identify customers with a high tendency to complain, providing a basis for formulating subsequent appeasement strategies and thus improving customer conversion rates. Using a pre-built labeling system, we can automatically label the causes of complaints, reducing the workload and time cost of manual labeling, improving the efficiency of identifying complaint causes, and thus quickly formulating appeasement strategies, thereby improving customer conversion rates.

[0174] By formulating soothing strategies based on the target complaint causes, agents can more accurately solve customer problems and provide personalized soothing measures, thereby improving customer conversion rates. By feeding back the target complaint causes and soothing strategies to agents, agents can continuously optimize their handling methods based on the feedback information, improve the efficiency and quality of complaint handling, and thereby improve customer conversion rates.

[0175] The non-Company software tools or components appearing in the embodiments of this application are merely examples and do not represent actual use.

[0176] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0177] Obviously, the embodiments described above are only some of the embodiments of the present application, rather than all of the embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application has been described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions described in the aforementioned specific embodiments, or to make equivalent replacements for some of the technical features therein. Any equivalent structure made using the contents of the present application specification and the accompanying drawings, directly or indirectly used in other related technical fields, is also within the scope of patent protection of the present application.

Claims

1. A method for predicting the probability of customer complaints, characterized in that: The method comprises: Obtain customers' historical call data and customer behavior data; Extract features from the historical call data using a multi-task classification model to obtain target global features, and calculate the target global features using the classification layer of the multi-task classification model to output a scene probability distribution, an emotion probability distribution, and a demand probability distribution; Integrate the scenario probability distribution, the emotion probability distribution, the demand probability distribution, and the customer behavior data to obtain a composite feature vector, and process the composite feature vector using a binary classification prediction model to obtain a complaint occurrence probability value; Determine whether the customer is a high-complaint-tendency customer based on the complaint occurrence probability value. If the customer is a high-complaint-tendency customer, label the complaint reason using a pre-built labeling system to obtain a target complaint reason. A soothing strategy is formulated based on the target complaint cause, and the target complaint cause and the soothing strategy are fed back to the agent.

2. The method for predicting the probability of customer complaints according to claim 1, wherein: After obtaining the customer's historical call data and customer behavior data, the method further includes: Extracting timbre features from the historical call data to obtain timbre features to be identified; Matching the to-be-identified timbre feature with a pre-stored timbre feature library to obtain the customer timbre feature and the customer service timbre feature; Performing a speech-to-text operation on the historical call data using a preset speech conversion technology to obtain historical call text; The historical call text is annotated according to the customer's voice characteristics and the customer service's voice characteristics to obtain annotated historical call text.

3. The method for predicting the probability of customer complaints according to claim 1, wherein: Before extracting features from the historical call data using the multi-task classification model, the method further includes: Pre-build an initial multi-task classification model; Obtaining labeled training data and unlabeled training data, performing label prediction on the unlabeled data to obtain a confidence value, filtering the unlabeled data according to the confidence value to obtain pseudo-labeled data, mixing the labeled data and the unlabeled data to obtain a dataset, and dividing the dataset into a training set, a test set, and a validation set; The initial multi-task classification model is trained using the training set, global training features of the training set are extracted through a feature extraction layer, attention weights between classification tasks in each global training feature are calculated through a cross attention layer, and weighted fusion is performed based on the attention weights. The probability distribution of the weighted fused global training features is output through a classification layer to obtain a trained initial multi-task classification model; Validating the trained initial multi-task classification model using the validation set to obtain an evaluation result, and optimizing the trained initial multi-task classification model based on the evaluation result by calling a corresponding optimization method; The optimized initial multi-task classification model is tested using the test set to obtain a trained multi-task classification model.

4. The method for predicting the probability of customer complaints according to claim 1, wherein: The step of extracting features from the historical call data using a multi-task classification model to obtain target global features includes: Using a word segmentation tool to perform a word segmentation operation on the historical call data to obtain a word segmentation sequence, and performing an encoding operation on the word segmentation sequence to obtain a word segmentation encoding sequence; Extracting features from the word segmentation coding sequence using a multi-head self-attention mechanism and a feedforward neural network in the feature extraction layer of the multi-task classification model to obtain global features; Constructing a classification task set, and calculating, based on each classification task in the classification task set, an attention score of the classification task and other classification tasks using the global feature and a preset attention calculation formula; The global feature is adjusted based on the attention score to obtain the target global feature.

5. The method for predicting the probability of customer complaints according to claim 1, wherein: The step of calculating the target global features through the classification layer of the multi-task classification model and outputting the scene probability distribution, the emotion probability distribution and the appeal probability distribution includes: Inputting the target global feature into the classification layer of the multi-task classification model, and computing the target global feature in parallel through each sub-classifier of the classification layer; Performing a linear transformation on the target global feature using the weight matrix and bias term of each sub-classifier to obtain the original score of each category; According to the original scores of each category, a preset conversion function is used to perform probability distribution conversion on the original scores to obtain the scene probability distribution, the emotion probability distribution and the appeal probability distribution.

6. The method for predicting the probability of customer complaints according to claim 1, wherein: The composite feature vector is processed by the binary classification prediction model to obtain a complaint occurrence probability value, including: Inputting the composite feature vector into the binary classification prediction model, and outputting a hierarchical prediction score according to each leaf node in the decision tree of the binary classification prediction model; Based on the weight corresponding to each leaf node, the corresponding hierarchical prediction score is weighted and summed according to the weight to obtain the prediction score of the decision tree; The predicted score is converted using a preset probability conversion function to obtain the complaint occurrence probability value.

7. The method for predicting the probability of customer complaints according to claim 1, wherein: When the customer is a customer with a high tendency to complain, the steps of labeling the complaint reason through a pre-built labeling system to obtain the target complaint reason include: When the customer is a customer with a high complaint tendency, obtaining the customer's historical call data and historical complaint reasons; Identifying the reasons for the pending complaint from the historical call data, summarizing the reasons for the pending complaint and the historical complaint reasons to obtain the reasons for the complaint to be marked; The complaint reason to be labeled is labeled using a rule-based labeling method in the pre-built labeling system to obtain a first labeled complaint reason; Based on the complaint reasons to be labeled that cannot be labeled by the rule-based labeling method, labeling is performed using the pre-built machine learning labeling method in the labeling system to obtain a second labeled complaint reason; The target complaint reason is filtered out from the first labeled complaint reason and the second labeled complaint reason according to the complaint category and timestamp of the first labeled complaint reason and the second labeled complaint reason.

8. A device for predicting the probability of customer complaints, characterized in that: The device comprises: Data acquisition module, used to obtain customers' historical call data and customer behavior data; a multidimensional distribution calculation module, configured to extract features from the historical call data using a multi-task classification model to obtain target global features, calculate the target global features using the classification layer of the multi-task classification model, and output a scene probability distribution, an emotion probability distribution, and an appeal probability distribution; A complaint occurrence probability prediction module is used to integrate the scenario probability distribution, the emotion probability distribution, the appeal probability distribution, and the customer behavior data to obtain a composite feature vector, and process the composite feature vector using a binary classification prediction model to obtain a complaint occurrence probability value; A complaint reason determination module is used to determine whether the customer is a high-complaint-prone customer based on the complaint occurrence probability value. If the customer is a high-complaint-prone customer, the complaint reason is labeled using a pre-built labeling system to obtain a target complaint reason; The appeasement strategy formulation module is used to formulate an appeasement strategy based on the target complaint cause and feed back the target complaint cause and the appeasement strategy to the seat staff.

9. A computer device, characterized in that: The computer device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method for predicting the customer complaint probability according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for predicting the probability of customer complaints according to any one of claims 1 to 7 is implemented.

Images