Customer complaint responsibility analysis method and device, readable storage medium, program product

Abstract

This application discloses a customer complaint liability determination analysis method and apparatus, a readable storage medium, and a program product. The method includes: querying a preset knowledge base based on the complaint content of a target customer and a preset retrieval enhancement model to obtain a set of candidate evidence related to the complaint content, wherein the preset knowledge base stores historical business data of different customers; semantically aligning the complaint content and each candidate piece of evidence in the candidate evidence set based on a cross-attention model to obtain a multimodal fusion feature vector corresponding to each candidate piece of evidence for the complaint content; and predicting the probability of each type of liability cause corresponding to the complaint content based on the multimodal fusion feature vector of each candidate piece of evidence and a preset liability cause classifier, wherein the preset liability cause classifier includes a classification space of multiple predefined types of liability causes.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a customer complaint liability determination analysis method and apparatus, a readable storage medium, and a program product. Background Technology

[0002] In the existing customer service system, when a customer files a complaint, the complaint handling personnel need to query the customer's historical business data, and then, based on the analysis of the above data, determine which responsible department or the customer's own reasons caused the complaint (such as billing errors, network quality, misunderstanding of the package, business violations, etc.).

[0003] Existing automated complaint liability determination methods use a plain text classification model to directly determine the cause of liability based on the complaint text. However, relying solely on the complaint text results in a single data dimension, making it difficult to obtain accurate customer complaint liability determination results. Summary of the Invention

[0004] The purpose of this application is to provide a customer complaint liability determination analysis method and apparatus, readable storage medium, and program product to solve the problem of inaccurate existing customer complaint liability determination results.

[0005] To solve the above-mentioned technical problems, this specification is implemented as follows: Firstly, it provides a method for analyzing and assigning responsibility for customer complaints, including: Based on the complaint content of the target customer and the preset retrieval enhancement model, a preset knowledge base is queried to obtain a set of candidate evidence related to the complaint content. The preset knowledge base stores historical business data of different customers. Based on the cross-attention model, the complaint content and each candidate piece of evidence in the candidate evidence set are semantically aligned to obtain the multimodal fusion feature vectors of the complaint content corresponding to each candidate piece of evidence. Based on the multimodal fusion feature vectors of each candidate piece of evidence and a preset cause-of-responsibility classifier, the probability of each type of cause of responsibility corresponding to the complaint content is predicted. The preset cause-of-responsibility classifier includes a classification space of multiple predefined types of cause of responsibility.

[0006] Optionally, the multimodal fusion feature vectors of each candidate piece of evidence corresponding to the complaint content are obtained, including: The text of the complaint is input into a pre-trained first text encoder for semantic encoding to obtain the semantic feature vector corresponding to the complaint content; Each candidate piece of evidence in the candidate evidence set is input into a pre-trained second text encoder for semantic encoding to obtain the semantic feature vector corresponding to each candidate piece of evidence. The semantic feature vector corresponding to the complaint content and the semantic feature vector corresponding to each candidate piece of evidence in the candidate evidence set are input into the cross-attention model for semantic alignment, so as to output the multimodal fusion feature vector of each candidate piece of evidence corresponding to the complaint content.

[0007] Optionally, after predicting the probability of each type of cause of liability corresponding to the content of the complaint, the method further includes: Input the semantic feature vectors corresponding to each candidate piece of evidence in the candidate evidence set into the preset evidence scoring model to obtain the evidence score set of the degree of support of the corresponding candidate piece of evidence for the target responsibility cause of the type with the highest probability. Construct an evidence chain corresponding to the cause of the target responsibility based on the evidence score set; Based on the stated reasons for the target responsibility and the corresponding chain of evidence, a responsibility determination analysis report corresponding to the content of the complaint is generated.

[0008] Optionally, before inputting the semantic feature vectors corresponding to each candidate piece of evidence in the candidate evidence set into the preset evidence scoring model, the method further includes training a responsibility classification model containing the first text encoder, the second text encoder, the cross-attention model, and the preset responsibility cause classifier, and training the preset evidence scoring model, specifically including: An objective function is constructed based on the classification loss corresponding to the responsibility classification model, which includes a first text encoder, a second text encoder, a cross-attention model, and a preset responsibility cause classifier, the scoring loss corresponding to the preset evidence scoring model, and the regularization loss of the trainable parameters corresponding to the responsibility classification model and the preset evidence scoring model. With the overall loss of minimizing the classification loss, scoring loss, and regularization loss as the optimization objective, the trainable parameters of the responsibility classification model and the preset evidence scoring model are adjusted until the trainable parameters converge or reach the preset number of iterations, thereby training the final responsibility classification model and the preset evidence scoring model.

[0009] Optionally, constructing an evidence chain corresponding to the cause of the target responsibility based on the evidence score set includes: Compare the score of each piece of evidence in the evidence score set with a preset score threshold; Identify at least one key piece of evidence whose score exceeds the preset score threshold; The at least one key piece of evidence is identified as the chain of evidence corresponding to the cause of the target's responsibility.

[0010] Optionally, after obtaining the set of evidence scores for the degree of support of each semantic feature vector for the target responsibility cause corresponding to the type with the highest probability, the following are also included: Calculate the average evidence score corresponding to the at least one key piece of evidence; If the maximum probability of the preset cause of responsibility classifier corresponding to the target cause of responsibility is greater than the first confidence threshold and the average evidence score is greater than the preset score threshold, then the target cause of responsibility predicted by the preset cause of responsibility classifier is adopted. If the maximum value of the preset responsibility cause classifier corresponding to the target responsibility cause is less than the second confidence threshold, then the responsibility cause corresponding to the complaint content is determined to be the customer's own cause, and the second confidence level is less than the first confidence threshold; If the maximum value of the preset responsibility cause classifier corresponding to the target responsibility cause is not greater than the first confidence threshold and not less than the second confidence threshold, then a manual review process is triggered to process the complaint content.

[0011] Optionally, it also includes: Based on the aforementioned reasons for the target responsibility, a related query can be performed to determine the corresponding responsibility attribution and responsible department; The contract terms of the target customer, historical cases similar to the complaints of the target customer, and / or rules of liability determination are consulted to determine the basis for determining the cause of liability of the target customer. Add the content of the complaint, the attribution of responsibility, the responsible department, and the basis for the judgment to the liability determination analysis report.

[0012] In a second aspect, a customer complaint liability determination analysis device is provided, including a processor and a memory, wherein the memory stores a program or instructions that can run on the processor, and the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0013] Thirdly, a readable storage medium is provided that stores a program or instructions which, when executed by a processor, implement the steps of the method described in the first aspect.

[0014] Fourthly, a computer program product is provided, comprising a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform the steps of the method described in the first aspect.

[0015] In this embodiment, a set of candidate evidence related to the complaint content is obtained by querying a preset knowledge base based on the complaint content of the target customer and a preset retrieval enhancement model. The preset knowledge base stores historical business data of different customers. A cross-attention model is used to semantically align the complaint content and each candidate piece of evidence in the candidate evidence set, resulting in multimodal fusion feature vectors corresponding to each candidate piece of evidence for the complaint content. Based on the multimodal fusion feature vectors of each candidate piece of evidence and a preset responsibility cause classifier, the probability of each type of responsibility cause corresponding to the complaint content is predicted. The preset responsibility cause classifier includes a classification space of multiple predefined types of responsibility causes. Therefore, based on the target customer's full lifecycle business data, multimodal or multi-source evidence (including a large amount of background data and long-term context of the complaint content) related to the complaint content can be queried. After semantic alignment with the input content, the complaint content incorporates the most relevant information from all candidate evidence, enabling accurate prediction of the responsibility cause of the complaint content. This also achieves a high degree of automation in complaint liability determination, significantly improving processing efficiency, reducing labor costs, and shortening the response cycle. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the customer complaint liability determination analysis method according to an embodiment of this application.

[0017] Figure 2 This is a schematic diagram of the execution modules corresponding to each step of the customer complaint liability determination analysis method in this application embodiment.

[0018] Figure 3 This is a structural block diagram of the customer complaint liability determination analysis device according to an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. The drawing numbers in this application are only used to distinguish the various steps in the solution and are not used to limit the execution order of the various steps. The specific execution order is subject to the description in the specification.

[0020] To address the problems existing in the prior art, embodiments of this application provide a method for analyzing and determining liability in customer complaints, such as... Figure 1As shown, the process includes steps 102 to 106.

[0021] Step 102: Based on the complaint content of the target customer and the preset retrieval enhancement model, query the preset knowledge base to obtain a set of candidate evidence related to the complaint content. The preset knowledge base stores historical business data of different customers.

[0022] Customer complaints can be received in the form of complaint tickets. When processing these tickets, the complaint content is input into a pre-defined retrieval enhancement model, such as a Retrieval-Augmented Generation (RAG) model. This model can query a pre-defined knowledge base to retrieve candidate evidence related to the complaint, accurately recall relevant evidence, and construct a candidate evidence set. The pre-defined knowledge base stores historical business data for different customers. This knowledge base may include external knowledge bases, contract term databases, historical ticket databases, historical case databases, business rule databases, and business databases recording billing records, call logs, tariffs, and service plans for different customers. The retrieved evidence consists of events or records directly and / or indirectly related to the complaint content, reflecting the reasons or causes that might have led the target customer to file the complaint.

[0023] Step 104: Based on the cross-attention model, perform semantic alignment between the complaint content and each candidate piece of evidence in the candidate evidence set to obtain the multimodal fusion feature vectors of the complaint content corresponding to each candidate piece of evidence.

[0024] Semantic alignment ensures that the complaint content and candidate evidence convey the same meaning. This step can be achieved by fusing the information from the complaint content and candidate evidence using a cross-attention model, resulting in multimodal fusion feature vectors corresponding to each candidate piece of evidence for the same complaint content. For example, if the same complaint content corresponds to n candidate pieces of evidence, semantic alignment between the complaint content and each of the retrieved candidate pieces of evidence can yield n multimodal fusion feature vectors corresponding to each candidate piece of evidence for the same complaint content.

[0025] Based on the solution provided in the above embodiments, optionally, in step 104 above, obtaining the multimodal fusion feature vector of each candidate evidence corresponding to the complaint content includes: inputting the text of the complaint content into a pre-trained first text encoder for semantic encoding to obtain the semantic feature vector corresponding to the complaint content; inputting each candidate piece of evidence in the candidate evidence set into a pre-trained second text encoder for semantic encoding to obtain the semantic feature vector corresponding to each candidate piece of evidence; inputting the semantic feature vector corresponding to the complaint content and the semantic feature vector corresponding to each candidate piece of evidence in the candidate evidence set into a cross-attention model for semantic alignment to output the multimodal fusion feature vector of each candidate piece of evidence corresponding to the complaint content.

[0026] In the above embodiments, features are constructed for the complaint content and candidate evidence, respectively.

[0027] The original complaints submitted by target customers through multiple channels (such as telephone, mobile app, website, etc.) are input into a pre-trained text encoder for semantic encoding to obtain the semantic feature vector corresponding to the complaint content. If the target customer submits the original complaint content in voice form, it is first converted into text using Automatic Speech Recognition (ASR) technology.

[0028] Specifically, the steps to obtain the semantic feature vector of the text corresponding to the complaint content are described below.

[0029]

[0030] Among them, input The text representing the content of the complaint. For text encoders based on pre-trained deep learning models, such as Transformer, the output is... Indicates the corresponding input The semantic feature vector of the complaint content after semantic encoding.

[0031] Specifically, the steps for obtaining the semantic feature vectors corresponding to each candidate piece of evidence are described below.

[0032]

[0033] Among them, input For candidate evidence set The multiple sources of evidence can include contract terms, historical work orders, billing records, image recognition text, automatic speech-to-text conversion, etc. For a pre-trained text encoder, the output is... Indicates the corresponding input The semantic feature vector of the candidate evidence modality (text / structure) after semantic encoding.

[0034] By performing the above semantic encoding process on all candidate evidence in the candidate evidence set, a set of semantic feature vectors of the candidate evidence set can be output. .

[0035] Then, based on the semantic feature vectors corresponding to the complaint content and the candidate evidence set obtained from the above steps, semantic alignment is performed using a cross-attention model to obtain the multimodal fusion feature vectors of each candidate evidence corresponding to the complaint content.

[0036] Specifically, the semantic alignment between the complaint text and multimodal candidate evidence is achieved through a cross-attention model, resulting in a multimodal fusion feature vector as shown in the following formula:

[0037] in, This represents the multimodal fusion feature vector of the complaint content and corresponding candidate evidence. This represents a cross-attention model (e.g., a Transformer layer).

[0038] The following are the semantic feature vectors corresponding to the text containing the complaint input. semantic feature vectors corresponding to candidate evidence Taking this as an example, we will describe the semantic alignment processing of the cross-attention model.

[0039] First, a projection transformation is performed on the semantic feature vector. and semantic feature vectors Each is achieved through a learnable linear transformation matrix (usually 1). ) are mapped to the query (Q), key (K), and value (V) space.

[0040] Then, the attention weights are calculated. The semantic feature vectors are then calculated. With each semantic feature vector The similarity of the keys (usually expressed as a dot product) is normalized using a normalized exponential function (Softmax) to obtain the attention weights as shown in the following formula:

[0041] Attention weight Reflects candidate evidence The importance of understanding the content of complaints from target customers.

[0042] Using attention weights Candidate evidence value vector We perform weighted summation to obtain a context-aware fusion representation, i.e., a multimodal fusion feature vector.

[0043] Using the algorithm described above, a fusion feature matrix can be obtained, which includes the multimodal fusion feature vectors corresponding to all candidate evidence in the candidate evidence set. This matrix is ​​a single dense vector. In standard cross-attention, if the semantic feature vector corresponding to the complaint content is a single vector, then the output multimodal fusion feature vector is also a single vector. Fusion Feature Matrix It dynamically integrates the most relevant information from all candidate evidence and aligns it semantically with the content of the complaint.

[0044] Step 106: Based on the multimodal fusion feature vectors of each candidate evidence and a preset cause-of-responsibility classifier, predict the probability of each type of cause of responsibility corresponding to the complaint content. The preset cause-of-responsibility classifier includes a classification space of multiple predefined types of cause of responsibility.

[0045] The fusion feature matrix output in step 104 above Multimodal fusion feature vector This information is input into a preset liability cause classifier, which can then predict the probability of each type of liability cause corresponding to the complaint content.

[0046] The classification output of the preset cause-of-fault classifier is shown in the following formula:

[0047] in, Represents the classification weight matrix; This represents a multimodal fusion feature vector; This represents the classification bias vector, with dimensions equal to the number of responsibility causes of a predefined type. This indicates the probability that the complaint content corresponds to each type of cause of liability, or the predictive confidence level of each type of cause of liability.

[0048] Different probabilities reflect the degree to which the corresponding cause of responsibility is likely to cause the target customer's complaint. The higher the probability, the more likely the complaint is to fall under the corresponding category of cause of responsibility. Usually, the cause of responsibility with the highest probability is determined as the reason for assigning responsibility for the target customer's complaint.

[0049] The default cause-of-fault classifier can be a fully connected layer (classification head), and the classification weight matrix of this layer... and bias terms Used to map to the classification space of responsibility cause types.

[0050] Through the above steps, based on the business data of the target customer throughout the entire life cycle, we can query multimodal or multi-source evidence (including a large amount of background data and long-term context of the complaint) related to the complaint content in various dimensions. After semantic alignment with the input content, the complaint content integrates the most relevant information from all candidate evidence, thereby enabling accurate prediction of the cause of responsibility for the complaint content.

[0051] In one specific embodiment, optionally, after predicting the probability of each type of cause of responsibility corresponding to the complaint content, the method further includes: inputting the semantic feature vector corresponding to each candidate evidence in the candidate evidence set into a preset evidence scoring model to obtain an evidence score set of the degree of support of each candidate feature for the target cause of responsibility of the type with the highest probability; constructing an evidence chain corresponding to the target cause of responsibility based on the evidence score set; and generating a responsibility determination analysis report corresponding to the complaint content based on the target cause of responsibility and the corresponding evidence chain.

[0052] In this embodiment, each candidate piece of evidence in the set of candidate evidence related to the complaint is further scored to obtain a corresponding evidence score. The evidence score is used to quantify the degree to which each candidate piece of evidence supports the determination of responsibility for the current complaint, providing a basis for the subsequent generation of interpretable responsibility determination analysis reports, calibration of responsibility rules, and manual review.

[0053] Specifically, candidate evidence is scored using the following formula:

[0054] in, This represents the scoring weight matrix. Represents the scoring bias scalar ( A scoring task applied to a single candidate piece of evidence (dimension 1). Represented as the Sigmoid activation function, Indicate candidate evidence Evidence score.

[0055] The input to the pre-defined evidence scoring model is a single candidate piece of evidence. Corresponding semantic feature vector ,in Then, the preset evidence scoring model is a lightweight evidence scoring network, which consists of a linear layer with a sigmoid activation function. The linear transformation parameters include the scoring weight matrix. and scoring bias scalar Where d represents the encoder dimension of the evidence scoring network, Indicates spatial range.

[0056] Candidate evidence Evidence score The higher the score, the better the candidate evidence. The more relevant and credible the evidence is to the determination of responsibility for the current complaint of the target customer, the greater the set of evidence scores for all n candidate pieces of evidence. It can be used to construct candidate evidence. The chain of evidence for the cause of responsibility for the type of target with the highest probability.

[0057] The pre-defined evidence scoring model is independent of the pre-defined cause-of-responsibility classifier, focusing on fine-grained evaluation at the evidence level, which complements the global cause-of-responsibility determination of the pre-defined cause-of-responsibility classifier.

[0058] Specifically, constructing the evidence chain corresponding to the cause of the target responsibility based on the evidence score set includes: comparing the scores of each piece of evidence in the evidence score set with a preset score threshold; determining at least one key piece of evidence whose score exceeds the preset score threshold; and determining the at least one key piece of evidence as the evidence chain corresponding to the cause of the target responsibility.

[0059] As mentioned above, the higher the evidence score, the more relevant and credible the candidate evidence is to the determination of responsibility for the current complaint of the target customer, meaning it is more likely to reflect the cause of the complaint. Thus, candidate evidence with scores exceeding a preset threshold may be considered key evidence supporting the cause of the target's responsibility. This key evidence constitutes a chain of evidence supporting the determination of the cause of the target's responsibility, reflecting the investigation results and indicating the key fields triggering the determination of responsibility (such as "Article 3 of the Contract Terms" and "Abnormal Traffic on Bill 2024-07-01").

[0060] By combining the reasons for the target's responsibility with the corresponding chain of evidence, a responsibility determination analysis report is generated corresponding to the complaint content of the target customer. Key evidence can be highlighted, which improves the interpretability and compliance of the customer system. When responding to the complaint ticket of the target customer, the responsibility reasons and chain of evidence are automatically populated to generate a responsibility determination analysis report, achieving intelligent responsibility determination with high accuracy and strong interpretability. This significantly improves the automation level of the customer service system and meets the compliance requirements of heavily regulated industries, which is beneficial to compliance audits and customer communication.

[0061] After automatically predicting the cause of responsibility for a complaint using the aforementioned model, there may be conflicts between the preset liability determination rules and the model. In such cases, a step of calibrating the preset liability determination rules can be introduced.

[0062] In one specific embodiment, after obtaining the evidence score set of the degree of support for the target responsibility cause corresponding to the type with the highest probability for each semantic feature vector, the method further includes: calculating the average evidence score corresponding to the at least one key piece of evidence; if the maximum probability of the preset responsibility cause classifier corresponding to the target responsibility cause is greater than a first confidence threshold and the average evidence score is greater than a preset score threshold, then the target responsibility cause predicted by the preset responsibility cause classifier is adopted; if the maximum probability of the preset responsibility cause classifier corresponding to the target responsibility cause is less than a second confidence threshold, then the responsibility cause corresponding to the complaint content is determined to be the customer's own cause, and the second confidence is less than the first confidence threshold; if the maximum probability of the preset responsibility cause classifier corresponding to the target responsibility cause is neither greater than the first confidence threshold nor less than the second confidence threshold, then a manual review process is triggered to process the complaint content.

[0063] This embodiment describes how to calibrate liability determination rules based on the evidence scores of each candidate piece of evidence. The calibration of liability determination rules can utilize a rule engine, whose decision logic deeply integrates the probabilities of various types of liability causes automatically output by a preset liability cause classifier. Evidence scores of candidate evidence This forms a multi-layered calibration and fallback mechanism. Among them, This represents the k-th element predicted by the pre-defined cause-of-fact classifier. , The pre-defined cause-of-fault classifier indicates that the complaint falls under the category of... The probability of each type of cause of liability. .

[0064] Depending on the scenario, specific rules include strong rule priority coverage, high model confidence adoption, default fallback rules, and / or manual review triggering.

[0065] In scenarios where strong rules take precedence, if the customer system detects that the target customer has uploaded explicit "attachments related to liability determination" (such as officially stamped complaint receipts, court judgments, etc.), then regardless of the liability cause prediction result output by the preset liability cause classifier, it will directly determine that "the company is liable." This is the highest priority hard rule.

[0066] In scenarios where the model adopts a high-confidence approach, the maximum probability output by the pre-defined cause-of-fact classifier is... ( Indicates the maximum confidence threshold, for example ), and the mean score of the key evidence supporting the most probable cause of liability ( supports )> ( This indicates a preset score threshold, for example... If the prediction result of the preset cause of responsibility classifier is adopted, the cause of responsibility corresponding to the highest probability will be output directly.

[0067] In the default fallback scenario, if the target customer uploads no liability-related attachments, and the maximum probability predicted by the preset liability cause classifier is lower than the minimum confidence threshold... (For example If the result is negative, it will be determined by default to the target customer's own reasons (such as misunderstanding of the package, operational errors, etc.).

[0068] In scenarios triggered by manual review, the customer system will automatically transfer the target customer's complaint ticket to the manual review queue if any of the following situations occur: The maximum probability predicted by the pre-defined cause-of-fact classifier is in the middle range. The pre-defined cause of responsibility classifier predicts a cause of responsibility that conflicts with a key business rule (e.g., the predicted cause of responsibility is a "network quality" problem, but backend data shows that the network indicators in that area are completely normal); there is a serious discrepancy in the evidence scores of key evidence (e.g., both the affirmative and negative evidence scores are very high).

[0069] In addition, the rule engine combines "automated accountability + auditable traceability + light human intervention" to calibrate the model prediction results and achieve rule priority coverage or two-way feedback optimization in key scenarios.

[0070] Optionally, the method further includes: querying the corresponding liability attribution and responsible department based on the cause of the target liability; querying the contract terms of the target customer, historical cases similar to the complaint content of the target customer, and / or judgment rules to determine the basis for judgment of the cause of the target liability; and adding the complaint content, the attribution of liability, the responsible department, and the basis for judgment to the liability determination analysis report.

[0071] The customer service system can use the model's predicted cause of responsibility to query the corresponding attribution of responsibility and the responsible department, obtaining a responsibility conclusion including attribution, cause of responsibility, and responsible department. It can also query the target customer's contract terms, historical cases similar to the target customer's complaints, and / or liability judgment rules to determine the basis for liability judgment, including contract terms, historical cases, and rules. Furthermore, it can obtain a description of the complaint based on the complaint content. By combining the above complaint description, the verification results of the evidence chain, the basis for liability judgment, the liability conclusion, and customer satisfaction and feedback, a standardized liability determination analysis report is generated.

[0072] Therefore, by combining the automatic output of liability analysis reports, the entire liability determination process can be automated, reducing human intervention and improving the accuracy and compliance of liability determination.

[0073] As described above, the customer complaint liability determination analysis method of this application embodiment can automatically output the liability cause prediction result. The corresponding model includes a first text encoder for the complaint text, a second text encoder for candidate evidence, a cross-attention model for semantic alignment, a preset liability cause classifier for outputting the liability cause prediction probability, and a preset evidence scoring model for scoring candidate evidence. Before practical application, the above models need to be trained. The training method for the overall model is described below as a whole.

[0074] Optionally, before inputting the semantic feature vectors corresponding to each candidate piece of evidence in the candidate evidence set into the preset evidence scoring model, the method further includes training a responsibility classification model containing the first text encoder, the second text encoder, the cross-attention model, and the preset responsibility cause classifier, and training the preset evidence scoring model. Specifically, this includes: constructing an objective function based on the classification loss corresponding to the responsibility classification model containing the first text encoder, the second text encoder, the cross-attention model, and the preset responsibility cause classifier, the scoring loss corresponding to the preset evidence scoring model, and the regularization loss of the trainable parameters corresponding to the responsibility classification model and the preset evidence scoring model; adjusting the trainable parameters of the responsibility classification model and the preset evidence scoring model with the optimization objective of minimizing the overall loss of the classification loss, the scoring loss, and the regularization loss, until the trainable parameters converge or reach a preset number of iterations, thereby training the final responsibility classification model and the preset evidence scoring model.

[0075] First, we design the loss function for the entire model. The objective function for training the overall model is:

[0076] in, This represents the cross-entropy classification loss corresponding to the responsibility classification model, which includes a first text encoder, a second text encoder, a cross-attention model, and a preset responsibility cause classifier. This represents the scoring loss corresponding to the preset evidence scoring model. This represents the regularization loss of the trainable parameters corresponding to the responsibility classification model and the preset evidence scoring model, such as the L2 regularization term. The weighting coefficients represent the classification loss. The weighting coefficients representing the scoring loss. These represent the weights of the regularization term, used to control the impact of each component on the total loss. The proportion of contribution.

[0077] The classification loss (cross-entropy) is expressed as follows:

[0078] in, A category index representing the cause of liability, assuming there is A predefined category of liability cause (e.g., "billing error", "network quality", "package misunderstanding", "customer's own reasons", etc.) then .

[0079] The first element in the one-hot encoded vector of the true label There are 10 elements. For a specific sample of complaint content, there is only one true cause of responsibility. Therefore, in the one-hot encoded vector, the corresponding category is 10 elements. All other categories .

[0080] Describes the first in the predicted probability distribution The element, output by the preset cause-of-responsibility classifier, represents the element that the model considers to belong to the category of the first element. The probability of a cause of liability.

[0081] This represents the classification loss, which drives the model to continuously optimize the accuracy of its cause-of-fact predictions by minimizing the cross-entropy between the true label and the predicted probability distribution. When the model predicts the probability of the true category... The closer it is to 1, the smaller the loss value.

[0082] The scoring loss of evidence (contrastive learning) is expressed as follows:

[0083] in, This indicates the number of positive and negative sample pairs in the training batch.

[0084] Indicates the first The evidence score represents the support of each positive example (i.e., evidence that contributes to the determination of the correct cause of liability).

[0085] Indicates the first The evidence score for the support of each negative example of evidence (i.e. evidence that is irrelevant to or contradicts the determination of the correct cause of liability).

[0086] By employing the concept of contrastive learning, we ensure that the scores of positive evidence are always higher than those of negative evidence by a fixed interval, thereby improving the accuracy of evidence ranking.

[0087] The regularization loss (L2 weight decay) is expressed as follows:

[0088] This option is used to prevent overfitting and improve the model's generalization ability.

[0089] in, This represents the set of all trainable parameters in the overall model. In the embodiments of this application, The parameters include the following components or sub-models: the weights of the text encoder for the complaint content, the weights of the text encoder for candidate evidence, the query (Q), key (K), and value (V) projection matrices in the cross-attention model, and the classification weight matrix of the cause-of-fact classifier. and classification bias vector The scoring weight matrix of the pre-defined evidence scoring model and scoring bias scalar .

[0090] Represents the trainable parameter vector The square of the L2 norm, i.e., the sum of the squares of all trainable parameters. This operation penalizes large weights, encouraging the model to learn smaller, smoother parameter values. Regularization loss is added to the total loss function. The optimization process not only focuses on prediction accuracy, but also suppresses model complexity, thereby improving the model's generalization ability on unseen complaint data and effectively alleviating the overfitting problem.

[0091] The embodiments of this application can employ multimodal fusion modeling using graph neural networks (GNNs) and cross-attention (e.g., Transformers). Through the cross-attention mechanism, the semantic alignment of the complaint content text with candidate evidence from multiple sources is achieved, and joint modeling is performed based on temporal causal chains to output the responsibility cause category and the support score of each candidate piece of evidence.

[0092] Graph neural networks start with the business events triggered by customer complaints corresponding to each historical complaint ticket, and combine them with ticket logs, tariff change records, and package contract terms to construct temporal / causal chains. Through the propagation formula of the graph neural network, causal dependency modeling is achieved between multiple pieces of evidence represented by nodes, thereby determining the attribution of responsibility.

[0093] The propagation formula for a graph neural network is shown below:

[0094] in, Represents a node (This represents a specific event or piece of evidence, such as "high data charges were incurred on July 1, 2024") Represents a node After passing through the graph neural network... The semantic feature vector is updated after each round of training or message passing, and it incorporates information from its neighbors.

[0095] N(v) represents the node In the evidence graph, the set of first-order neighbor nodes is used. The evidence graph consists of multiple evidence-corresponding nodes and corresponding edges, obtained through training iterations. These neighbors represent nodes... Other events / evidence that have a direct causal or temporal relationship.

[0096] Representing neighboring nodes In the Layer embedding vector. Indicates the first The learnable weight matrix of a layered graph neural network is used to perform a linear transformation on neighbor information. Represents the normalization coefficient, usually a node. and nodes A certain combination of degrees, for example Its function is to prevent gradient explosion and balance the message strength from nodes of different degrees. This represents a nonlinear activation function (such as ReLU) used to introduce nonlinear expressive power.

[0097] This propagation formula iteratively aggregates information from neighboring nodes, so that the final representation of each piece of evidence not only contains its own information, but also encodes its context and logical relationship in the entire causal chain, thus providing a deeper reasoning basis for determining responsibility.

[0098] Then, combining the evidence graph obtained from the graph neural network, a cross-attention mechanism is used to achieve semantic alignment between the text of the complaint and different evidence from multiple sources. A graph neural network is used to model the causal evidence graph, and the cross-attention mechanism is combined to achieve deep integration between the complaint content and multi-source evidence, improving the accuracy and robustness of the model's determination of causal responsibility.

[0099] This application also proposes a novel mechanism for bidirectional calibration of rules and models. If the model's output prediction of the cause of responsibility is inconsistent with existing liability judgment rules, the liability judgment rules are corrected. If a liability judgment rule is invalid for an extended period, it is automatically fed back to the model for retraining. This forms an adaptive closed-loop optimization process, avoiding the problem of rigid liability judgment rules.

[0100] In this embodiment, a set of candidate evidence related to the complaint content is obtained by querying a preset knowledge base based on the complaint content of the target customer and a preset retrieval enhancement model. The preset knowledge base stores historical business data of different customers. A cross-attention model is used to semantically align the complaint content and each candidate piece of evidence in the candidate evidence set, resulting in multimodal fusion feature vectors corresponding to each candidate piece of evidence for the complaint content. Based on the multimodal fusion feature vectors of each candidate piece of evidence and a preset responsibility cause classifier, the probability of each type of responsibility cause corresponding to the complaint content is predicted. The preset responsibility cause classifier includes a classification space of multiple predefined types of responsibility causes. Therefore, based on the target customer's full lifecycle business data, multimodal or multi-source evidence (including a large amount of background data and long-term context of the complaint content) related to the complaint content can be queried. After semantic alignment with the input content, the complaint content incorporates the most relevant information from all candidate evidence, enabling accurate prediction of the responsibility cause of the complaint content. This also achieves a high degree of automation in complaint liability determination, significantly improving processing efficiency, reducing labor costs, and shortening the response cycle.

[0101] Optionally, such as Figure 2 As shown in the figure, this application embodiment also provides a customer complaint liability determination analysis system, the overall architecture of which is composed of a complaint access module 1, a data collection module 2, a retrieval enhancement module 3, a fusion liability determination module 4, a rule engine 5, and a liability report generation module 6.

[0102] like Figure 2 As shown, the system first receives text or voice complaints submitted by customers through multiple channels via the complaint access module 1, and then uses ASR technology to transcribe the voice into text. Subsequently, the data acquisition module 2 automatically associates and extracts background data from multiple sources, including the customer's work order records, business processing information, billing records, contract terms and attachments. Building upon this foundation, the retrieval enhancement module 3 uses complaint content as the query, combining vector semantic retrieval and rule matching mechanisms to accurately retrieve relevant evidence from the knowledge base, contract database, and historical case database, constructing a candidate evidence pool (set). The fusion accountability module 4 employs a multimodal model combining Transformer and Graph Neural Network (GNN), achieving semantic alignment between complaint text and multi-source evidence through a cross-attention mechanism and jointly modeling the evidence graph, outputting the category of responsibility cause and the support score of each piece of evidence. The rule engine 5 calibrates the prediction results of the fusion accountability module 4, achieving rule priority coverage or bidirectional feedback optimization in key scenarios. Finally, the accountability report generation module 6 automatically generates a standardized accountability analysis report containing responsibility attribution, accountability basis, evidence chain, and key field annotations, automatically incorporating the responsibility cause when processing responses, achieving highly accurate and highly interpretable intelligent accountability, significantly improving the automation level and compliance capabilities of the customer service system.

[0103] The modules in the customer complaint liability analysis system described in this specification can achieve the following: Figure 1 To avoid repetition, the various processes implemented in the corresponding steps of the method embodiment will not be described again here.

[0104] Optionally, such as Figure 3 As shown in the figure, this application embodiment also provides a customer complaint liability determination analysis device 2000, including a processor 2400 and a memory 2200. The memory 2200 stores a program or instructions that can be run on the processor 2400. When the program or instructions are executed by the processor 2400, they implement the various steps of the above-described customer complaint liability determination analysis method embodiment and can achieve the same technical effect. To avoid repetition, they will not be described again here.

[0105] This application also provides a readable storage medium storing a program or instructions. When executed by a processor, the program or instructions implement the various processes of any of the above-described customer complaint liability analysis method embodiments and achieve the same technical effect. To avoid repetition, further details are omitted here. The readable storage medium includes computer-readable storage media, such as read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] This application also provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program. The computer program is operable to enable a computer to execute the various processes of any of the above-described customer complaint liability analysis method embodiments and achieve the same technical effect. To avoid repetition, it will not be described again here.

[0107] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0108] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they 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 this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0109] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for analyzing liability determination in customer complaints, characterized in that, include: Based on the complaint content of the target customer and the preset retrieval enhancement model, a preset knowledge base is queried to obtain a set of candidate evidence related to the complaint content. The preset knowledge base stores historical business data of different customers. Based on the cross-attention model, the complaint content and each candidate piece of evidence in the candidate evidence set are semantically aligned to obtain the multimodal fusion feature vectors of the complaint content corresponding to each candidate piece of evidence. Based on the multimodal fusion feature vectors of each candidate piece of evidence and a preset cause-of-responsibility classifier, the probability of each type of cause of responsibility corresponding to the complaint content is predicted. The preset cause-of-responsibility classifier includes a classification space of multiple predefined types of cause of responsibility.

2. The method according to claim 1, characterized in that, The multimodal fusion feature vectors of each candidate piece of evidence corresponding to the complaint content are obtained, including: The text of the complaint is input into a pre-trained first text encoder for semantic encoding to obtain the semantic feature vector corresponding to the complaint content; Each candidate piece of evidence in the candidate evidence set is input into a pre-trained second text encoder for semantic encoding to obtain the semantic feature vector corresponding to each candidate piece of evidence. The semantic feature vector corresponding to the complaint content and the semantic feature vector corresponding to each candidate piece of evidence in the candidate evidence set are input into the cross-attention model for semantic alignment, so as to output the multimodal fusion feature vector of each candidate piece of evidence corresponding to the complaint content.

3. The method according to claim 2, characterized in that, After predicting the probability of each type of cause of liability corresponding to the content of the complaint, it also includes: Input the semantic feature vectors corresponding to each candidate piece of evidence in the candidate evidence set into the preset evidence scoring model to obtain the evidence score set of the degree of support of the corresponding candidate piece of evidence for the target responsibility cause of the type with the highest probability. Construct an evidence chain corresponding to the cause of the target responsibility based on the evidence score set; Based on the stated reasons for the target responsibility and the corresponding chain of evidence, a responsibility determination analysis report corresponding to the content of the complaint is generated.

4. The method according to claim 3, characterized in that, Before inputting the semantic feature vectors corresponding to each candidate piece of evidence in the candidate evidence set into the preset evidence scoring model, the method further includes the steps of training a responsibility classification model containing the first text encoder, the second text encoder, the cross-attention model, and the preset responsibility cause classifier, and training the preset evidence scoring model, specifically including: An objective function is constructed based on the classification loss corresponding to the responsibility classification model, which includes a first text encoder, a second text encoder, a cross-attention model, and a preset responsibility cause classifier, the scoring loss corresponding to the preset evidence scoring model, and the regularization loss of the trainable parameters corresponding to the responsibility classification model and the preset evidence scoring model. With the overall loss of minimizing the classification loss, scoring loss, and regularization loss as the optimization objective, the trainable parameters of the responsibility classification model and the preset evidence scoring model are adjusted until the trainable parameters converge or reach the preset number of iterations, thereby training the final responsibility classification model and the preset evidence scoring model.

5. The method according to claim 3, characterized in that, Based on the set of evidence scores, construct the evidence chain corresponding to the cause of the target responsibility, including: Compare the score of each piece of evidence in the evidence score set with a preset score threshold; Identify at least one key piece of evidence whose score exceeds the preset score threshold; The at least one key piece of evidence is identified as the chain of evidence corresponding to the cause of the target's responsibility.

6. The method according to claim 5, characterized in that, After obtaining the set of evidence scores for the degree of support of each semantic feature vector for the target responsibility cause of the type with the highest probability, it also includes: Calculate the average evidence score corresponding to the at least one key piece of evidence; If the maximum probability of the preset responsibility cause classifier corresponding to the target responsibility cause is greater than the first confidence threshold and the average evidence score is greater than the preset score threshold, then the target responsibility cause predicted by the preset responsibility cause classifier is adopted. If the maximum value of the preset responsibility cause classifier corresponding to the target responsibility cause is less than the second confidence threshold, then the responsibility cause corresponding to the complaint content is determined to be the customer's own cause, and the second confidence level is less than the first confidence threshold; If the maximum value of the preset responsibility cause classifier corresponding to the target responsibility cause is not greater than the first confidence threshold and not less than the second confidence threshold, then a manual review process is triggered to process the complaint content.

7. The method according to claim 3, characterized in that, Also includes: Based on the aforementioned reasons for the target responsibility, a related query can be performed to determine the corresponding responsibility attribution and responsible department; The contract terms of the target customer, historical cases similar to the complaints of the target customer, and / or rules of liability determination are consulted to determine the basis for determining the cause of liability of the target customer. Add the content of the complaint, the attribution of responsibility, the responsible department, and the basis for the judgment to the liability determination analysis report.

8. A customer complaint liability determination analysis device, characterized in that, It includes a processor and a memory, the memory storing a program or instructions that can run on the processor, the program or instructions being executed by the processor to implement the steps of the method as described in any one of claims 1-7.

9. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, The computer program product includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform the steps of the method as described in any one of claims 1-7.

Images