Anomaly processing method and device based on multi-dimensional classifier chain, equipment and medium

Abstract

This invention relates to the field of intelligent decision-making technology and can be applied to business scenarios such as fintech and healthcare. It discloses an anomaly handling method, apparatus, device, and medium based on a multi-dimensional classifier chain, comprising: acquiring a voice data stream and generating a text data sequence with role identifiers, while simultaneously extracting voice acoustic features; performing semantic parsing on the text data sequence and determining dialogue structure features; fusing three types of features to form a unified input feature vector; performing chain-like prediction based on label dependencies using a multi-dimensional classifier chain model and outputting multi-dimensional classification prediction results; determining the anomaly level based on the prediction results and matching the target strategy to execute the corresponding action. This invention improves the accuracy of multi-dimensional prediction by fusing multi-source features and utilizing a classifier chain model to model label dependencies, thereby more reliably identifying anomaly levels and triggering matching strategies, achieving more accurate customer service analysis and recovery processing.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision-making technology, and in particular to an anomaly handling method, apparatus, device, and medium based on a multidimensional classifier chain. Background Technology

[0002] In the fintech business, customer service centers need to process a large amount of voice interaction information from customers, which often contains important signals such as emotional fluctuations, inquiry intentions, potential complaint behaviors, and transaction risk tendencies. Traditional quality management methods usually rely on manual sampling or coarse-grained analysis methods based on keywords. Due to the limited sampling rate, the overall coverage is extremely low, resulting in a large number of potential service risks going unidentified in a timely manner. At the semantic understanding level, existing methods struggle to handle complex language structures with ambiguous expressions, implicit emotions, or cross-sentence connections, making it difficult to accurately grasp the customer's true intentions. In terms of tag prediction, multi-dimensional tags such as emotion, intention, and complaint risk are often processed in isolation, lacking inter-dimensional dependency modeling, making the prediction results unable to reflect the inherent logical relationships between multi-dimensional information. Furthermore, in terms of real-time response, due to the lack of automation, customer service centers often cannot determine the level of abnormality in real time based on the overall risk situation of the call content, making it difficult to take appropriate handling strategies in a timely manner.

[0003] In the healthcare sector, patient service centers also face the challenge of handling massive volumes of voice interactions and insufficient manual review when dealing with inquiries, complaints, emotional fluctuations, and potential risk events. Traditional quality control methods rely heavily on manual listening or template-based rule detection, which struggles to identify complex information such as escalating emotions, heightened anxieties, and implicit semantics in medication risk consultations. Regarding dialogue structure analysis, healthcare services often involve multiple rounds of questioning, explanation, and confirmation. Existing technologies lack the ability to effectively process structural features such as the number of rounds of communication between different roles, interruptions, and silences, making it difficult to judge potential risk trends from dialogue interaction patterns. Furthermore, there are clear correlations between dimensions such as medical complaints, anxiety about follow-up visits, and service dissatisfaction, but traditional prediction methods typically treat these dimensions independently, ignoring the logical dependencies between them, leading to inaccurate overall risk assessments. Simultaneously, the lack of automated strategy triggering mechanisms based on multi-dimensional prediction results makes it difficult for service systems to intervene quickly in inquiries involving emotional abnormalities or potential complaints. Summary of the Invention

[0004] The main objective of this invention is to provide an anomaly handling method, apparatus, device, and storage medium based on a multidimensional classifier chain, aiming to solve the technical problems of existing technologies being unable to perform chain-like modeling of the conditional dependencies between multidimensional tags in service interactions, unable to combine fusion features to achieve reliable dynamic prediction, and difficult to identify anomaly levels and trigger corresponding strategies in a timely manner.

[0005] To achieve the above objectives, the present invention provides an anomaly handling method based on a multidimensional classifier chain, comprising: Acquire the voice data stream during the service interaction process, perform role channel differentiation and speech-to-text processing on the voice data stream, generate a text data sequence containing role identifiers, and extract the speech acoustic features of the voice data stream. Semantic parsing is performed on the text data sequence to generate a semantic representation vector, and the dialogue structure features in the service interaction process are determined based on the role identifier; The speech acoustic features, the semantic representation vector, and the dialogue structure features are fused together to obtain a unified input feature vector. The pre-built multidimensional classifier chain model is invoked. The multidimensional classifier chain model is built based on a preset multidimensional label set. The construction process includes determining the degree of conditional dependence between the labels of each dimension in the multidimensional label set to establish a ranking chain for label prediction, and building a corresponding binary classification model set for each dimension label based on the label decomposition strategy. The unified input feature vector is input into the multidimensional classifier chain model, and the binary classification model set is called sequentially according to the order of the sorting chain for prediction. During the prediction process, the confidence index is used to control the transmission of the prediction results of the preceding node to the following node, and the multidimensional classification prediction result is output. Anomaly level is determined based on the multidimensional classification prediction results, a corresponding target strategy is matched based on the anomaly level, and a corresponding action is executed according to the target strategy.

[0006] Furthermore, to achieve the above objectives, the present invention provides an anomaly handling device based on a multidimensional classifier chain, comprising: The voice processing module is used to acquire the voice data stream during the service interaction process, perform role channel differentiation and voice transcription processing on the voice data stream, generate a text data sequence containing role identifiers, and extract the voice acoustic features of the voice data stream. The semantic analysis module is used to perform semantic parsing on the text data sequence to generate semantic representation vectors, and to determine the dialogue structure features in the service interaction process based on the role identifier; The feature fusion module is used to fuse the speech acoustic features, the semantic representation vector and the dialogue structure features to obtain a unified input feature vector; The classifier calling module is used to call a pre-built multidimensional classifier chain model. The multidimensional classifier chain model is built based on a preset multidimensional label set. The construction process includes determining the degree of conditional dependence between the labels of each dimension in the multidimensional label set to establish a sorting chain for label prediction, and building a corresponding binary classification model set for each dimension label based on the label decomposition strategy. The classification prediction module is used to input the unified input feature vector into the multidimensional classifier chain model, call the binary classification model set sequentially according to the order of the sorting chain for prediction, use the confidence index to control the transmission of the prediction results of the preceding node to the subsequent node during the prediction process, and output the multidimensional classification prediction result. The strategy execution module is used to determine the anomaly level based on the multidimensional classification prediction results, match the corresponding target strategy based on the anomaly level, and execute the corresponding action according to the target strategy.

[0007] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and an exception handling program based on a multidimensional classifier chain stored in the memory and executable on the processor, wherein when the exception handling program based on the multidimensional classifier chain is executed by the processor, it implements the steps of the exception handling method based on the multidimensional classifier chain as described above.

[0008] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing an exception handling program based on a multidimensional classifier chain, wherein the exception handling program based on the multidimensional classifier chain, when executed by a processor, implements the steps of the exception handling method based on the multidimensional classifier chain as described above.

[0009] Beneficial Effects: This invention relates to the field of intelligent decision-making technology and can be applied to business scenarios such as fintech and healthcare. It discloses an anomaly handling method, apparatus, device, and medium based on a multi-dimensional classifier chain, including: acquiring a voice data stream and generating a text data sequence with role identifiers, while simultaneously extracting voice acoustic features; performing semantic parsing on the text data sequence and determining dialogue structure features; fusing three types of features to form a unified input feature vector; performing chain-like prediction based on label dependencies using a multi-dimensional classifier chain model and outputting multi-dimensional classification prediction results; determining the anomaly level based on the prediction results and matching the target strategy to execute corresponding actions. This invention improves the accuracy of multi-dimensional prediction by fusing multi-source features and using a classifier chain model to model label dependencies, thereby more reliably identifying anomaly levels and triggering matching strategies, achieving more accurate customer service analysis and recovery processing. Attached Figure Description

[0010] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of an application environment for an anomaly handling method based on a multidimensional classifier chain according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating an embodiment of the anomaly handling method based on a multidimensional classifier chain according to the present invention. Figure 3This is a schematic diagram of the functional modules of a preferred embodiment of the anomaly handling device based on a multidimensional classifier chain of the present invention; Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0011] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0012] The anomaly handling method based on a multidimensional classifier chain provided in this invention can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can acquire the voice data stream from the client and generate a text data sequence with role identifiers, while extracting the voice acoustic features; it performs semantic parsing on the text data sequence and determines the dialogue structure features; it fuses the three types of features to form a unified input feature vector; it performs chain prediction based on label dependencies using a multi-dimensional classifier chain model and outputs multi-dimensional classification prediction results; based on the prediction results, it determines the anomaly level and matches the target strategy to execute the corresponding action. This invention improves the accuracy of multi-dimensional prediction by fusing multi-source features and using a classifier chain model to model label dependencies, thereby more reliably identifying anomaly levels and triggering matching strategies, achieving more accurate customer service analysis and recovery processing. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The following detailed description of specific embodiments further illustrates this invention.

[0013] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the anomaly handling method based on a multidimensional classifier chain provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0014] like Figure 2 As shown, the anomaly handling method based on a multidimensional classifier chain proposed in this invention includes the following steps: S10, acquire the voice data stream during the service interaction process, perform role channel differentiation and speech transcription processing on the voice data stream, generate a text data sequence containing role identifiers, and extract the speech acoustic features of the voice data stream. In this embodiment, acquiring the voice data stream during service interaction is a process of collecting continuous sound signals, typically generated by customers and service personnel in the call system. The voice data stream contains information such as time-domain waveforms, energy distribution, silent segments, and background noise for subsequent analysis. Role-channel differentiation of the voice data stream means separating the overall audio into independent trajectories corresponding to different speakers. This differentiation can rely on the physical channels of the call system or on voiceprint features to identify different sound sources, ensuring that text and behavioral cues are accurately mapped to specific participants. Speech-to-text processing of the voice data stream converts continuous audio segments into a text sequence. Acoustic models capture the acoustic features of the sound, and language models infer the linguistic content composed of words, generating corresponding timestamps to ensure the text sequence is temporally consistent with the audio.

[0015] Generating a text data sequence containing role identifiers involves arranging the transcribed text content chronologically and associating it with previously obtained role information, ensuring each text segment has a clear speaker attribute. At this stage, the text sequence forms a structured representation, providing a linguistic foundation for subsequent semantic and behavioral judgments. Extracting speech acoustic features involves quantifying the rhythm, volume, frequency, and fundamental frequency variations of the original audio signal. For example, volume fluctuations reflect emotional intensity, speech rate changes reflect communication status, and fundamental frequency variations reflect intonation characteristics. These parameters are calculated using a short-time analysis window, allowing for continuous output without interrupting the original speech structure, thus obtaining a complete description of the behavioral signal.

[0016] Voice data streams can be acquired through call recording systems, real-time VoIP audio packets, or device microphones, and can be adapted to different sampling rates, noise environments, or echo conditions. Role channel differentiation can be achieved through preset dual-channel configurations, or through a combination of voiceprint clustering and speech activity detection to obtain finer-grained speaker segmentation. In speech-to-text processing, an end-to-end model or a concatenated mode combining an acoustic model and a language model can be selected based on business characteristics; in scenarios with a large amount of technical terminology, a domain-specific vocabulary can be introduced to improve recognition quality.

[0017] Text data sequences can be generated by recombining them based on role segmentation points while maintaining the original chronological order, or by integrating discontinuous dialogues using a uniform time-slice approach to adapt to different call rhythms. Acoustic feature extraction can be performed using MFCC, energy envelope, fundamental frequency curve, or a combination of multiple features, depending on device conditions. The feature analysis window size, overlap ratio, and number of filters can all be adjusted based on response speed and feature stability.

[0018] This embodiment synchronously acquires and forms structured outputs of audio content, speaker identity, and vocal behavior signals, ensuring that each text segment has a clear source and complete acoustic background. This provides a more accurate and stable input basis for subsequent semantic processing and anomaly detection, thereby reducing analytical biases caused by information loss and role confusion.

[0019] S20, perform semantic parsing on the text data sequence to generate a semantic representation vector, and determine the dialogue structure features in the service interaction process based on the role identifier; In this embodiment, semantic parsing of the text data sequence is a process of analyzing the language structure and extracting semantic information from text content arranged chronologically. The text data sequence originates from the previous transcription step, where each segment of text carries role identification and chronological order information to reconstruct semantic cues in the communication process. The goal of semantic parsing is to identify expressive intent, emotional cues, key information points, and other content from text fragments, forming semantic representation vectors that can be computed and used by the model.

[0020] Semantic parsing typically involves operations such as word segmentation, part-of-speech tagging, and contextual modeling. By converting word sequences in text into representations in a continuous vector space, semantic representation vectors can reflect the dependencies between words, the continuous logic between sentences, and the relationships across sentences. Semantic representation vectors can capture abstract expressions such as changes in customer needs, service personnel response methods, and potential conflict points, giving language content computational properties and facilitating subsequent model processing.

[0021] Determining the dialogue structure features in service interactions based on role identification refers to quantifying the behavioral patterns of the communication process using the speaker identities and temporal sequence information marked in the text. Dialogue structure features typically include changes in turn order, the speaking rhythm of both parties, the frequency of interruptions, and the distribution of silent periods. Turn order can be determined by changes in roles within consecutive speaking segments; interruptions can be detected by overlapping timestamps or rapid switching of voice activities; and the proportion of silent periods can be estimated by the intervals between text segments and audio pauses. These features reflect the dynamic structure of the communication process, providing an analytical basis for assessing the smoothness of communication, changes in the attitudes of both parties, and potential points of conflict.

[0022] The semantic representation vectors and dialogue structure features mentioned above are logically complementary: the semantic representation vectors describe the language content itself, while the dialogue structure features describe the communication behavior. Together, they constitute a deep expression of language information in the service interaction process.

[0023] Semantic parsing can be accomplished using a deep neural network-based encoding model. Text is input into the encoding network sentence by sentence or segment by segment to generate corresponding vector representations. When it is necessary to identify specialized terms or industry expressions, business vocabulary can be added to the vocabulary to enhance representation capabilities. For scenarios with rapidly changing semantics, a context window approach can be used to enable the model to capture semantic relationships across sentence dependencies.

[0024] The determination of dialogue structure features can be achieved through a combination of timestamp differences, role change statistics, and voice activity detection. Speaking turns can be obtained by recording role changes between adjacent text segments; interruptions can be identified by analyzing whether role switching occurs during the other party's speech; the proportion of silent periods can be determined by analyzing the time intervals between text segments, identifying pauses based on preset thresholds, and accumulating the ratio of silent time to the overall interaction duration. In noise-sensitive environments, smoothing techniques can be used to optimize the stability of silent interval identification, thereby ensuring the reliability of dialogue structure features.

[0025] This embodiment constructs semantic representation vectors at the text level and simultaneously generates dialogue structure features, enabling language content and communication behavior to be expressed in a unified form. This provides a highly consistent input basis for subsequent analysis and enhances the ability to identify ambiguous expressions, emotional transition signals, and abnormal communication patterns.

[0026] S30, the speech acoustic features, the semantic representation vector and the dialogue structure features are fused to obtain a unified input feature vector; In this embodiment, the speech acoustic features, semantic representation vectors, and dialogue structure features originate from different information sources, reflecting sound performance, text semantics, and communication behavior patterns, respectively. The fusion of these three types of features involves a series of transformations and combinations to unify them into the same vector space, enabling them to participate in subsequent judgments and calculations in a consistent manner.

[0027] Speech acoustic features typically represent acoustic variations such as intonation, speech rate, and volume fluctuations using multidimensional numerical values. These data differ from semantic representation vectors in scale and statistical distribution. Semantic representation vectors, expressed in a high-dimensional vector space, convey the intent, emotion, or logical relationships inherent in the text, exhibiting a higher degree of abstraction compared to acoustic features. In contrast, dialogue structure features use limited dimensions to characterize dynamic communication attributes such as turn order, the proportion of silent intervals, and interruptions, differing in granularity from both semantic and acoustic features.

[0028] To eliminate modal differences among the three types of features and improve overall expressiveness, they can be mapped to a compatible vector space. Linear transformations are used to adjust the scale and dimension of different feature vectors, enabling subsequent fusion stages to operate within the same dimensional structure. Fusion processing requires more than simple concatenation; it necessitates the introduction of weight adjustment mechanisms to ensure differentiated contributions from different features across various interaction scenarios. Gating structures dynamically adjust the contribution levels of speech, semantic, and structural information based on the characteristics of the input content, resulting in a more expressive weighted vector. The weighted fusion result is then subjected to a nonlinear transformation to ensure separability of feature relationships in space and enhance the ability to capture complex patterns, thereby obtaining a unified input feature vector.

[0029] The formation of a unified input feature vector logically connects the three types of information: acoustic features provide clues to emotional changes, semantic features provide linguistic connotations, and dialogue structure features provide the context of communication behavior. The fused vector formed by these three features has a more comprehensive ability to characterize the interaction content, facilitating multi-dimensional judgments by the subsequent model.

[0030] The fusion process can be accomplished through matrix projection. The speech acoustic vector, semantic representation vector, and dialogue structure vector are input into different linear transformation networks, ensuring they maintain dimensionality. The projected features are then input into a gating network. This network generates three sets of learnable coefficients through parallel computation branches, representing the adaptive weights required for each feature class. The weights are then element-wise multiplied with their respective projected vectors to obtain weighted features. Finally, the three features are summed to form the fused intermediate vector.

[0031] When stronger feature representation is required, multi-layer nonlinear networks can be used to transform the weighted fusion vector layer by layer, enabling it to exhibit sensitivity to mood shifts, semantic conflicts, or round anomalies in a high-dimensional space. If the feature distribution differs across different business scenarios, the projection dimension or the depth of the gating structure can be adjusted to achieve stable fusion quality. In resource-constrained systems, the gating network can be simplified by assigning balanced weights to the three types of features through distribution normalization, thereby reducing the computational cost of the fusion process.

[0032] This embodiment transforms, weights, and nonlinearly combines three types of information to enable acoustic changes, semantic content, and interactive behavior to be expressed collaboratively in a unified vector space. This enhances the ability to characterize complex communication patterns, reduces the one-sidedness caused by single features, and provides a more stable and sensitive input basis for subsequent judgments.

[0033] S40, Invoke the pre-built multidimensional classifier chain model. The multidimensional classifier chain model is built based on a preset multidimensional label set. The construction process includes determining the degree of conditional dependency between the labels of each dimension in the multidimensional label set to establish the ranking chain of label prediction, and building a corresponding binary classification model set for each dimension label based on the label decomposition strategy. In this embodiment, calling the pre-built multidimensional classifier chain model requires a structured learning system driven by a multidimensional label set. In order for the model to form ordered association judgments across multiple label dimensions, it is necessary to systematically model the relationships between labels during the construction phase and fix these relationships in an ordered chain so that subsequent label predictions can proceed in a defined order.

[0034] The multidimensional label set originates from the definitions of different attributes in business scenarios, such as emotion, intent, behavior, and abnormal tendencies. Each dimension label has an independent meaning, but in real-world interaction scenarios, conditional dependencies often exist. For example, changes in emotion may affect the expression of intent, and behavioral deviations may be accompanied by abnormal tendencies. Therefore, it is necessary to calculate the degree of conditional dependency between different dimension labels so that the label ranking reflects the true direction of association. The degree of conditional dependency is usually constructed using entropy-related statistics, representing the magnitude of change in the distribution of one label when the value of another label changes. Using this metric, a directional dependency ranking can be formed, ensuring that the prediction task follows an order from strong dependency to weak dependency, and from basic features to complex features.

[0035] After obtaining the ranking chain, the prediction task for each dimension label needs to be broken down into finer-grained classification tasks, thus forming a set of binary classification models. The label decomposition strategy maps labels that might originally contain multiple categories into multiple binary classification tasks, allowing each task to focus on only one splitting dimension of the label. For example, if a label has multiple possible states, multiple binary classifiers can be built using a one-to-one split approach, with each classifier determining whether a label belongs to a particular state. The model set built through binary classification enables a more stable learning process and provides clearer output boundaries in chained inference.

[0036] Calling a pre-built multi-dimensional classifier chain model means directly using the already trained sorting chain and set of binary classification models at the current input stage. When called, the model accesses the corresponding dimension's binary classification model set sequentially according to the order of the sorting chain, ensuring the prediction process follows a fixed label dependency flow. The calling action itself relies on the model structure formed in the earlier construction stage, which achieves collaborative operation between multi-dimensional tasks through label dependency modeling and decomposition strategies.

[0037] In the construction phase, joint statistics can be performed on all dimensions of the labels in the training dataset, and the dependency between each pair of labels can be calculated using conditional entropy. After the dependency matrix is ​​processed by a ranking algorithm or search program, a label prediction chain is formed. The ranking method can strictly rely on the conditional entropy from high to low, or the sequence can be optimized through evolutionary search to make the prediction chain outperform the sequence arranged simply according to dependency strength.

[0038] For label decomposition, a one-to-one splitting approach can be used to map the label space into several mutually exclusive binary classification tasks, and an independent classifier model can be trained for each task. The classifier structure can be a linear model, a deep model, or a tree model, as long as it can stably distinguish the split labels. Multiple binary classification models are organized along the label dimension so that they can be called in the order of the sorting chain.

[0039] During model building, various variations can be configured. For example, the calculation method for conditional dependencies can be changed, using mutual information or symmetric uncertainty as the source of dependencies. The label splitting strategy can also be altered, adjusting the structure of the binary classification set through hierarchical splitting or data-driven splitting. In some business scenarios, the number of model sets may be large; in such cases, sharing low-level feature networks can reduce the overall computational load.

[0040] This embodiment models the dependencies between labels and constructs a sorting chain, enabling the prediction task to proceed in a reasonable order during execution and avoiding interference between predictions of labels from different dimensions. Through label decomposition and the construction of a binary classification model set, the multidimensional prediction task achieves higher stability and interpretability, improving overall judgment ability and providing a structured foundation for subsequent multidimensional judgment steps.

[0041] S50, the unified input feature vector is input into the multi-dimensional classifier chain model, and the binary classification model set is called sequentially according to the order of the sorting chain for prediction. During the prediction process, the confidence index is used to control the transmission of the prediction results of the preceding node to the subsequent node, and the multi-dimensional classification prediction result is output. In this embodiment, the unified input feature vector is fed into the multidimensional classifier chain model, serving as the entry point for initiating the prediction process during multidimensional label inference. The unified input feature vector, having undergone preprocessing, has already incorporated acoustic features, semantic information, and structural information, thus providing a consistent representational basis for each label dimension. When feeding this vector into the model, it is necessary to ensure that its shape remains consistent with that of the pre-training stage so that the parameters within the model can correctly respond to the input signal.

[0042] Calling the set of binary classification models sequentially according to the order of the sorting chain is a key operational mode of the chained prediction structure. The sorting chain records the dependency order between different dimension labels. Each label dimension is built upon the prediction result of the previous dimension, so this dependency order must be strictly followed to ensure causal consistency in the multidimensional inference process. Each dimension label may correspond to multiple binary classification models, which are organized in the form of a set, allowing the models to evaluate different states in parallel and obtain the most certain label output.

[0043] In the prediction process, using a confidence metric to control the propagation of prediction results from preceding nodes to subsequent nodes is a key mechanism in chain structures to avoid mispropagation. The confidence metric stems from the classification model's trustworthiness of the current predicted label and can reflect probability output, boundary distance, or classification strength. Only when the confidence level meets the propagation requirement will the predicted label of the preceding node be encoded and appended to the input feature space to influence the judgment of subsequent nodes. If the confidence level is insufficient, the current input features remain unchanged, allowing subsequent nodes to make predictions based on more stable fundamental information, thereby reducing cascading bias caused by mispropagation.

[0044] Outputting multidimensional classification prediction results means that each dimension in the entire chain is predicted, and all labels are integrated into a complete classification vector in a fixed order.

[0045] In practical operation, a unified input feature vector can be input into the initial entry node of the model, allowing the model to start prediction from the first dimension of the sorted chain. After the prediction calculation for each dimension is completed, the label probability distribution for that dimension is obtained, and the confidence index is extracted from it. If the confidence score meets the propagation condition, the label can be encoded into a vector form and concatenated into the current input vector, and then the expanded features are input into the binary classification model set for the next dimension. If the confidence score is insufficient, no concatenation is performed, and the original features are directly fed into the next dimension.

[0046] The calculation of confidence can take many forms. For example, maximum class probability can be used, cross-model consistency can be used as the source of confidence, or a combination of internal boundary distance metrics within the model can be used to improve stability. The label encoding method can also be adjusted, for example, using fixed-length dense vectors, sparse vectors, or embedding vectors to adapt to different model structures. During the input vector expansion process, linear mapping can also be used to compress the dimension of newly added labels, keeping the overall input size within a computationally manageable range.

[0047] When executing a binary classification model set sequentially, weighted ensembles can be applied to the models within the set to make the output more robust. In some scenarios, to reduce computational burden, all computations of the model set can be skipped for dimensions with insufficient confidence, and only the coarsest-grained model can be used for approximate judgment. The ranking chain can also be retrained and generated according to business differences to maintain optimal label dependencies in different scenarios.

[0048] This embodiment calls the binary classification model set sequentially according to the sorted chain and introduces confidence control, making the transmission of information between label dimensions more stable and reducing the risk of the cumulative spread of erroneous labels in the chain. By dynamically deciding whether to pass previous results, the prediction process can adapt to different input qualities, improving the overall reliability and accuracy of multi-dimensional label inference.

[0049] S60, determine the anomaly level based on the multidimensional classification prediction result, match the corresponding target strategy based on the anomaly level, and execute the corresponding action according to the target strategy.

[0050] In this embodiment, determining the anomaly level based on the multidimensional classification prediction results requires structured parsing of the labels in each dimension of the prediction results and identifying label features that reflect service anomaly tendencies. The multidimensional classification prediction results consist of multiple labels, each derived from judgments in different dimensions, such as mood changes, intent shifts, or agent behavior. These labels are used to form a set of indicators that can be used for anomaly identification. The determination of the anomaly level depends on the combination relationships between the labels. Typically, mapping rules or scoring mechanisms are used to transform label combinations into high, medium, or low anomaly levels, giving the anomaly state a quantifiable expression.

[0051] Based on the anomaly level matching target strategy, it is necessary to construct a set of strategies for different anomaly levels, establishing a correspondence between levels and strategies. Each strategy contains a set of actions that can be executed by the system to take appropriate handling measures in different anomaly scenarios. The strategy set can include different categories of actions such as external processing, information sending, and internal process triggering. Through level mapping, the system can immediately select the action path after anomaly identification.

[0052] Executing the corresponding action based on the target strategy is the triggering step in the entire process. Executing an action requires retrieving the specific operation item corresponding to the target strategy from the strategy set, converting it into an executable instruction, and implementing it within the established business environment. For example, triggering an external connection process requires writing the customer identifier into the processing queue; sending a message requires constructing the message content and scheduling the sending module; generating a work order requires submitting the record to the internal processing system. Executing actions also requires ensuring that the completion status of the action is traceable for subsequent service adjustments or quality monitoring.

[0053] In actual operation, the multidimensional classification prediction results can be read first, and the key labels can be converted into input parameters and sent to the anomaly level parsing component. The parsing component can calculate the anomaly level based on rule mapping, vector distance, or weighted scoring, and output one of the three levels. After the level is determined, the level signal is passed to the policy matching component, which retrieves the corresponding policy and converts the policy content into an executable structure.

[0054] The action execution module can call different business interfaces based on the policy content. For example, when the policy includes external processing actions, the customer identifier, processing scenario, and triggering reason can be written to the external queue; when the policy includes information sending actions, the sending content can be encapsulated into a standard message structure and the message distribution module can be scheduled; when the policy includes internal process actions, the work order creation interface can be called and relevant fields can be recorded. To adapt to different scenarios, an execution path selection mechanism can also be added, enabling the system to select a more suitable action execution method based on the current load, business period, or user characteristics.

[0055] This embodiment matches target strategies and executes actions based on anomaly levels, creating a tight linkage between anomaly identification and response, and achieving seamless connection between prediction results and subsequent processing flows; at the same time, the strategic execution of actions makes the processing path more adaptable and scalable.

[0056] In one embodiment, step S10 includes: S101, Receive raw audio data during service interaction as a voice data stream; S102, the voice data stream is differentiated by role channel to obtain independent voice channels corresponding to different speakers; S103, Perform speech-to-text processing on each independent speech channel to generate the initial text corresponding to the independent speech channel and the timestamp information of each sentence of the initial text; S104, Assign a role identifier to each independent voice channel and associate the role identifier with the initial text transcribed from the independent voice channel; S105, based on the timestamp information, the initial text assigned the role identifier is merged in a time sequence to form a text data sequence containing the role identifier; S106 extracts speech acoustic features from each independent speech channel, including volume variation features, speech rate variation features, and pitch fluctuation features.

[0057] In this embodiment, the voice data stream during the service interaction process can be understood as a sequence of audio signals continuously generated by the customer and service personnel during a single call. This data is typically collected by a front-end voice access platform or recording system using a unified encoding format and cached or stored in chronological order. The original audio data retains complete waveform information during the acquisition phase, without undergoing processes that disrupt the temporal structure, such as silence trimming or resampling compression, to allow for subsequent fine-grained reconstruction of speaker rotation and emotional change trends.

[0058] Role-based channel differentiation addresses situations involving multiple participants in service interactions. It divides the voice data stream into several independent voice channels using signal source information or acoustic separation mechanisms. One implementation utilizes a dual-channel or multi-channel recording architecture, mapping the client and service personnel to different physical channels, achieving channel-level separation at the audio access stage. Another implementation uses a speaker separation model for single-channel recordings, clustering and labeling continuous audio segments based on spectral characteristics, voiceprint features, and short-time energy distribution, grouping segments with the same speaker into the same independent voice channel. Regardless of the implementation, role-based channel differentiation results in several independent voice channels corresponding one-to-one with different speakers, providing independent signal sources for subsequent role identification allocation and behavior analysis.

[0059] Independent speech channels undergo speech-to-text processing, transforming continuous speech waveforms into discrete initial text sequences. This processing employs a recognition engine combining acoustic and language modeling, performing feature extraction, acoustic unit recognition, and language decoding for each independent speech channel in chronological order, converting the audio signal into word-level or sentence-level text output. During transcription, to ensure alignment between the text and audio on the timeline, the recognition results are segmented into sentences, and a timestamp is recorded for each initial sentence. The timestamp information includes at least the start and end times of the sentence within the current independent speech channel. In this way, the initial text not only carries semantic content but also retains a precise correspondence with the original speech signal.

[0060] Role identifiers are used to mark the functional identity of each independent voice channel in service interactions, such as customer, frontline customer service, quality inspector, etc. In one implementation, role identifiers can be associated with agent accounts, extension numbers, or access line configurations in the business system, directly determining whether the channel is a customer-side channel or a service personnel-side channel based on its source. In another implementation, role identifiers can be bound to conversation access records or manually configured. Once a role identifier is assigned to an independent voice channel, it needs to be written into a structured record of each initial text transcribed from that channel, so that each text record simultaneously carries three types of fields: semantic content, timestamp information, and role identity.

[0061] When merging initial text with role identifiers based on timestamp information, text records from different independent voice channels are placed in a unified timeline and sorted from earliest to latest according to their start time, constructing a cross-role text data sequence. A stable sorting strategy can be adopted during the merging process; when sentences from different channels have the same or very similar start times, the original channel order or internal sequence number is used as the ordering basis to avoid distorting the rhythm of the actual dialogue. Each element in the time-merged text data sequence contains text content, timestamp information, and role identifiers, and the sequence as a whole reflects the complete evolution trajectory of the service interaction process in the time dimension, including phenomena such as alternating speaking, cross-interruption, and prolonged silence.

[0062] Speech acoustic feature extraction is performed on independent speech channels, aiming to capture three complementary types of acoustic information: volume variation features, speech rate variation features, and pitch fluctuation features. Volume variation features are obtained by calculating energy or sound pressure level indices from short-time audio frames, constructing a time series from the results of consecutive frames, and extracting statistical quantities such as average value, amplitude of change, and rise and fall slopes from this series to characterize dynamic changes in service interactions, such as emotional excitement, low tone, and sudden increases in volume. Speech rate variation features are obtained by calculating the number of words, phrases, or sentences per unit time based on the transcribed initial text and timestamp information. These indicators are then aggregated by sliding a window throughout the call to obtain the speech rate variation curves in different time segments, and quantitative descriptions such as stability and acceleration or deceleration trends are extracted. Pitch fluctuation features are obtained by fundamental frequency detection and formant analysis to estimate pitch curves and formant positions in the short-time spectrum. Smoothing and differencing operations are performed on the curves to obtain morphological indicators such as pitch rise, fall, and sustained high pitches, and the variation patterns of pitch during the call are characterized by statistical distribution, extreme value positions, and fluctuation frequencies. Volume variation features, speech rate variation features, and pitch fluctuation features can be aligned along a unified time axis to form a multidimensional vector sequence. Then, through pooling, dimensionality reduction, or aggregation, a speech acoustic feature representation can be formed for subsequent analysis, making each independent speech channel comparable and quantifiable at the acoustic level.

[0063] In terms of time, text data sequences and speech acoustic features can be aligned by sharing timestamp information, synchronizing each text segment and its corresponding acoustic performance. For example, text sentences can be associated with adjacent acoustic frame sets through the intersection of time periods to generate a joint record carrying role identifiers, text content, and acoustic features.

[0064] This embodiment achieves unified organization of textual and speech acoustic information on the same timeline by performing role channel differentiation, speech transcription processing, role identification association, temporal merging, and extraction of multiple types of acoustic features on the voice data stream during service interaction. This ensures that each speech maintains a precise correspondence in the three dimensions of content, time, and role, thereby providing a complete, quantifiable, and clearly structured basic data representation for subsequent interaction quality assessment and anomaly level identification based on multi-dimensional input.

[0065] In one embodiment, step S20 above includes: S201, Perform word segmentation on the text data sequence to obtain the word segmentation result; S202, Determine the keyword weight vector based on the word segmentation results; S203, The text data sequence is converted into a deep semantic vector using a semantic embedding model; S204, Generate a semantic representation vector based on the keyword weight vector and the deep semantic vector; S205, based on the temporal sequence information of sentences in the text data sequence and the role identifier, divide the speaking rounds; S206, Based on the time sequence information and the role identifier, count the number of interruptions during the service interaction process; S207, Based on the time sequence information, determine the proportion of silent time periods in the service interaction process; S208, the number of speaking turns, the number of interruptions, and the percentage of silence time are used as dialogue structure features.

[0066] In this embodiment, the text data sequence is a set of sentences or phrases arranged in chronological order. Each record contains at least text content, time information, and a role identifier generated in the previous stage. The time information may include the start and end times of the sentence in the call, and the role identifier is used to distinguish between the customer and the service personnel, and can also distinguish between different types of service personnel if necessary. Semantic parsing performs content analysis and structural encoding on the text data sequence, transforming the interactive content, originally in natural language form, into a vector representation suitable for subsequent computational processing.

[0067] Word segmentation processes break down each sentence in a text data sequence into words or sub-words. During segmentation, multi-granular dictionaries, statistical models, and contextual judgments can be combined to accurately segment proper nouns, business terms, and sentiment words. Simultaneously, the position index of each word within the sentence is preserved, ensuring that subsequent calculations can trace back to the original sentence structure and role information. The segmentation results can be organized into several word sequences grouped by sentence number and role identifier. Each word entry includes its form, position, and optional part-of-speech tag.

[0068] Keyword weight vectors quantify the importance of different words based on word segmentation results. In implementation, a base weight for each word can be calculated using a combination of word frequency and inverse document frequency. This weight is then combined with sentiment vocabularies, complaint-related vocabularies, and business domain vocabularies to add weight gains to words strongly correlated with emotional fluctuations, risk expression, and service attitude. Alternatively, methods such as attention distribution or gradient attribution can be used to extract information reflecting contribution from the internal calculations of the pre-trained semantic model. These values ​​are then combined with the base weights to form keyword weight vectors whose dimensions correspond one-to-one with the vocabularies, highlighting the language units that truly influence judgment in the current call scenario.

[0069] Semantic embedding models vectorize text data sequences, mapping discrete word sequences into deep semantic vectors in a continuous space. In implementation, a sequence modeling structure can be used to encode the word segmentation results of each sentence, obtaining a sequence of word vectors that includes both word meaning and contextual dependencies. Then, through pooling, aggregation, or a dedicated sentence vector generation unit, a fixed-dimensional deep semantic vector is generated for each sentence. At the dialogue level, sentence vectors from the same conversation can be input into the sequence modeling structure again in chronological order to generate a high-dimensional representation covering the entire interaction process, used to comprehensively characterize information such as customer needs, problem types, and service attitudes.

[0070] Semantic representation vectors are generated based on keyword weight vectors and deep semantic vectors. One implementation involves weighted summation or weighted pooling of word vectors for each sentence according to keyword weights, ensuring that words with higher weights in the keyword weight vectors contribute more to the sentence-level representation, thus constructing sentence-level semantic representation vectors. Another implementation introduces weighted gating units at the dialogue level, modulating the deep semantic vectors of different sentences with the keyword weight vectors, forming dialogue-level semantic representation vectors that distinguish between key and background information at the sentence level. Through this joint modeling approach, the semantic representation vectors not only include contextual semantic relationships but also explicitly incorporate keyword importance information, thereby enabling more sensitive encoding of implied complaints, transitional tones, and subtle refusals.

[0071] Speaking turns are divided based on the temporal sequence of sentences and role identifiers to describe the turn-taking structure among different participants in service interactions. In implementation, the text data sequence is traversed chronologically. When two consecutive records have the same role identifier and the time interval is less than a preset threshold, they are grouped into the same turn. When the role identifier changes or the time interval exceeds the threshold, a new speaking turn is initiated. The start time, end time, and the set of sentences included in each turn are recorded. In this way, a turn sequence covering the entire call process can be constructed, clearly defining the duration, content density, and role distribution of each turn, providing fundamental data for determining whether service personnel monopolize the speaking position for extended periods or whether customers repeatedly make the same requests.

[0072] Interruption counts are statistically analyzed based on temporal sequence information and role identification to detect instances where one party interrupts a conversation before the other has finished speaking. Based on pre-defined speaking rounds, the time boundaries between adjacent rounds are checked. If the start time of a subsequent round is earlier than the end time of the last sentence in the previous round, or if the interval between the start and end times of the last sentence is less than a pre-set overlap threshold, that round transition is recorded as an interruption event. More granular analysis can also be performed at the sentence level to determine whether the start time of a new sentence from one role falls within the duration window of a previous long sentence from another role, thus identifying forced interruptions. The count of these events is accumulated throughout the call to form the interruption count, a dialogue structure indicator, and can differentiate between customer interruptions and service personnel interruptions based on role.

[0073] The percentage of silent periods is calculated based on chronological information and reflects the proportion of time during which no effective speech occurs in service interactions. In practice, the start and end times of all sentences are sorted on a timeline using the call's start and end times as boundaries. The time interval between adjacent speeches is calculated, and intervals with intervals greater than a silence threshold are defined as silent periods. The sum of the durations of all silent periods is then compared to the total call duration to obtain the quantitative indicator of the percentage of silent periods. For the waiting time before connection begins or the tail end of the call before hanging up, inclusion in the statistics can be selected based on business settings to more accurately reflect the silence situation during the main phase of the service interaction. The percentage of silent periods, along with the number of speaking turns and interruptions, constitutes the dialogue structure features, describing the rhythm, pause distribution, and power allocation during the interaction. It can be stored in parallel with semantic representation vectors in a unified data structure. By associating the same dialogue identifier and time index, a computable mapping relationship is formed between semantic content and the interaction structure.

[0074] This embodiment processes text data sequences through word segmentation, keyword weight calculation, and semantic embedding to form semantic representation vectors that combine contextual information and the importance of key words. Based on role identification and time sequence information, it divides speaking turns, counts the number of interruptions, and the proportion of silent periods to establish dialogue structure features covering the content and structural layers. This allows each service interaction to have a refined vectorized description in three dimensions: semantic expression, role behavior, and rhythm distribution. As a result, it can more accurately distinguish between normal business communication and abnormal interaction patterns in subsequent analysis, providing a more reliable basic representation for identifying emotional abnormalities, communication imbalances, and potential complaint risks.

[0075] In one embodiment, step S30 above includes: S301, linearly project the speech acoustic features, the semantic representation vector and the dialogue structure features respectively to obtain projected acoustic features, projected semantic features and projected structure features; S302, input the projection acoustic features, the projection semantic features and the projection structural features into a preset gating network, and determine the adaptive weight coefficients corresponding to the projection acoustic features, the projection semantic features and the projection structural features respectively; S303, the projection acoustic features, projection semantic features and projection structural features are weighted and summed using the adaptive weighting coefficients to obtain weighted fusion features; S304, perform a nonlinear transformation on the weighted fusion features to obtain enhanced fusion features as a unified input feature vector.

[0076] In this embodiment, the speech acoustic features, semantic representation vectors, and dialogue structure features originate from the multi-channel processing results of the previous stage and are aligned in terms of time axis and session identifiers. However, the three types of data differ significantly in dimensionality, numerical range, and statistical properties. Directly concatenating them would make it difficult for the subsequent model to balance information from different sources during training and inference. To address this issue, a unified fusion processing flow is introduced, transforming the three types of features into a unified input feature vector through linear transformation, gated weighting, and nonlinear mapping. This vector represents the comprehensive state of a service interaction within the same vector space.

[0077] Speech acoustic features can include time-varying numerical sequences such as volume changes, speech rate changes, and intonation fluctuations, typically forming matrices or high-dimensional vectors on a frame-by-frame basis. The dimensionality is related to the acoustic analysis window and the feature dimension. Semantic representation vectors come from the semantic embedding module in the previous stage and are usually high-dimensional vectors with fixed dimensions, used to characterize customer needs, emotional tendencies, and service content. Dialogue structure features include aggregated indicators such as speaking turns, number of interruptions, and percentage of silent periods. These can be normalized and encoded to convert discrete counts and proportions into vector forms that can be jointly processed with other features. To achieve unified modeling in subsequent calculations, these three types of features need to be linearly projected so that their dimensions are mapped to the same target dimension, and their numerical ranges are constrained within an appropriate range.

[0078] Linear projection can be achieved through matrix multiplication. For example, a dedicated projection matrix can be configured for speech acoustic features to calculate projected acoustic features; another set of matrices or fully connected layers can be configured for semantic representation vectors to obtain projected semantic features; and a third set of projection parameters can be configured for dialogue structure features to obtain projected structure features. Each projection parameter can be trained independently, ensuring that data from different sources have the same dimensionality after mapping, but occupy different regions in the vector space, thus preserving the original information structure while facilitating unified processing. In the implementation, bias and regularization terms can be introduced to control the numerical distribution after projection, avoiding excessive amplification of a certain type of input on a numerical scale.

[0079] Gated networks are used to adaptively assign weights to projective acoustic features, projective semantic features, and projective structural features. The gated network can receive the concatenated results of the three types of projective features, or their statistical summaries, and then generate three weight coefficients through several layers of linear transformations and nonlinear activations. These three weight coefficients correspond to the projective acoustic features, projective semantic features, and projective structural features, respectively. The weight coefficients can be constrained by a normalization function to fall within a preset range. For example, a normalization function can be used to ensure that the sum of the three weights equals one, or a compression function can be used to ensure that the weights are between zero and one. The gated network can automatically adjust the contribution ratio of each feature to the final fused representation based on the differences in speech fluctuations, semantic content, and structural rhythm in different calls. For example, it can increase the weight of the acoustic part in emotionally charged calls, and increase the weight of the semantic and structural parts in complex complaint scenarios.

[0080] The weighted summation operation is performed after obtaining the adaptive weight coefficients, combining the projected acoustic features, projected semantic features, and projected structural features in a one-to-one correspondence along the vector dimensions. For each dimension, the products of the acoustic component, semantic component, and structural component with their respective weight coefficients are calculated and then summed to form a weighted fusion feature. In this structure, even if the three types of features contain noise or redundant information in some dimensions, this influence will be suppressed as long as the weights given by the gating network are low. Conversely, in dimensions that play a crucial role in the judgment result, the effective signal can be amplified through higher weights. The weighted summation process can be performed at the single-call level, or at the time-slice or speech-turn level, and then a globally weighted fusion representation is obtained through pooling or aggregation.

[0081] Nonlinear transformations are used to further reconstruct the space of the weighted fusion features, mapping the linearly combined vectors to a more expressive high-dimensional space to generate enhanced fusion features. Nonlinear transformations can employ single-layer or multi-layer fully connected networks with nonlinear activation functions, or residual or normalized structures to enhance gradient propagation stability. During computation, activation functions can suppress extreme values, highlighting discriminative feature dimensions and creating clearer distribution boundaries for different call scenarios in the enhanced fusion feature space. After this series of linear and nonlinear processing, the enhanced fusion features are used as a unified input feature vector, unified in dimension, numerical range, and information structure to a single representation format, facilitating direct integration by subsequent modules without requiring separate input channels for data from different sources.

[0082] This embodiment unifies multi-source data with large differences in dimensionality and numerical distribution into a unified input feature vector with consistent structure by performing linear projection, gated weighting, and nonlinear transformation on speech acoustic features, semantic representation vectors, and dialogue structure features. This allows for the aggregation of emotional signals, semantic content, and dialogue rhythm information without losing key information, preventing one type of input from dominating the overall expression in subsequent modeling. This improves the expression accuracy and separability of multi-source information fusion, providing a stable and discriminative input representation for the subsequent discrimination module.

[0083] In one embodiment, before invoking the pre-built multidimensional classifier chain model in step S40 above, the following is also included: S4001, Define a multi-dimensional tag set, which includes customer emotion dimension, customer intention dimension, agent behavior dimension, complaint anomaly dimension, and recovery strategy dimension; S4002, Obtain the training dataset with complete dimensional labels; S4003, Based on the training dataset, determine the conditional entropy value between any two dimension labels in the multidimensional label set as the degree of conditional dependency; S4004, Based on the conditional entropy value, run an evolutionary search strategy to determine the prediction order between dimension labels and establish a sorting chain for label prediction; S4005, For each dimension label in the sorting chain, a one-to-one decomposition strategy is adopted to train the corresponding set of binary classification models. S4006, use the validation dataset to analyze the prediction accuracy of each binary classifier in the binary classification model set, and mark the binary classifiers with prediction accuracy lower than a preset threshold as non-transfer nodes; S4007, Based on the sorting chain, the binary classification model set, and the label of the non-transitive node, a multidimensional classifier chain model is constructed.

[0084] In this embodiment, before calling the multidimensional classifier chain model, a classification structure reflecting the multidimensional label relationships of the business needs to be constructed during the offline training phase. First, a multidimensional label set is defined. This set divides the annotation results corresponding to a service interaction session into multiple semantic dimensions, each representing an independent but relevant business judgment perspective. The customer sentiment dimension can be used to classify emotional biases, such as stable, slightly fluctuating, or strongly dissatisfied; the customer intent dimension is used to characterize the customer's current focus, such as inquiries, complaints, insurance cancellation tendencies, and additional purchases; the agent behavior dimension describes the service personnel's performance in terms of script standards, compliant language, and resolution efficiency; the complaint anomaly dimension is used to mark whether the call has potential complaint risks or abnormal service situations; and the recovery strategy dimension records the intervention methods to be taken in similar situations. The multidimensional label set unifies these dimensions through a higher-level concept, thereby allowing multiple label vectors to be combined in the same label space to characterize the overall state of an interaction.

[0085] The training dataset consists of a large number of historical service interaction records. Each record contains at least the input features corresponding to the aforementioned speech processing and text parsing stages, as well as a label vector consistent with the multi-dimensional label set. During data collection, information such as call audio, quality inspection conclusions, complaint records, insurance cancellation results, and subsequent customer behavior can be jointly extracted from the recording system and business system, and this information is organized into a unified sample. During the annotation process, each interaction record can be labeled with emotion category, intent category, agent behavior level, complaint anomaly category, and corresponding recovery strategy category, forming a complete dimensional label annotation. To increase the representativeness of the label distribution, different channels, different business types, and different customer groups can be covered during sampling, so that the training dataset has sufficient coverage in the business space.

[0086] Conditional dependency is quantified using conditional entropy. For any two dimensions in a multidimensional label set, such as customer sentiment and complaint anomaly, the joint probability distribution and conditional probability distribution can be estimated based on the frequency of label combinations appearing in the training dataset. Conditional entropy measures the uncertainty of one dimension's label given the label of another dimension; a smaller value indicates a stronger dependency. By calculating the conditional entropy for any two dimension label pairs, an association matrix can be obtained, with dimensions as nodes and conditional entropy as edge weights, describing the conditional dependency structure between multidimensional labels. The degree of conditional dependency can not only be directly equated to the conditional entropy itself, but a relevance score can also be calculated based on the conditional entropy. For example, normalization or inverse transformation can be used to map a smaller conditional entropy to a higher dependency score.

[0087] Evolutionary search strategies are used to determine a prediction order that favors chained prediction among multi-dimensional labels. The prediction order can be viewed as a permutation encoding of each label dimension, with each permutation corresponding to a chained prediction order structure. Evolutionary search strategies can use a population to represent multiple candidate ordering chains, treating each chain as a chromosome, and iteratively updating it within the search space through operations such as crossover and mutation. When evaluating each ordering chain, the chained prediction order can be simulated on the training dataset. A simplified model is trained by concatenating existing features with the true or fitted labels of the preceding dimensions, calculating the prediction accuracy or joint loss function, and assigning fitness to the ordering chain based on this evaluation metric. After several generations of evolutionary search iterations, the prediction order that performs better in joint multi-dimensional label prediction can be selected and used as the ordering chain for label prediction. In this way, the label order no longer depends on empirical settings but is data-driven, derived based on the business label conditional dependency structure.

[0088] The one-to-one decomposition strategy is used to break down the multi-class determination of each dimension's label into a set of binary classification tasks, thereby improving the flexibility of prediction for each dimension. For a given dimension, if it contains multiple categories—for example, the recovery strategy dimension may contain intervention strategies of different levels or paths—a binary classifier can be built for each pair of categories. Each binary classifier is used to determine the boundary between two candidate categories. When training the binary classification model set, a unified input feature vector and the label information of the preceding dimensions can be used as input, and the category pairs of the current dimension can be used as output labels. The one-to-one decomposition strategy can improve the discriminative ability in cases of imbalanced category boundaries or complex category distributions, and it also facilitates the synthesis of the final category decision through voting or the maximum score rule during subsequent prediction.

[0089] The validation dataset is either partitioned from the training samples or constructed from independent samples to evaluate the prediction accuracy of each binary classifier in the binary classification model set. During validation, a unified input feature vector can be fed into the trained binary classifier without participating in parameter updates to predict the current class pair, and the consistency between the prediction result and the true label is statistically analyzed. Binary classifiers with an accuracy below a preset threshold are marked as non-transitive nodes. The labeling information of non-transitive nodes means that these classifiers do not participate in the backward propagation of label information in subsequent chain predictions; they are only used for local decision-making in this dimension, thereby reducing the interference of low-quality models on subsequent dimensions. The threshold can be set according to the business's tolerance for misclassification and label importance, using stricter labeling criteria for binary classifiers in key dimensions and slightly more lenient criteria for relatively minor dimensions.

[0090] After completing the above construction process, the multidimensional classifier chain model forms a unified structure, including a ranking chain for label prediction, a set of binary classification models for each dimension, and non-transitive node label information. The ranking chain specifies the order in which multidimensional labels are predicted. The set of binary classification models provides classification capabilities for each dimension based on a unified input feature vector and preceding labels. Non-transitive node labels are used to selectively filter out low-quality paths during information flow. This model structure can be saved as a parameter file, model configuration, and model weights, and can be directly loaded and used in subsequent runtime stages to achieve chain-like modeling of conditional dependencies between multidimensional labels.

[0091] This embodiment defines a multi-dimensional label set covering customer emotions, customer intentions, agent behavior, complaint anomalies, and recovery strategies in the offline stage. Based on the training dataset with complete dimensional label annotations, it calculates the conditional entropy between the labels of each dimension and uses an evolutionary search strategy to automatically determine the prediction order. Then, it trains a set of binary classification models with a one-to-one decomposition strategy and adds non-transitive node labels to low-accuracy binary classifiers. This can construct a multi-dimensional classifier chain model that takes into account both label dependency structure and model robustness. It can explicitly model the conditional dependencies between multi-dimensional labels, reduce the negative impact of low-quality sub-models on the overall prediction, and provide a stable and reliable basic classification structure for subsequent multi-dimensional joint judgment and anomaly level identification.

[0092] In one embodiment, step S50 above includes: S501, the unified input feature vector is used as the current input feature vector, and the current dimension label is determined according to the sorting chain; S502, the current input feature vector is input into the binary classification model set corresponding to the current dimension label for prediction, to obtain the current dimension predicted label and the current dimension confidence; S503, determine whether the confidence level of the current dimension is greater than or equal to the preset transmission threshold; S504, if the confidence level of the current dimension is greater than or equal to the preset transmission threshold, the predicted label of the current dimension is vectorized into a new feature vector, and the new feature vector is concatenated with the current input feature vector to form an updated input feature vector; S505, if the confidence level of the current dimension is less than the preset transmission threshold, then the current input feature vector is used as the updated input feature vector; S506, take the updated input feature vector as the current input feature vector, and determine the next dimension label as the current dimension label according to the sorting chain. Repeat the prediction, judgment and feature vector update steps until all dimension labels in the sorting chain have been predicted. S507 aggregates the current dimension prediction labels corresponding to all dimension labels to form a multidimensional classification prediction result.

[0093] In this embodiment, the unified input feature vector serves as the entry vector for the multidimensional classifier chain model. It can be understood as a high-dimensional numerical representation obtained by concatenating and transforming a single service interaction at the acoustic, semantic, and dialogue structure levels. In this stage, it is copied as the current input feature vector as the initial state for the chained prediction loop. The ranking chain is determined during the training phase of the multidimensional label set. Each position corresponds to a dimension label, such as customer sentiment dimension, customer intent dimension, agent behavior dimension, complaint anomaly dimension, or recovery strategy dimension. During runtime, the dimension label corresponding to the current position can be read from the ranking chain using an index, pointer, or iteration variable, and recorded as the current dimension label to indicate the prediction task to be completed in the current round.

[0094] There is a one-to-one correspondence between the binary classification model set and the current dimension label. Each set contains multiple binary classification models built for that dimension label, which can be formed based on different feature subspaces, different training subsets, or different model structures. During prediction, the current input feature vector is fed into the binary classification model set corresponding to the current dimension label in parallel or sequentially. Each binary classification model outputs the judgment result of the current dimension in a certain category and the corresponding confidence score. The predicted label and confidence score of the current dimension can be obtained by aggregating the outputs of the models in the set. The aggregation method can be majority voting, weighted average probability, or a weighted combination determined based on the performance of the validation set, converging the outputs of multiple models into a single label decision and a single confidence index, thus utilizing the stability of model ensemble while maintaining the consistency of the output format.

[0095] The current dimension confidence score measures the reliability of the predicted label for the current dimension. It can be derived from the probability output of the binary classification model ensemble on the positive and negative classes, log-likelihood values, inter-class distance, or other quantifiable criteria. This score is normalized and mapped to a fixed range for comparison with a preset propagation threshold. The preset propagation threshold is set during the model building phase using a validation dataset. Different thresholds can be applied to different dimension labels to control the propagation strength of label information in the chain structure, preventing unstable labels from being injected into subsequent dimensions when the confidence is too low. To determine whether the current dimension confidence score is greater than or equal to the preset propagation threshold, this value can be compared to the threshold immediately after each prediction round, providing a binary decision result that serves as a switch signal for subsequent feature concatenation or skipping.

[0096] When the confidence level meets the propagation condition, the predicted label for the current dimension needs to be converted into a numerical representation that the model can process, i.e., a new feature vector. This conversion can be achieved using one-hot encoding, mapping each possible label to a sparse vector space; or using label embedding, mapping the labels to a low-dimensional dense vector space, whose dimensions and distribution are fixed during training. After generating the new feature vector, it is concatenated with the current input feature vector along the feature dimension. Vector-level concatenation operations can be used to combine the two vectors sequentially into a longer updated input feature vector, ensuring that the concatenation order remains consistent across all samples so that the subsequent model can correctly identify the feature blocks corresponding to the labels in each dimension. When the confidence level does not meet the propagation condition, the current input feature vector is directly used as the updated input feature vector, which is equivalent to not introducing new label information at that dimension label, only retaining the feature state before entering this round, thereby blocking the influence of low-confidence results on subsequent dimensions.

[0097] After each round of prediction, the updated input feature vector is written back as the new current input feature vector. The index in the sorting chain shifts one position to the right, and the next dimension label is read as the new current dimension label. This triggers the cycle of binary classification model set prediction, confidence calculation, threshold judgment, and vector update again. Through this cyclical structure, the unified input feature vector propagates sequentially among different dimension labels. In each round, if the confidence condition is met, the new label information is encoded into the feature space, gradually forming a composite vector representation that simultaneously contains the original features and high-confidence label information. The loop termination condition can be designed as all dimension labels in the sorting chain having been predicted, or it can terminate early upon the occurrence of an abnormal interruption signal. In the case of normal termination, the current dimension prediction labels generated in each round are collected and combined according to the order in the sorting chain to form a multidimensional classification prediction result. The multidimensional classification prediction result can record the label value of each dimension in vector form, or it can use a structured record to save the label name, label value, and corresponding confidence score for interactive analysis and subsequent decision-making.

[0098] This embodiment introduces a multidimensional classifier chain structure based on a sorting chain onto a unified input feature vector. In each round of prediction, the confidence level output by the binary classification model set is compared with a preset propagation threshold. The current dimension prediction label with high confidence is vectorized into a new feature vector and concatenated into the current input feature vector. At the same time, the feature vector remains unchanged when the confidence level is insufficient. This allows explicit control over the propagation path and intensity of label information during multidimensional label prediction, enabling the conditional dependencies between labels to be gradually characterized. This reduces the risk of low-reliability labels accumulating and amplifying errors in the chain structure, thereby obtaining more stable and consistent multidimensional classification prediction results and providing a reliable label foundation for subsequent anomaly level identification and strategy decision-making.

[0099] In one embodiment, step S60 above includes: S601, parse the predicted label of the complaint anomaly dimension from the multidimensional classification prediction results; S602, determine the anomaly level based on the predicted label of the complaint anomaly dimension, the anomaly level including high anomaly level, medium anomaly level and low anomaly level; S603, if the anomaly level is a high anomaly level, then the target strategy is determined as the first strategy, the first strategy including creating a quality inspection work order and triggering the first external processing flow. S604, if the anomaly level is medium anomaly level, then the target strategy is determined as the second strategy, the second strategy includes sending reassurance information and triggering a second external connection processing flow, the second external connection processing flow is executed within a preset time limit; S605, if the anomaly level is a low anomaly level, then the target strategy is determined as a third strategy, the third strategy including sending notification information; S606, Execute the action corresponding to the target strategy and record the result of the strategy execution.

[0100] In this embodiment, the multidimensional classification prediction result can be understood as a set of labels generated for a single service interaction. Each dimension corresponds to an analytical perspective, such as customer sentiment, customer intent, agent behavior, complaint anomalies, and recovery strategies. Within this set, the predicted labels for the complaint anomaly dimension exist in discrete category form, and can be designed as level labels or risk type labels reflecting the degree of anomaly related to the complaint, such as potential complaints, explicit dissatisfaction, and escalating complaints. When the system enters the current processing stage, it reads the label value corresponding to the complaint anomaly dimension from the multidimensional classification prediction result, and combines this label value with its probability information as direct input for subsequent anomaly level calculation.

[0101] Anomaly levels provide a unified quantitative result for complaint-related anomalies, categorizing complex complaint tags into three discrete levels: high anomaly, medium anomaly, and low anomaly, facilitating a direct mapping to subsequent business strategies. Anomaly levels can be determined through pre-configured mapping rules, such as constructing a mapping table to map different complaint anomaly tags and their confidence intervals to the three levels; alternatively, probability scores can be combined, using interval division to categorize serious complaints with high confidence as high anomaly levels, moderate or ambiguous complaints as medium anomaly levels, and minor negative feedback or situations with unclear risk signals as low anomaly levels. During implementation, an anomaly level field can be generated for each service interaction instance, recording the corresponding calculation basis, including source tag, probability, and threshold interval, for traceability and auditing.

[0102] The target strategy describes the combination of response processes and business actions that the system needs to automatically trigger under a given anomaly level. For ease of management and expansion, strategy templates can be predefined for different levels, labeled as Strategy 1, Strategy 2, and Strategy 3, with specific content maintained in the configuration center. Strategy 1 targets high anomaly levels and handles highly sensitive or unsatisfactory interaction scenarios. Strategy content includes creating a quality inspection work order and triggering the first external processing flow. Creating a quality inspection work order generates a structured record through the quality inspection system interface. Record fields include call identifier, agent identifier, customer identifier, anomaly level, complaint anomaly dimension tag, and the time segment requiring review. These fields are written to the quality inspection task queue. The first external processing flow can be designed as a server-side scheduling flow. According to a preset priority, the call instance is pushed to personnel or system channels with processing permissions. Communication with the customer is conducted via telephone, audio / video channels, or other external channels. Flow nodes include task acceptance, external execution, and communication result recording. Timeout, reassignment, and escalation rules can be configured in the process engine.

[0103] The second strategy targets medium-level anomalies and handles interactions where there is significant dissatisfaction but has not yet escalated into a serious complaint. This strategy includes sending reassuring messages and triggering a second outbound processing flow. Reassuring messages can be generated through a message template system, pre-defined with expressions of concern for customer experience, a request for problem confirmation, and a brief explanation of the solution path. These messages are then pushed by the sending module based on the customer's preferred contact channel (SMS, in-app messages, email, or other electronic channels). The second outbound processing flow uses a similar process engine structure to the first, but can differ in priority, participating roles, and resource allocation. For example, a dedicated follow-up team or an automated outbound call robot may handle the main work. The preset time limit can be set to a relatively lenient time window that still falls within the customer's perception period. The start time of the outbound task is compared to the anomaly level generation time. If the outbound task is not completed within the preset time limit, timeout processing is triggered, including re-queuing, escalation to a high anomaly level, or triggering additional notifications. To ensure that time constraints are enforced, the second outbound processing flow needs to record the creation time, target completion time, and current status in the task data structure, and the timed scheduling module periodically scans for incomplete tasks.

[0104] The third strategy targets low-level anomalies and is used to handle interactions with weak risk signals or where only maintaining customer relationships is required. The strategy includes sending notification messages, which can use a templated structure and include elements such as a service summary, processing progress description, and prompts for subsequent contact channels. This provides customers with clear feedback without consuming excessive human resources. The unified messaging service module handles the sending of notification messages. After receiving the instruction to issue the third strategy, this module selects content matching the current business scenario from the template library, delivers it through the customer's preferred contact channels, and records the delivery result and reading status. For low-level anomalies, complex external processes are generally not triggered; instead, the intervention is completed through a one-time notification.

[0105] The strategy execution action depends on the previously determined target strategy. The target strategy can be viewed as a binding object of strategy identifier and strategy content. After anomaly level resolution, one of the first, second, or third strategies is assigned to the target strategy field. Subsequent execution units only need to read the target strategy field to schedule the corresponding execution logic, avoiding redundant level judgments across multiple submodules. During execution, each action node needs to record the strategy execution result in the log system or business database, including execution time, execution status, feedback information from external processes, customer response summaries, etc., thus forming a structured record for subsequent operational monitoring, quality inspection analysis, and retraining of the multidimensional classification model. This recording mechanism allows for tracing the complete chain from multidimensional classification prediction results to anomaly level, from anomaly level to target strategy, and from target strategy to actual business actions, ensuring that the decision-making and execution process is traceable and explainable.

[0106] This embodiment analyzes the predicted labels of the complaint anomaly dimension in the multidimensional classification prediction results and maps these labels to high, medium, and low anomaly levels. Combined with predefined first, second, and third strategies, it integrates the creation of quality inspection work orders, the sending of reassurance information, the triggering of different external processing flows, and the push of notification information into the target strategy field for driving. At the same time, it records the complete strategy execution results during the execution phase, which can form an automated closed loop from label recognition to response action in a single service interaction. This achieves a fine match between the severity of the anomaly and the intervention resource investment, improves the processing priority and response intensity of high-anomaly level interactions, reduces the escalation probability of medium-anomaly level interactions, and maintains the necessary communication frequency and transparency in low-anomaly level interactions. This provides a reusable strategy management foundation and traceable execution data for customer service quality inspection and trust recovery.

[0107] In one embodiment, an anomaly handling device based on a multidimensional classifier chain is provided, which corresponds one-to-one with the anomaly handling method based on a multidimensional classifier chain in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the anomaly handling device based on a multi-dimensional classifier chain of the present invention. The modules include a speech processing module 10, a semantic analysis module 20, a feature fusion module 30, a classifier invocation module 40, a classification prediction module 50, and a policy execution module 60. Detailed descriptions of each functional module are as follows: The voice processing module 10 is used to acquire the voice data stream during the service interaction process, perform role channel differentiation and voice transcription processing on the voice data stream, generate a text data sequence containing role identifiers, and extract the voice acoustic features of the voice data stream. The semantic analysis module 20 is used to perform semantic parsing on the text data sequence to generate a semantic representation vector, and to determine the dialogue structure features in the service interaction process based on the role identifier; The feature fusion module 30 is used to fuse the speech acoustic features, the semantic representation vector and the dialogue structure features to obtain a unified input feature vector. The classifier calling module 40 is used to call a pre-built multidimensional classifier chain model. The multidimensional classifier chain model is built based on a preset multidimensional label set. The construction process includes determining the degree of conditional dependence between the labels of each dimension in the multidimensional label set to establish a sorting chain for label prediction, and building a corresponding binary classification model set for each dimension label based on the label decomposition strategy. The classification prediction module 50 is used to input the unified input feature vector into the multidimensional classifier chain model, call the binary classification model set sequentially according to the order of the sorting chain for prediction, use the confidence index to control the transmission of the prediction results of the preceding node to the subsequent node during the prediction process, and output the multidimensional classification prediction result. The strategy execution module 60 is used to determine the anomaly level based on the multidimensional classification prediction result, match the corresponding target strategy based on the anomaly level, and execute the corresponding action according to the target strategy.

[0108] Specific limitations regarding the exception handling device based on multidimensional classifier chains can be found in the aforementioned limitations on the exception handling method based on multidimensional classifier chains, and will not be repeated here. Each module in the aforementioned exception handling device based on multidimensional classifier chains can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0109] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides deterministic and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements server-side functions or steps of an exception handling method based on a multi-dimensional classifier chain.

[0110] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements client-side functions or steps of an exception handling method based on a multi-dimensional classifier chain.

[0111] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Acquire the voice data stream during the service interaction process, perform role channel differentiation and speech-to-text processing on the voice data stream, generate a text data sequence containing role identifiers, and extract the speech acoustic features of the voice data stream. Semantic parsing is performed on the text data sequence to generate a semantic representation vector, and the dialogue structure features in the service interaction process are determined based on the role identifier; The speech acoustic features, the semantic representation vector, and the dialogue structure features are fused together to obtain a unified input feature vector. The pre-built multidimensional classifier chain model is invoked. The multidimensional classifier chain model is built based on a preset multidimensional label set. The construction process includes determining the degree of conditional dependence between the labels of each dimension in the multidimensional label set to establish a ranking chain for label prediction, and building a corresponding binary classification model set for each dimension label based on the label decomposition strategy. The unified input feature vector is input into the multidimensional classifier chain model, and the binary classification model set is called sequentially according to the order of the sorting chain for prediction. During the prediction process, the confidence index is used to control the transmission of the prediction results of the preceding node to the following node, and the multidimensional classification prediction result is output. Anomaly level is determined based on the multidimensional classification prediction results, a corresponding target strategy is matched based on the anomaly level, and a corresponding action is executed according to the target strategy.

[0112] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, and a computer program is stored thereon, which, when executed by a processor, performs the following steps: Acquire the voice data stream during the service interaction process, perform role channel differentiation and speech-to-text processing on the voice data stream, generate a text data sequence containing role identifiers, and extract the speech acoustic features of the voice data stream. Semantic parsing is performed on the text data sequence to generate a semantic representation vector, and the dialogue structure features in the service interaction process are determined based on the role identifier; The speech acoustic features, the semantic representation vector, and the dialogue structure features are fused together to obtain a unified input feature vector. The pre-built multidimensional classifier chain model is invoked. The multidimensional classifier chain model is built based on a preset multidimensional label set. The construction process includes determining the degree of conditional dependence between the labels of each dimension in the multidimensional label set to establish a ranking chain for label prediction, and building a corresponding binary classification model set for each dimension label based on the label decomposition strategy. The unified input feature vector is input into the multidimensional classifier chain model, and the binary classification model set is called sequentially according to the order of the sorting chain for prediction. During the prediction process, the confidence index is used to control the transmission of the prediction results of the preceding node to the following node, and the multidimensional classification prediction result is output. Anomaly level is determined based on the multidimensional classification prediction results, a corresponding target strategy is matched based on the anomaly level, and a corresponding action is executed according to the target strategy.

[0113] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0114] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0115] It should be noted that if any AI models, software tools, or components not belonging to this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use. The above-described embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

[0116] The user personal information involved in this application embodiment is all authorized (knowing and consenting) by the relevant parties or fully authorized by all parties, and the executing entity can obtain it through various open, legal and compliant means. The collection, storage, use, processing, transmission, provision and disclosure of the information, data and signals involved all comply with the relevant laws and regulations of the relevant countries and regions, and do not violate public order and good morals.

Claims

1. An anomaly handling method based on a multidimensional classifier chain, characterized in that, Includes the following steps: Acquire the voice data stream during the service interaction process, perform role channel differentiation and speech-to-text processing on the voice data stream, generate a text data sequence containing role identifiers, and extract the speech acoustic features of the voice data stream. Semantic parsing is performed on the text data sequence to generate a semantic representation vector, and the dialogue structure features in the service interaction process are determined based on the role identifier; The speech acoustic features, the semantic representation vector, and the dialogue structure features are fused together to obtain a unified input feature vector. The pre-built multidimensional classifier chain model is invoked. The multidimensional classifier chain model is built based on a preset multidimensional label set. The construction process includes determining the degree of conditional dependence between the labels of each dimension in the multidimensional label set to establish a ranking chain for label prediction, and building a corresponding binary classification model set for each dimension label based on the label decomposition strategy. The unified input feature vector is input into the multidimensional classifier chain model, and the binary classification model set is called sequentially according to the order of the sorting chain for prediction. During the prediction process, the confidence index is used to control the transmission of the prediction results of the preceding node to the following node, and the multidimensional classification prediction result is output. Anomaly level is determined based on the multidimensional classification prediction results, a corresponding target strategy is matched based on the anomaly level, and a corresponding action is executed according to the target strategy.

2. The anomaly handling method based on a multidimensional classifier chain as described in claim 1, characterized in that, Acquire the voice data stream during the service interaction process, perform role channel differentiation and speech-to-text processing on the voice data stream, generate a text data sequence containing role identifiers, and extract the speech acoustic features of the voice data stream, including: Receive raw audio data during service interaction as a voice data stream; The voice data stream is differentiated by role channel to obtain independent voice channels corresponding to different speakers; Speech transcription is performed on each independent speech channel to generate the initial text corresponding to the independent speech channel and the timestamp information of each sentence of the initial text; Assign a role identifier to each independent voice channel and associate the role identifier with the initial text transcribed from the independent voice channel; Based on the timestamp information, the initial texts assigned with the role identifiers are sequentially merged to form a text data sequence containing the role identifiers; Speech acoustic features, including volume variation features, speech rate variation features, and intonation fluctuation features, are extracted from each independent speech channel.

3. The anomaly handling method based on a multidimensional classifier chain as described in claim 1, characterized in that, Semantic parsing is performed on the text data sequence to generate semantic representation vectors, and the dialogue structure features in the service interaction process are determined based on the role identifier, including: The text data sequence is segmented into words to obtain the segmentation results; Determine the keyword weight vector based on the word segmentation results; The text data sequence is converted into a deep semantic vector using a semantic embedding model; Based on the keyword weight vector and the deep semantic vector, a semantic representation vector is generated; Based on the temporal sequence information of sentences in the text data sequence and the role identifier, the speaking turns are divided; Based on the time sequence information and the role identifier, the number of interruptions during the service interaction process is counted; Based on the time sequence information, determine the percentage of silent time periods during the service interaction process; The number of speaking turns, the number of interruptions, and the percentage of silent periods are used as dialogue structure features.

4. The anomaly handling method based on a multidimensional classifier chain as described in claim 1, characterized in that, The speech acoustic features, the semantic representation vector, and the dialogue structure features are fused to obtain a unified input feature vector, including: The speech acoustic features, the semantic representation vector, and the dialogue structure features are linearly projected to obtain projected acoustic features, projected semantic features, and projected structure features, respectively. The projection acoustic features, projection semantic features, and projection structural features are input into a preset gating network to determine adaptive weight coefficients corresponding to the projection acoustic features, projection semantic features, and projection structural features, respectively. The projection acoustic features, projection semantic features, and projection structural features are weighted and summed using the adaptive weighting coefficients to obtain weighted fusion features. The weighted fusion features are subjected to a nonlinear transformation to obtain enhanced fusion features, which are then used as a unified input feature vector.

5. The anomaly handling method based on a multidimensional classifier chain as described in claim 1, characterized in that, Before invoking the pre-built multidimensional classifier chain model, the following is also included: Define a multidimensional tag set, which includes customer emotion dimension, customer intent dimension, agent behavior dimension, complaint anomaly dimension, and recovery strategy dimension; Obtain the training dataset with complete dimensional labels; Based on the training dataset, the conditional entropy value between any two dimensional labels in the multidimensional label set is determined as the degree of conditional dependency. Based on the conditional entropy value, an evolutionary search strategy is run to determine the prediction order among dimensional labels and establish a sorting chain for label predictions; For each dimension label in the sorting chain, a one-to-one decomposition strategy is adopted to train the corresponding set of binary classification models. The prediction accuracy of each binary classifier in the binary classification model set is analyzed using the validation dataset, and binary classifiers with prediction accuracy below a preset threshold are marked as non-transfer nodes. Based on the sorting chain, the binary classification model set, and the labels of the non-transitive nodes, a multidimensional classifier chain model is constructed.

6. The anomaly handling method based on a multidimensional classifier chain as described in claim 1, characterized in that, The unified input feature vector is input into the multidimensional classifier chain model, and the binary classification model set is called sequentially according to the order of the sorted chain for prediction. During the prediction process, the confidence index is used to control the transmission of the prediction results of the preceding node to the following node, and the multidimensional classification prediction result is output, including: The unified input feature vector is used as the current input feature vector, and the current dimension label is determined according to the sorting chain; The current input feature vector is input into the binary classification model set corresponding to the current dimension label for prediction, and the predicted label and confidence of the current dimension are obtained. Determine whether the confidence level of the current dimension is greater than or equal to a preset transmission threshold; If the confidence level of the current dimension is greater than or equal to the preset transmission threshold, the predicted label of the current dimension is vectorized into a new feature vector, and the new feature vector is concatenated with the current input feature vector to form an updated input feature vector. If the confidence level of the current dimension is less than the preset transmission threshold, then the current input feature vector is used as the updated input feature vector. The updated input feature vector is used as the current input feature vector, and the next dimension label is determined according to the sorting chain as the current dimension label. The prediction, judgment and feature vector update steps are repeated until all dimension labels in the sorting chain have been predicted. The predicted labels for the current dimension corresponding to all dimension labels are collected to form a multidimensional classification prediction result.

7. The anomaly handling method based on a multidimensional classifier chain as described in claim 1, characterized in that, Anomaly levels are determined based on the multidimensional classification prediction results, corresponding target strategies are matched based on the anomaly levels, and corresponding actions are executed according to the target strategies, including: The predicted labels for the complaint anomaly dimension are parsed from the multidimensional classification prediction results; The anomaly level is determined based on the predicted label of the anomaly dimension of the complaint, and the anomaly level includes high anomaly level, medium anomaly level and low anomaly level; If the anomaly level is a high anomaly level, then the target strategy is determined as the first strategy, which includes creating a quality inspection work order and triggering a first external processing flow. If the anomaly level is medium anomaly level, then the target strategy is determined as the second strategy. The second strategy includes sending reassurance information and triggering a second external connection processing flow, which is executed within a preset time limit. If the anomaly level is a low anomaly level, then the target strategy is determined to be the third strategy, which includes sending notification information; Execute the action corresponding to the target strategy and record the result of the strategy execution.

8. An anomaly handling device based on a multidimensional classifier chain, characterized in that, The anomaly handling device based on the multidimensional classifier chain includes: The voice processing module is used to acquire the voice data stream during the service interaction process, perform role channel differentiation and voice transcription processing on the voice data stream, generate a text data sequence containing role identifiers, and extract the voice acoustic features of the voice data stream. The semantic analysis module is used to perform semantic parsing on the text data sequence to generate semantic representation vectors, and to determine the dialogue structure features in the service interaction process based on the role identifier; The feature fusion module is used to fuse the speech acoustic features, the semantic representation vector and the dialogue structure features to obtain a unified input feature vector; The classifier calling module is used to call a pre-built multidimensional classifier chain model. The multidimensional classifier chain model is built based on a preset multidimensional label set. The construction process includes determining the degree of conditional dependence between the labels of each dimension in the multidimensional label set to establish a sorting chain for label prediction, and building a corresponding binary classification model set for each dimension label based on the label decomposition strategy. The classification prediction module is used to input the unified input feature vector into the multidimensional classifier chain model, call the binary classification model set sequentially according to the order of the sorting chain for prediction, use the confidence index to control the transmission of the prediction results of the preceding node to the subsequent node during the prediction process, and output the multidimensional classification prediction result. The strategy execution module is used to determine the anomaly level based on the multidimensional classification prediction results, match the corresponding target strategy based on the anomaly level, and execute the corresponding action according to the target strategy.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and an exception handling program based on a multidimensional classifier chain stored in the memory and executable on the processor. When executed by the processor, the exception handling program based on the multidimensional classifier chain implements the steps of the exception handling method based on a multidimensional classifier chain as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores an exception handling program based on a multidimensional classifier chain. When the exception handling program based on the multidimensional classifier chain is executed by the processor, it implements the steps of the exception handling method based on the multidimensional classifier chain as described in any one of claims 1-7.

Images