A multi-label classification method and system for a dispatching scene
By extracting features and enhancing label knowledge from work order texts in the dispatching scenario, and dynamically generating classification thresholds, the problem of mis-dispatch and missed dispatch caused by fixed thresholds is solved. Differentiated decision-making and semantic alignment of multi-label classification are realized, improving the accuracy of dispatching and business adaptability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CAPINFO CO LTD
- Filing Date
- 2025-10-29
- Publication Date
- 2026-05-22
AI Technical Summary
Existing technologies suffer from errors and omissions in order dispatching due to fixed thresholds, weak generalization ability due to ignoring label semantics, severe noise interference due to failure to distinguish feature importance, and lack of dynamic focusing ability on key intent fields, resulting in a lack of logical consistency and accuracy in order dispatching results.
By extracting features from work order texts to generate global text semantic representations, and combining them with pre-constructed business tag semantic embedding vectors for tag knowledge enhancement, classification thresholds are dynamically generated. An end-to-end collaborative mechanism of importance neural networks and tag knowledge enhancement is adopted to achieve semantic alignment and differentiated decision-making.
It significantly reduced the false negative rate of high-value tags and the false positive rate of low-value tags, improved the model's ability to generalize to diverse user expressions and the recognition accuracy of key fields, and enhanced the business adaptability of dispatching and the overall prediction accuracy.
Smart Images
Figure CN121301576B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of label classification technology, and in particular to a multi-label classification method and system for dispatch scenarios. Background Technology
[0002] In intelligent work order dispatching systems for government services, urban management, or enterprise operations and maintenance, multi-label text classification is the core technology for automatic work order dispatching. A work order typically contains multiple structured or unstructured fields such as user request, event type, and location, and may be associated with multiple responsible units simultaneously. Therefore, a multi-label classification model is required for joint prediction. Existing technologies mainly employ three methods: First, the independent binary classifier method, which trains an independent binary classification model for each responsible unit. While simple to implement, it completely ignores the business correlation between labels, leading to a lack of logical consistency in dispatch results. Second, joint classification models based on deep neural networks, while capable of outputting multi-label probabilities end-to-end, generally use a uniform threshold for judgment, failing to adapt to the reality of vastly different label importance in dispatch scenarios. For example, "property management company," as the primary responsible party, must be assigned a dispatch, while "district-level department" plays only a supporting role; a fixed threshold easily leads to the omission of high-value labels or the misassignment of low-value labels. Third, while label embedding or knowledge enhancement methods attempt to introduce label semantic information, they are mostly used only to improve classification scores without being linked to decision thresholds. Furthermore, when faced with diverse user expressions, the model's lack of deep understanding of label semantics results in insufficient generalization ability. In addition, the work order text itself contains a large number of redundant or irrelevant fields. Traditional models treat all content equally, lacking the ability to dynamically focus on key intent fields, further reducing prediction accuracy. In summary, existing technologies face three common defects in order dispatch scenarios: (1) fixed thresholds lead to mis-dispatch and missed dispatch; (2) ignoring label semantics leads to weak generalization ability; and (3) failure to distinguish the importance of features leads to severe noise interference. Among these, the inability to achieve differentiated and refined order dispatch decisions due to the use of fixed thresholds is the most critical bottleneck restricting the practicality of the system. Summary of the Invention
[0003] In view of this, this invention proposes a multi-label classification method and system for work order dispatching scenarios, which can solve problems such as mis-dispatch and missed dispatch of work orders caused by fixed thresholds in work order dispatching scenarios. This invention provides the following technical solution:
[0004] A multi-label classification method for dispatch scenarios, the method comprising:
[0005] Feature extraction is performed on the input work order text to generate a global text semantic representation;
[0006] Based on the pre-constructed semantic embedding vectors of each preset business tag, the global text semantic representation is enhanced with tag knowledge to achieve alignment between the text representation and the tag semantics in the semantic space.
[0007] Based on the aligned text semantic representation, the original confidence level corresponding to each preset business label is calculated by the main classifier.
[0008] For each preset business tag and the work order text, a classification threshold corresponding to the business tag is dynamically generated based on the global text semantic representation, the semantic embedding vector of the business tag, the original confidence score, and the business prior information of the business tag.
[0009] The original confidence level is determined based on the classification threshold, and the multi-label classification result corresponding to the work order text is output to achieve differentiated work order dispatch decision.
[0010] Optionally, the step of extracting features from the input work order text to generate a global text semantic representation includes:
[0011] The multiple structured fields contained in the work order text are pooled separately to obtain vector representations of each field;
[0012] The importance score of each field is calculated based on learnable parameters, and the field weight is obtained through normalization. The field vectors are then weighted and fused to obtain the field-level representation.
[0013] Differentiable sparse gating mechanism is applied to the word sequence of the work order text to generate a soft mask for each word, and the word vector is weighted and pooled based on the soft mask to obtain word-level representation;
[0014] The field-level representation and the lexical-level representation are fused together to form the global text semantic representation.
[0015] Optionally, the application of a differentiable sparse gating mechanism to the lexical sequence of the work order text is implemented in the following way:
[0016] Calculate the original importance score for each lexical unit;
[0017] Random noise is introduced into the original importance score and processed by a smoothing function controlled by a temperature parameter to obtain a continuous relaxation gate value;
[0018] The continuously relaxed gate value is linearly stretched and clipped to the 0 to 1 range to obtain the final soft mask;
[0019] The temperature parameter is used to control the sparsity of the mask and is gradually reduced during training.
[0020] Optionally, the classification threshold for dynamically generating the business tag includes:
[0021] Construct a fusion feature vector, which includes the global text semantic representation, the semantic embedding vector of the business tag, the original confidence score, the prior probability or business cost weight of the business tag, external knowledge injection information, and the contextual features of the work order.
[0022] The fused feature vector is input into the threshold sub-network, which outputs an initial threshold.
[0023] The initial threshold is mapped to a preset bounded interval through an activation function to obtain the final classification threshold.
[0024] Optionally, the global text semantic representation is generated through an importance neural network layer, and the original confidence score is calculated based on the text representation enhanced with label knowledge.
[0025] The method employs end-to-end joint training to enable the network parameters involved in generating global text semantic representations, performing label knowledge enhancement, and dynamically generating classification thresholds to be updated collaboratively under a unified optimization objective.
[0026] Optionally, the method further includes:
[0027] The network parameters used to generate the original confidence score and the classification threshold are optimized using a joint loss function;
[0028] The joint loss function includes shifted binary cross-entropy loss and marginal safety margin loss;
[0029] The shifted binary cross-entropy loss is calculated based on the difference between the original confidence level and the classification threshold, and is used to jointly optimize the main classifier and the threshold generation process.
[0030] The marginal safety margin loss is used to impose a positive safety margin constraint on positive class samples and a negative safety margin constraint on negative class samples, so as to force the prediction results away from the decision boundary and improve classification stability.
[0031] This invention further discloses a multi-label classification system for dispatch scenarios, including:
[0032] The feature extraction module is used to extract features from the input work order text and generate a global text semantic representation.
[0033] The tag knowledge enhancement module is used to enhance the global text semantic representation with tag knowledge based on the semantic embedding vectors of each preset business tag, so as to achieve the alignment of text representation and tag semantics in the semantic space.
[0034] The main classifier module is used to calculate the original confidence level corresponding to each preset business label based on the aligned text semantic representation;
[0035] The dynamic threshold generation module is used to dynamically generate a classification threshold corresponding to each preset business tag and the work order text, based on the global text semantic representation, the semantic embedding vector of the business tag, the original confidence score, and the business prior information of the business tag.
[0036] The classification decision module is used to determine the original confidence level based on the classification threshold and output the multi-label classification result corresponding to the work order text to realize differentiated work order dispatch decision.
[0037] Optionally, the storage medium stores a computer program that, when executed by a processor, implements the method described above.
[0038] The present invention further discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described method.
[0039] The present invention further discloses a computer program product, including a computer program that implements the above-described method when executed by a processor.
[0040] According to the technical solution of the present invention, in the multi-label classification process of the dispatch scenario, feature extraction is performed on the work order text to generate a global text semantic representation. The semantic embedding vectors of each preset business label are combined to perform label knowledge enhancement to achieve semantic alignment. The original confidence is calculated based on the aligned representation. Then, for each label-sample pair, a personalized classification threshold is dynamically generated by integrating the global text semantic representation, label semantic embedding, original confidence, and business prior information. Finally, the multi-label result is output based on the threshold. This effectively solves the core bottleneck problem of mis-dispatch and missed dispatch caused by the use of a uniform fixed threshold in the prior art. This dynamic threshold mechanism can adaptively adjust the decision boundary, significantly reducing the false negative rate on high-value labels and effectively suppressing the false negative rate on low-value or easily confused labels. At the same time, this mechanism, together with label knowledge enhancement and importance perception feature extraction, forms an end-to-end collaborative link of "noise reduction and focusing - semantic alignment - differentiated decision-making". This not only improves the model's ability to generalize to diverse user expressions, but also enhances the recognition accuracy of key fields and the overall business adaptability of order dispatch, thereby enabling intelligent order dispatch in complex, multi-field, multi-label, and noisy real-world scenarios. Attached Figure Description
[0041] For illustrative purposes and not limiting, the present invention will now be described in conjunction with embodiments and accompanying drawings, wherein:
[0042] Figure 1 This is a flowchart illustrating the multi-label classification method for dispatch scenarios in an embodiment of the present invention.
[0043] Figure 2 This is a schematic diagram of the structure of a multi-label classification system for dispatch scenarios in an embodiment of the present invention;
[0044] Figure 3 This is a schematic diagram of the structure of the electronic device in an embodiment of the present invention. Detailed Implementation
[0045] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application.
[0046] It should be noted that, where there is no conflict, the embodiments and features of the embodiments in this application can be combined with each other. The embodiments of this application will be described in detail below with reference to the accompanying drawings.
[0047] refer to Figure 1 This embodiment discloses a multi-label classification method for dispatch scenarios, the method including the following steps:
[0048] S100: Extract features from the input work order text to generate a global text semantic representation.
[0049] Specifically, firstly, the original work order text is split according to its structured fields, such as "event description," "location," "request type," and "submission time." The text within each field is then encoded into a sequence of lexical embeddings using a pre-trained language model. ,in The field number is used; subsequently, pooling operations (such as mean pooling or max pooling) are performed on the word embedding sequence of each field to obtain a fixed-dimensional vector representation of each field. Then, the importance score for each field was calculated. ,in, For learnable weight vectors, The biases are learnable and the field weights are obtained by normalization using the softmax function. This is then integrated into a field-level representation. , Meanwhile, the word sequence of the entire work order text. Apply a token-level differentiable sparse gating mechanism: calculate the original importance score for each lexical unit. ,in and These are learnable parameters. Gumbel noise is then introduced and controlled by temperature parameters. Smoothing function processing of control, i.e., sampling Calculate the continuous relaxation gate value ,in, This is the sigmoid function. Next, regarding... Perform linear stretching: ,in, , is the lower bound hyperparameter for stretching. To stretch the upper bound hyperparameter, the final soft mask is: Based on this mask, weighted pooling of the word vectors is performed to obtain word-level representations: ,in, To prevent division by zero of extremely small positive numbers. Finally, and By concatenating or weighted fusion, a high signal-to-noise ratio global text semantic representation is formed. This process dynamically suppresses irrelevant fields and redundant words through field-level and token-level gating, significantly improving the expressive power of key intents compared to existing technologies, and providing high-quality input for subsequent tag knowledge enhancement and dynamic threshold generation.
[0050] S200: Based on the pre-constructed semantic embedding vectors of each preset business tag, the global text semantic representation is enhanced with tag knowledge to achieve alignment between the text representation and the tag semantics in the semantic space.
[0051] Specifically, for each preset business tag A semantic description text is constructed, which comprehensively includes tag names, common aliases, business description documents, and Top-K typical work order fragments retrieved from historical work orders, ensuring comprehensive semantic coverage and business representativeness. Subsequently, the aforementioned description text is input into the tag encoder, which shares a vocabulary with the work order encoder. After pooling, the semantic embedding vector of the tag is obtained. All tags Together, they form the tag knowledge base. During the training phase, the global text semantic representation generated in the preceding steps is combined with the semantic embedding vectors of each tag. Interactive methods are employed, for example, calculating initial relevance scores through dot product, concatenation followed by MLP, or bilinear fusion, and using these scores as part of the input to the main classifier. Furthermore, a semantic alignment training objective is introduced, such as adding a representation alignment loss term to the loss function. Its form can be a contrastive loss or a cosine similarity regularization term, which forces positive sample pairs to... and In semantic space, negative sample pairs are placed closer together, while negative sample pairs are placed further apart. This alignment mechanism enables the model to understand the deep semantic relationships between diverse user expressions, allowing for correct recall based on semantic similarity even if the tag name is not explicitly present in the work order. Finally, the aligned text representations are used by the main classifier to calculate the original confidence scores for each tag. This enables the classification boundary to vary "depending on the semantics of the label," significantly improving the model's generalization ability and recall stability in cross-regional, new expression, or paraphrasing scenarios.
[0052] S300: Based on the aligned text semantic representation, the original confidence level corresponding to each preset business label is calculated by the main classifier.
[0053] Specifically, the aligned text semantic representation obtained after being enhanced by tag knowledge. This serves as the input to the main classifier; the main classifier employs a parameterized multilayer perceptron (MLP) structure, typically containing one or more fully connected layers, each followed by a ReLU activation function and an optional Dropout layer, for each preset business label. The main classifier calculates its original confidence score through a separate output head (or a linear projection layer with shared weights). All tags This constructs the original confidence vector for the multi-label system. This vector is output directly without sigmoid normalization to preserve its original numerical range, which is then used by the subsequent dynamic thresholding module for difference comparison. It is worth noting that this main classifier shares the training objective with the label knowledge enhancement module and updates its parameters synchronously during end-to-end joint optimization, thus... It not only reflects the degree of matching between text content and tags, but also implicitly integrates prior semantic information of tags, thus providing a highly discriminative input basis for dynamic threshold adjustment.
[0054] S400: For each preset business tag and the work order text, dynamically generate the classification threshold corresponding to the business tag based on the global text semantic representation, the semantic embedding vector of the business tag, the original confidence level, and the business prior information of the business tag.
[0055] First, construct a feature vector that integrates information from multiple sources. Its specific form is: ,in, Prior information for the label can be historical statistical prior probabilities or manually set business cost weights. Inject vectors into external knowledge, including inter-label constraints from knowledge graphs or rule engines. The work order's contextual characteristics include information such as urgency, customer level, origin region, or submission time. Then, Input to a parameterized thresholding subnetwork This sub-network employs a multilayer perceptron (MLP) structure, and outputs a dynamic threshold for the current sample label pair. Then it is mapped and linearly scaled to a preset range using the Sigmoid function. ,Right now This mechanism enables the threshold to be dynamically adjusted based on the semantic clarity of the sample, the importance of the label, business rules, and external constraints, thereby achieving refined decision-making that varies "from sample to sample and from label to label," significantly alleviating the misassignment and omission problems caused by fixed thresholds.
[0056] S500: Based on the classification threshold, determine the original confidence level and output the multi-label classification result corresponding to the work order text to achieve differentiated work order dispatch decision.
[0057] For each preset business tag The original confidence scores output by the main classifier. With dynamically generated classification thresholds The comparison is performed using a decision rule in the form of an indicator function: ,in, The indicator function outputs 1 when the condition is met (indicating the tag is activated and should be assigned an order, positive class), and 0 otherwise (indicating no order assigned, negative class). This decision-making mechanism abandons the rigid decision-making method of the traditional uniform threshold, and instead dynamically adjusts the decision boundary based on the semantic clarity of each tag sample pair, the business importance of the tag, historical prior probability, external knowledge constraints, and work unit information, thereby realizing a differentiated order assignment strategy of "preventing missed judgments for high-value tags and preventing false judgments for low-value tags". Finally, the above decision is performed in parallel on all tags to obtain a binary multi-tag vector. As the final output of the dispatch system, it directly drives the subsequent work order distribution process, significantly improving the accuracy of dispatch, business adaptability, and user experience.
[0058] Furthermore, in one specific embodiment of the present invention, to achieve end-to-end joint training of the dynamic threshold module and the main classifier, a differentiable composite loss function is used to uniformly optimize all learnable parameters. This composite loss function consists of two parts:
[0059] First, a shifted binary cross-entropy loss is used: ,in, The raw confidence score output by the main classifier. The classification threshold is dynamically generated. For the true label, BCE is the binary cross-entropy loss function, which will... The difference is used as a valid input, allowing the main classifier and the threshold sub-network to adapt to each other during training, so as to jointly optimize the main classifier and the threshold network.
[0060] Secondly, the marginal safety margin loss is introduced: ,in, and As a preset safety margin hyperparameter, this loss is mandatory: for positive class samples... Must be greater than Negative class samples must be smaller than This pushes the prediction results away from the decision boundary, i.e. This significantly reduces prediction oscillations in the critical region and improves the robustness and stability of the model in actual order dispatching.
[0061] The final definition of the total loss function is: ,in, The weighting coefficients for the marginal loss are determined through optimization on the validation set. During training, the AdamW optimizer is used to update the parameters synchronously, ensuring that the high signal-to-noise ratio representation of the important neural network layer output, the semantic alignment brought by the label knowledge enhancement, and the personalized decision-making ability of the dynamic threshold co-evolve under a unified goal. Ultimately, this achieves the technical effect of a complete solution that combines noise reduction focusing, semantic generalization, and differentiated boundaries.
[0062] In summary, this invention achieves multi-label classification for dispatch scenarios through end-to-end collaboration of importance neural networks, label knowledge enhancement, and dynamic threshold adjustment: It effectively suppresses redundant and noisy information in work orders through a two-level gating mechanism at the field and token levels, significantly improving the signal-to-noise ratio of key intents; it encodes labels as semantic embeddings and trains the model through semantic alignment to understand the inherent connections between diverse expressions, greatly enhancing generalization ability; it integrates text representation, label semantics, original confidence, business priors, external knowledge, and contextual features to adaptively generate personalized thresholds for each label-sample pair, and combines shifted binary cross-entropy loss and marginal safety margin loss for end-to-end optimization, forcing predictions away from the decision boundary, thereby significantly reducing the missed dispatch rate of high-value labels and the mis-dispatch rate of low-value labels under differentiated business rules. These three elements constitute a complete technical chain of noise reduction focusing, semantic generalization, and differentiated decision-making. They are indispensable and together solve the problems of misjudgment, missed judgment, and insufficient generalization caused by fixed thresholds, ignoring label semantics, and feature noise interference in existing technologies. This enables the model to have both accuracy, stability, and business adaptability in complex, multi-field, and multi-label real-world order dispatching scenarios.
[0063] refer to Figure 2This embodiment further discloses a multi-label classification system for dispatch scenarios, specifically including a feature extraction module 21, a label knowledge enhancement module 22, a main classifier module 23, a dynamic threshold generation module 24, and a classification decision module 25. Detailed descriptions follow:
[0064] The feature extraction module 21 is used to extract features from the input work order text and generate a global text semantic representation.
[0065] The tag knowledge enhancement module 22 is used to enhance the global text semantic representation with tag knowledge based on the semantic embedding vectors of each preset business tag, so as to achieve the alignment of text representation and tag semantics in the semantic space.
[0066] The main classifier module 23 is used to calculate the original confidence level corresponding to each preset business label based on the aligned text semantic representation through the main classifier.
[0067] The dynamic threshold generation module 24 is used to dynamically generate a classification threshold corresponding to each preset business tag and the work order text, based on the global text semantic representation, the semantic embedding vector of the business tag, the original confidence level, and the business prior information of the business tag.
[0068] The classification decision module 25 is used to determine the original confidence level based on the classification threshold and output the multi-label classification result corresponding to the work order text to realize differentiated work order dispatch decision.
[0069] Figure 3 A schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention, such as... Figure 3 As shown, the electronic device 50 includes: a processor 501, a memory 502, and a bus 503;
[0070] The processor 501 and the memory 502 communicate with each other via the bus 503; the processor 501 is used to call the program instructions in the memory 502 to execute the methods provided in the above-described embodiments.
[0071] This embodiment provides a non-transitory computer-readable storage medium that stores computer instructions that cause a computer to execute the methods provided in the above-described embodiments.
[0072] Those skilled in the art will understand that all or part of the steps of the above-described method implementation can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above-described method implementation. The aforementioned storage medium includes various storage media capable of storing program code, such as ROM, RAM, magnetic disk, or optical disk.
[0073] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0074] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of each embodiment or some parts of the embodiments.
[0075] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can occur depending on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.
Claims
1. A multi-label classification method for dispatch scenarios, characterized in that, The method includes: Feature extraction is performed on the input work order text to generate a global text semantic representation; Based on the pre-constructed semantic embedding vectors of each preset business tag, the global text semantic representation is enhanced with tag knowledge to achieve alignment between the text representation and the tag semantics in the semantic space. Based on the aligned text semantic representation, the original confidence level corresponding to each preset business label is calculated by the main classifier. For each preset business tag and the work order text, a classification threshold corresponding to the business tag is dynamically generated based on the global text semantic representation, the semantic embedding vector of the business tag, the original confidence score, and the business prior information of the business tag. This includes: constructing a fusion feature vector, which includes the global text semantic representation, the semantic embedding vector of the business tag, the original confidence score, the prior probability or business cost weight of the business tag, external knowledge injection information, and the contextual features of the work order; inputting the fusion feature vector into a threshold sub-network to output an initial threshold; and mapping the initial threshold to a preset bounded interval through an activation function to obtain the final classification threshold. The original confidence level is determined based on the classification threshold, and the multi-label classification result corresponding to the work order text is output to achieve differentiated work order dispatch decision.
2. The multi-label classification method for dispatch scenarios according to claim 1, characterized in that, The step of extracting features from the input work order text to generate a global text semantic representation includes: The multiple structured fields contained in the work order text are pooled separately to obtain vector representations of each field; The importance score of each field is calculated based on learnable parameters, and the field weight is obtained through normalization. The field vectors are then weighted and fused to obtain the field-level representation. A differentiable sparse gating mechanism is applied to the word sequence of the work order text to generate a soft mask for each word, and the word vector is weighted and pooled based on the soft mask to obtain a word-level representation. The field-level representation and the lexical-level representation are fused together to form the global text semantic representation.
3. The multi-label classification method for dispatch scenarios according to claim 2, characterized in that, The application of a differentiable sparse gating mechanism to the lexical sequence of the work order text includes: Calculate the original importance score for each lexical unit; Random noise is introduced into the original importance score and processed by a smoothing function controlled by a temperature parameter to obtain a continuous relaxation gate value; The continuously relaxed gate value is linearly stretched and clipped to the 0 to 1 range to obtain the final soft mask; The temperature parameter is used to control the sparsity of the mask and is gradually reduced during training.
4. The multi-label classification method for dispatch scenarios according to claim 1, characterized in that, The global text semantic representation is generated through an importance neural network layer, and the original confidence score is calculated based on the text representation enhanced with label knowledge.
5. The multi-label classification method for dispatch scenarios according to claim 4, characterized in that, The method further includes: The network parameters used to generate the original confidence score and the classification threshold are optimized using a joint loss function; The joint loss function includes shifted binary cross-entropy loss and marginal safety margin loss; The shifted binary cross-entropy loss is calculated based on the difference between the original confidence level and the classification threshold, and is used to jointly optimize the main classifier and the threshold generation process. The marginal safety margin loss is used to impose a positive safety margin constraint on positive class samples and a negative safety margin constraint on negative class samples, so as to force the prediction results away from the decision boundary and improve classification stability.
6. A multi-label classification system for dispatch scenarios, characterized in that, include: The feature extraction module is used to extract features from the input work order text and generate a global text semantic representation. The tag knowledge enhancement module is used to enhance the global text semantic representation with tag knowledge based on the semantic embedding vectors of each preset business tag, so as to achieve the alignment of text representation and tag semantics in the semantic space. The main classifier module is used to calculate the original confidence level corresponding to each preset business label based on the aligned text semantic representation; A dynamic threshold generation module is used to dynamically generate a classification threshold corresponding to each preset business tag and the work order text, based on the global text semantic representation, the semantic embedding vector of the business tag, the original confidence score, and the business prior information of the business tag. This includes: constructing a fused feature vector, which includes the global text semantic representation, the semantic embedding vector of the business tag, the original confidence score, the prior probability or business cost weight of the business tag, external knowledge injection information, and the contextual features of the work order; inputting the fused feature vector into a threshold sub-network to output an initial threshold; and mapping the initial threshold to a preset bounded interval through an activation function to obtain the final classification threshold. The classification decision module is used to determine the original confidence level based on the classification threshold and output the multi-label classification result corresponding to the work order text to realize differentiated work order dispatch decision.
7. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1-5.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method of any one of claims 1-5.
9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-5.