Hot line work order multistage classification model adjustment method and device, medium and equipment
By preprocessing and feature extraction of hotline work order data, and adjusting the model by combining hierarchical structure information and domain knowledge, the problem of error accumulation in multi-level classification of hotline work orders in traditional text classification technology has been solved, achieving higher classification accuracy and policy response capability.
Patent Information
- Application Number
- CN202511055283.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-30
- Publication Date
- 2025-11-21
AI Technical Summary
Traditional text classification techniques struggle to handle the semantic complexity and diversity of hotline work orders, resulting in poor classification performance. Furthermore, they suffer from error accumulation issues in multi-level classification systems and are unable to capture fine-grained semantic differences and respond to policy changes.
By preprocessing, labeling, and extracting features from work order data, prompt information containing hierarchical structure information is constructed. Joint features and domain knowledge are used to adjust the parameters of a multi-level classification model, and low-rank adaptive and prefix tuning methods are used for model fine-tuning.
It improves the accuracy and adaptability of multi-level classification models, enabling them to accurately capture complex semantic features, dynamically respond to policy changes, and output multi-level classification results.
Smart Images
Figure CN120994834A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of large models, in particular to a hotline work order multi-level classification model adjustment method and device, medium and equipment. BACKGROUND
[0002] The content of the hotline work order has semantic complexity due to user expression, intention diversity and context dependence: user expression diversity and ambiguity usually refers to colloquial expression, ambiguous reference and redundant information; multiple intentions are mixed in implicit requirements, users may propose multiple demands, and emotional agitation leads to text logic confusion; domain specificity and terminology difference; context dependence and cross-session association.
[0003] Traditional text classification technology is difficult to handle the semantic complexity and diversity of expression of work order text, so the classification effect is poor, and there is error accumulation problem in multi-level classification system, and the error of lower level classification will be amplified with the deepening of level. In addition, work order classification needs to consider the field knowledge and dynamic policy changes, and traditional models are difficult to capture fine-grained semantic differences in multi-level classification scenarios. Based on large model, the classification of work order can effectively solve the predicament of traditional text classification technology. The semantic understanding ability formed by relying on massive data training of large model can accurately capture the complex semantic features in work order text, eliminate classification standard deviation, for example, through the context of the unified label attribution of ambiguous expressions such as "manhole cover damage", avoid the phenomenon of "different labels for the same problem". At the same time, the large model can dynamically update the knowledge base and respond to policy changes in real time to ensure that the classification standard keeps pace with the times. In addition, the large model can convert unstructured data into structured labels, provide strong support for quantitative analysis of hot issues of people's livelihood, help the government make accurate decisions, improve governance efficiency, and truly realize data-driven fine management, reduce the burden of grassroots at the same time, and improve service efficiency and quality.
[0004] However, the general large model in the classification method based on the general large model adopts single work order information in training, which leads to poor classification effect of the model obtained by training. SUMMARY
[0005] In view of at least one of the above technical problems, the present application embodiment provides a hotline work order multi-level classification model adjustment method and device, medium and equipment.
[0006] According to a first aspect, the method for adjusting a multi-level classification model of a hotline work order provided by an embodiment of the present application comprises: pre-processing first work order data and performing labeling processing on the pre-processed first work order data; performing feature extraction on the labeled first work order data, corresponding user attribute data and corresponding historical work order data to obtain joint features corresponding to the first work order data; constructing prompt information containing hierarchical structure information; and adjusting parameters of a multi-level classification model of a hotline work order according to the joint features and the prompt information to obtain an adjusted multi-level classification model of the hotline work order.
[0007] In one embodiment, the pre-processing of the first work order data comprises: desensitizing processing of private information in the first work order data; wherein the desensitizing processing process comprises at least one of the following: generalization processing of address information, one-way encryption processing of names and contact information, and deletion of ID numbers and mobile phone numbers; and cleaning processing of the desensitized first work order data.
[0008] In one embodiment, the multi-level classification comprises first-level classification, second-level classification and third-level classification; and the multi-level classification model of the hotline work order comprises a first classification head structure corresponding to the first-level classification, a second classification head structure corresponding to the second-level classification and a third classification head structure corresponding to the third-level classification.
[0009] In one embodiment, the first classification head structure comprises a full connection layer, the second classification head structure comprises a conditional full connection layer, and the third classification head structure comprises a gated attention classifier.
[0010] In one embodiment, the construction of the prompt information containing hierarchical structure information comprises: constructing classification requirements corresponding to the first-level classification, the second-level classification and the third-level classification to generate the prompt information; and capturing prompt features of different classification levels from the prompt information by using a hierarchical related attention module; and accordingly, the parameter adjustment of the multi-level classification model of the hotline work order according to the joint features and the prompt information to obtain the adjusted multi-level classification model of the hotline work order comprises: adjusting parameters of the multi-level classification model of the hotline work order according to the joint features and the prompt features of different classification levels to obtain the adjusted multi-level classification model of the hotline work order.
[0011] In one embodiment, the method further comprises: retrieving relevant knowledge of the first work order data from a domain knowledge base; and accordingly, the parameter adjustment of the multi-level classification model of the hotline work order according to the joint features and the prompt information to obtain the adjusted multi-level classification model of the hotline work order comprises: adjusting parameters of the multi-level classification model of the hotline work order according to the joint features, the prompt information and the relevant knowledge to obtain the adjusted multi-level classification model of the hotline work order.
[0012] In one embodiment, the parameter adjustment on the hotline ticket multi-level classification model comprises: performing parameter adjustment on the hotline ticket multi-level classification model by using a low-rank self-adaption and prefix optimization method.
[0013] According to a second aspect, an embodiment of the present application provides a hotline ticket multi-level classification model adjustment device, comprising: a preprocessing module configured to preprocess first ticket data and perform annotation processing on the preprocessed first ticket data; a feature extraction module configured to perform feature extraction on the annotated first ticket data, corresponding user attribute data, and corresponding historical ticket data to obtain joint features corresponding to the first ticket data; a prompt construction module configured to construct prompt information containing hierarchical structure information; and a model adjustment module configured to perform parameter adjustment on a hotline ticket multi-level classification model according to the joint features and the prompt information to obtain an adjusted hotline ticket multi-level classification model.
[0014] In one embodiment, the preprocessing module is specifically configured to: perform desensitization processing on private information in the first ticket data; wherein the desensitization processing process comprises at least one of the following: generalization processing on address information, one-way encryption processing on names and contact information, and deletion of ID numbers and mobile phone numbers; and perform cleaning processing on the desensitization-processed first ticket data.
[0015] In one embodiment, the multi-level classification comprises primary classification, secondary classification, and tertiary classification; and the hotline ticket multi-level classification model comprises a first classification head structure corresponding to the primary classification, a second classification head structure corresponding to the secondary classification, and a third classification head structure corresponding to the tertiary classification.
[0016] In one embodiment, the first classification head structure comprises a full connection layer, the second classification head structure comprises a conditional full connection layer, and the third classification head structure comprises a gated attention classifier.
[0017] In one embodiment, the prompt construction module is specifically configured to: construct classification requirements corresponding to the primary classification, the secondary classification, and the tertiary classification to generate the prompt information; and capture prompt features of different classification levels from the prompt information by using a hierarchical related attention module; and correspondingly, the model adjustment module is specifically configured to: perform parameter adjustment on the hotline ticket multi-level classification model according to the joint features and the prompt features of different classification levels to obtain an adjusted hotline ticket multi-level classification model.
[0018] In one embodiment, the device further comprises:
[0019] a knowledge retrieval module configured to retrieve relevant knowledge of the first ticket data from a domain knowledge base;
[0020] Correspondingly, the model adjustment module is specifically configured to: perform parameter adjustment on the hotline work order multi-level classification model according to the joint feature, the prompt information, and the related knowledge, to obtain an adjusted hotline work order multi-level classification model.
[0021] In one embodiment, the model adjustment module is specifically configured to: perform parameter adjustment on the hotline work order multi-level classification model by using a low-rank self-adaptive and prefix optimization method.
[0022] According to a third aspect, an embodiment of the present application provides a computer readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method provided in the first aspect.
[0023] According to a fourth aspect, an embodiment of the present application provides a computing device including a memory and a processor, the memory having executable code stored therein, and the processor, when executing the executable code, implements the method provided in the first aspect.
[0024] The hotline work order multi-level classification model adjustment method and device, medium, and equipment provided by the embodiments of the present application can extract features from the labeled first work order data, corresponding user attribute data, and corresponding historical work order data, to obtain joint features corresponding to the first work order data, so that multi-granularity work order feature information can be obtained, which is conducive to improving the model adjustment quality. Moreover, prompt information containing hierarchical structure information is constructed, and model adjustment is performed based on the prompt information, which can improve the effect of model adjustment. The prompt information includes multi-level classification prompt information, so that the model trained can output multi-level classification results when applied. BRIEF DESCRIPTION OF DRAWINGS
[0025] Figure 1 FIG. 1 is a flowchart of a hotline work order multi-level classification model adjustment method according to an embodiment of the present application;
[0026] Figure 2 FIG. 2 is a structural block diagram of a hotline work order multi-level classification model adjustment device according to an embodiment of the present application. DETAILED DESCRIPTION
[0027] In a first aspect, an embodiment of the present application provides a hotline work order multi-level classification model adjustment method, referring to Figure 1 The method includes the following steps:
[0028] S110, pre-processing the first work order data and labeling the pre-processed first work order data.
[0029] In one embodiment, the pre-processing of the first work order data in S110 can include:
[0030] 1. Desensitizing the privacy information in the first work order data; wherein the desensitization process includes at least one of the following: generalization processing for address information, one-way encryption processing for name and contact information, and deleting ID number and mobile phone number.
[0031] Among them, the ID number and mobile phone number can be matched by regular expression.
[0032] Among them, the name and contact information can be one-way encrypted by using a hash algorithm.
[0033] Among them, the address is generalized, for example, using XX City, XX District, XX Road to replace the detailed address.
[0034] 2. Cleaning the first work order data after desensitization.
[0035] Among them, the cleaning process can include removing duplicate text, noise symbols, invalid content, etc.
[0036] As can be seen, the above method can obtain high-quality first work order data, which helps to improve the effect of model training.
[0037] Among them, the labeling process, for example, a total of 10,000 labeled data, involving three-level classification labels. Since the training data requires high-quality label information, it is necessary to accurately align more than 1,500 categories. Data labeling has passed multiple rounds of quality inspection and label correction, and the double-blind consistency rate reaches 98%, and the quality inspection accuracy rate is 99%.
[0038] S120, extracting features from the labeled first work order data, corresponding user attribute data and corresponding historical work order data to obtain joint features corresponding to the first work order data.
[0039] It can be understood that the previous work order classification mostly only classifies based on the content of the work order to be classified, and the classification information is single, which is not conducive to feature extraction of the model, so the classification accuracy is not high. The embodiment adopts joint features, that is, the first work order data, corresponding user attribute data and corresponding historical work order data are jointly encoded to obtain corresponding joint features, thereby obtaining multi-granularity work order feature information.
[0040] S130, constructing prompt information containing hierarchical structure information.
[0041] Exemplarily, the prompt information is: "[first-level classification] [second-level classification] [third-level classification] Please judge the most appropriate classification according to the following content: {joint features}".
[0042] S140, adjusting the hotline work order multi-level classification model according to the joint features and the prompt information to obtain an adjusted hotline work order multi-level classification model.
[0043] In one embodiment, the multi-level classification includes a first-level classification, a second-level classification and a third-level classification; the hot line ticket multi-level classification model includes a first classification head structure corresponding to the first-level classification, a second classification head structure corresponding to the second-level classification and a third classification head structure corresponding to the third-level classification.
[0044] The first classification head structure can include a full connection layer, the second classification head structure can include a conditional full connection layer, and the third classification head structure can include a gated attention classifier.
[0045] Exemplarily, the full connection layer performs first-level classification by using the following formula:
[0046]
[0047] Wherein h is an upstream feature, y1 is a probability distribution of first-level classification, W1 is a first weight, and b1 is a bias. Through the formula, the parameter sharing gradient isolation from the downstream task can be maintained, the interference of back propagation on the upstream feature is blocked, the sparse processing is performed, and L1 regularization (λ=0.01) is applied to the output layer to improve the feature decoupling ability.
[0048] Exemplarily, the conditional full connection layer realizes second-level classification by using the following formula:
[0049]
[0050] Wherein y2 is a probability distribution of second-level classification, and G is a gating network, which adopts a lightweight multi-layer perception.
[0051] Exemplarily, the gated attention classifier realizes third-level classification by using the following formula:
[0052] F feature =W1h+W2y1+W3y2
[0053]
[0054] Wherein y3 is a probability distribution of third-level classification; α i is an adjustment coefficient for dynamically adjusting the contribution of each layer by a learnable parameter; W2 is a second weight, and W3 is a third weight. By parallel computing 4 groups of attention gates (3 weights in each group), the gradient stability is maintained by residual connection; during the dynamic pruning training process, the sub-modules with a contribution degree less than 0.1 are automatically closed.
[0055] In one embodiment, the prompt information containing hierarchical structure information constructed in S130 includes:
[0056] 1, build the classification requirements corresponding to the primary classification, the secondary classification and the tertiary classification, and generate the prompt information.
[0057] 2, adopt a hierarchical attention module to capture prompt features of different classification levels from the prompt information.
[0058] That is, a hierarchical attention (HRA) module is used to dynamically capture semantic features of different classification levels:
[0059] F = HRA (P)
[0060] Where F is the semantic feature, P is the prompt information, and HRA is the hierarchical attention module.
[0061] Correspondingly, the parameter adjustment of the hotline work order multi-level classification model according to the joint feature and the prompt information in S140 to obtain the adjusted hotline work order multi-level classification model includes:
[0062] According to the joint feature and the prompt feature of different classification levels, the parameter adjustment of the hotline work order multi-level classification model is performed to obtain the adjusted hotline work order multi-level classification model.
[0063] That is, the joint feature and the prompt feature are used to fine-tune the hotline work order multi-level classification model to obtain a hotline work order multi-level classification model with better classification effect.
[0064] In one embodiment, the method provided by the present embodiment can further include:
[0065] Retrieving relevant knowledge of the first work order data from a domain knowledge base;
[0066] Wherein the domain knowledge base includes relevant domain knowledge of government hotline.
[0067] Wherein a text similarity retrieval method can be used for retrieval operation in the domain knowledge base.
[0068] Of course, after policy changes and classification standard updates, knowledge enhancement is needed to respond to changes in real time. First, encode the text information corresponding to the updated knowledge to obtain knowledge features, and add the knowledge features to the domain knowledge base.
[0069] Correspondingly, the parameter adjustment of the hotline work order multi-level classification model according to the joint feature and the prompt information to obtain the adjusted hotline work order multi-level classification model includes:
[0070] According to the joint feature, the prompt information and the relevant knowledge, the parameter adjustment of the hotline work order multi-level classification model is performed to obtain the adjusted hotline work order multi-level classification model.
[0071] Since the joint features, prompt information and domain-related knowledge are considered when the model is adjusted, the effect of model adjustment can be improved.
[0072] In one embodiment, the parameter adjustment of the hotline ticket multi-level classification model comprises:
[0073] The parameter adjustment of the hotline ticket multi-level classification model is performed by using a low-rank adaptive and prefix tuning method.
[0074] The low-rank adaptive method is LoRA (Low-Rank Adaptation).
[0075] The prefix tuning method is Prefix-Tuning.
[0076] It can be understood that the above method can improve the accuracy and efficiency of model adjustment.
[0077] After the adjustment is completed, the model obtained by the adjustment can be used to perform multi-level classification on the to-be-classified ticket.
[0078] As described above, the joint features corresponding to the first ticket data are obtained by performing feature extraction on the labeled first ticket data, corresponding user attribute data and corresponding historical ticket data, and it can be seen that multi-granularity ticket feature information can be obtained, which is beneficial to improve the model adjustment quality. Moreover, the prompt information containing hierarchical structure information is constructed, and the model is adjusted based on the prompt information, which can improve the effect of model adjustment. The prompt information includes multi-level classification prompt information, so that the model trained can output multi-level classification results when applied. In addition, the related domain knowledge retrieved based on single ticket data is also used in model adjustment, which can improve the combination of the model and the domain knowledge, and further improve the model training effect.
[0079] In a second aspect, a hotline ticket multi-level classification model adjustment device is provided, comprising:
[0080] The preprocessing module is configured to preprocess the first ticket data and label the preprocessed first ticket data.
[0081] The feature extraction module is configured to perform feature extraction on the labeled first ticket data, corresponding user attribute data and corresponding historical ticket data to obtain joint features corresponding to the first ticket data.
[0082] The prompt construction module is configured to construct prompt information containing hierarchical structure information.
[0083] The model adjustment module is configured to adjust parameters of the hotline work order multi-level classification model according to the joint feature and the prompt information, and obtain an adjusted hotline work order multi-level classification model.
[0084] In one embodiment, the preprocessing module is specifically configured to perform desensitization processing on the private information in the first work order data, and the desensitization processing process includes at least one of the following: generalization processing on address information, one-way encryption processing on names and contact information, and deletion of ID numbers and mobile phone numbers; and the first work order data after the desensitization processing is subjected to cleaning processing.
[0085] In one embodiment, the multi-level classification includes first-level classification, second-level classification and third-level classification; and the hotline work order multi-level classification model includes a first classification head structure corresponding to the first-level classification, a second classification head structure corresponding to the second-level classification, and a third classification head structure corresponding to the third-level classification.
[0086] In one embodiment, the first classification head structure includes a full connection layer, the second classification head structure includes a conditional full connection layer, and the third classification head structure includes a gated attention classifier.
[0087] In one embodiment, the prompt construction module is specifically configured to construct classification requirements corresponding to the first-level classification, the second-level classification and the third-level classification, and generate the prompt information; and a hierarchical related attention module is used to capture prompt features of different classification levels from the prompt information; and correspondingly, the model adjustment module is specifically configured to adjust parameters of the hotline work order multi-level classification model according to the joint feature and the prompt features of different classification levels, and obtain an adjusted hotline work order multi-level classification model.
[0088] In one embodiment, the device further includes:
[0089] The knowledge retrieval module is configured to retrieve relevant knowledge of the first work order data from a domain knowledge base.
[0090] Correspondingly, the model adjustment module is specifically configured to adjust parameters of the hotline work order multi-level classification model according to the joint feature, the prompt information and the relevant knowledge, and obtain an adjusted hotline work order multi-level classification model.
[0091] In one embodiment, the model adjustment module is specifically configured to adjust parameters of the hotline work order multi-level classification model by using a low-rank adaptive and prefix optimization method.
[0092] It can be understood that the explanations, specific embodiments, beneficial effects, examples and the like of the device provided in the embodiments of the present application can refer to the corresponding parts in the method provided in the first aspect, and will not be repeated here.
[0093] In a third aspect, an embodiment of the present application provides a computer readable medium, having stored thereon computer instructions, which when executed by a processor, cause the processor to perform the method provided in the first aspect.
[0094] Specifically, a system or an apparatus equipped with a storage medium on which a software program code for realizing the functions of any of the above embodiments is stored, and a computer (or CPU or MPU) of the system or the apparatus can be provided to read out and execute the program code stored in the storage medium.
[0095] In this case, the program code read out from the storage medium itself realizes the functions of any of the above embodiments, and therefore the program code and the storage medium storing the program code constitute a part of the present application.
[0096] Embodiments of the storage medium for providing the program code include a floppy disk, a hard disk, a magneto-optical disk, an optical disk (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), a magnetic tape, a non-volatile memory card, and a ROM. Alternatively, the program code can be downloaded from a server computer via a communication network.
[0097] Further, it should be clear that not only the program code read out by the computer, but also the operating system or the like operating on the computer based on the instructions of the program code can perform part or all of the actual operations to realize the functions of any of the above embodiments.
[0098] Further, it can be understood that the program code read out from the storage medium is written into a memory provided in an expansion board inserted into the computer or a memory provided in an expansion module connected to the computer, and then part or all of the actual operations are performed based on the instructions of the program code by a CPU or the like installed on the expansion board or the expansion module to realize the functions of any of the above embodiments.
[0099] It can be understood that the explanation of the content, the specific embodiments, the beneficial effects, the examples, and the like of the computer readable medium provided by the embodiments of the present application can refer to the corresponding parts in the method provided in the first aspect, and will not be described here.
[0100] In a fourth aspect, an embodiment of the present application provides a computing device, including a memory and a processor, the memory having stored thereon executable code which, when executed by the processor, implements the method in any of the embodiments of the specification.
[0101] It can be understood that the explanation, specific implementation, advantages, examples and the like of the content related to the computing device provided by the embodiments of the present application can refer to the corresponding parts in the method provided by the first aspect, and will not be described here.
[0102] Each of the embodiments in the specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each of the embodiments mainly explains the difference from other embodiments. Especially, for the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can refer to the part of the method embodiments.
[0103] Those skilled in the art can realize that the functions described in the above one or more examples can be realized by hardware, software, a plug-in or any combination thereof. When realized by software, the functions can be stored in a computer readable medium or transmitted as one or more instructions or codes on the computer readable medium.
[0104] The above specific embodiments further explain the purpose, technical solutions and advantages of the present application. It should be understood that the above description is only a specific embodiment of the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement and the like made on the basis of the technical solutions of the present application shall be included in the protection scope of the present application.
Claims
1. A method for adjusting a multi-level classification model for hotline work orders, characterized in that, include: The data of the first work order is preprocessed, and the preprocessed data of the first work order is labeled. Feature extraction is performed on the labeled first work order data, the corresponding user attribute data, and the corresponding historical work order data to obtain the joint features corresponding to the first work order data. Construct prompts that include hierarchical structure information; Based on the joint features and the prompt information, the parameters of the hotline work order multi-level classification model are adjusted to obtain the adjusted hotline work order multi-level classification model.
2. The method according to claim 1, characterized in that, The preprocessing of the first work order data includes: The privacy information in the first work order data is anonymized; the anonymization process includes at least one of the following: generalizing the address information, unidirectionally encrypting the name and contact information, and deleting the ID card number and mobile phone number; The data from the first work order after the anonymization process is cleaned.
3. The method according to claim 1, characterized in that, The multi-level classification includes a primary classification, a secondary classification, and a tertiary classification; the hotline work order multi-level classification model includes a first classification header structure corresponding to the primary classification, a second classification header structure corresponding to the secondary classification, and a third classification header structure corresponding to the tertiary classification.
4. The method according to claim 3, characterized in that, The first classification head structure includes a fully connected layer, the second classification head structure includes a conditionally fully connected layer, and the third classification head structure includes a gated attention classifier.
5. The method according to claim 3, characterized in that, The prompt information that includes hierarchical structure information includes: Construct the classification requirements corresponding to the first-level category, the second-level category, and the third-level category, and generate the prompt information; A hierarchical relevance attention module is used to capture prompt features of different classification levels from the prompt information; Accordingly, the step of adjusting the parameters of the hotline work order multi-level classification model based on the joint features and the prompt information to obtain the adjusted hotline work order multi-level classification model includes: Based on the joint features and the prompt features of different classification levels, the parameters of the hotline work order multi-level classification model are adjusted to obtain the adjusted hotline work order multi-level classification model.
6. The method according to claim 1, characterized in that, Also includes: Retrieve relevant knowledge about the first work order data from the domain knowledge base; Correspondingly, the parameter adjustment of the hotline work order multi-level classification model based on the joint features and the prompt information to obtain the adjusted hotline work order multi-level classification model includes: Based on the joint features, the prompt information, and the relevant knowledge, the parameters of the hotline work order multi-level classification model are adjusted to obtain the adjusted hotline work order multi-level classification model.
7. The method according to claim 1, characterized in that, The parameter adjustment of the multi-level classification model for hotline work orders includes: The parameters of the multi-level classification model for hotline work orders are adjusted using low-rank adaptive and prefix tuning methods.
8. A device for adjusting a multi-level classification model for hotline work orders, characterized in that, include: The preprocessing module is used to preprocess the first work order data and to annotate the preprocessed first work order data. The feature extraction module is used to extract features from the labeled first work order data, the corresponding user attribute data, and the corresponding historical work order data to obtain the joint features corresponding to the first work order data. The prompt building module is used to construct prompt messages that include hierarchical structure information; The model adjustment module is used to adjust the parameters of the hotline work order multi-level classification model according to the joint features and the prompt information, so as to obtain the adjusted hotline work order multi-level classification model.
9. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed in a computer, causes the computer to perform the method described in any one of claims 1 to 7.
10. A computing device, characterized in that, The method includes a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method described in any one of claims 1 to 7.