Public service hotline work order identification method and system based on bert-base-history model

Abstract

The invention relates to the technical field of text classification, and discloses a public service hotline work order identification method and system based on a bert-based-history model, and the method comprises the steps: obtaining a hotline work order data set, carrying out the feature cleaning and standardized data labeling, and dividing the data set after the standardized data labeling into a training set, a verification set and a test set; constructing a hot line work order identification model based on a bert-based-chinese model, training the hot line work order identification model by using a training set and a verification set, and optimizing hyper-parameters of the hot line work order identification model by using Optuna in the training process; and inputting the test set into the trained hotline work order identification model to obtain a hotline work order identification result. According to the invention, accurate identification of various hotline work orders can be realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of text classification, in particular to a public service hotline work order recognition method and system based on a bert-base-chinese model. BACKGROUND

[0002] In the field of public services, the number of hotline work orders is huge, covering information consultation, complaint reporting and suggestion. The existing hotline work order processing mainly relies on manual classification and simple rule matching, which is low in efficiency and prone to errors. In order to improve the processing efficiency and accuracy of hotline work orders, the method of using text classification is used. However, the traditional text classification method only captures word frequency information and cannot handle the ambiguity and context dependence of Chinese semantics, especially when facing hotline work orders, the accuracy cannot be guaranteed.

[0003] In order to better complete Chinese semantic processing, dynamic models (such as Embeddings from Language Models (ELMo)) and pre-training language models based on Transformer architecture (such as Bidirectional Encoder Representations from Transformers (BERT) and its variants RoBERTa, ALBERT) have appeared in the prior art. These bidirectional encoding mechanisms can consider context information, thereby improving semantic understanding. However, these methods cannot effectively handle the flexibility of Chinese grammar and semantic ambiguity when facing hotline work orders, which are large in quantity and have unique language grammar in public services, due to the lack of standardized labels, resulting in low classification accuracy (usually less than 85%) and weak generalization ability. SUMMARY

[0004] Therefore, the technical problem to be solved by the present application is to overcome the deficiencies in the prior art and provide a public service hotline work order recognition method and system based on a bert-base-chinese model, which can accurately recognize various hotline work orders.

[0005] To solve the above technical problems, the present application provides a public service hotline work order recognition method based on a bert-base-chinese model, comprising: obtaining a hotline work order dataset and performing feature cleaning and standardized data labeling, and dividing the dataset after standardized data labeling into a training set, a validation set and a test set; A hotline work order recognition model based on the bert-base-chinese model is constructed. The hotline work order recognition model is trained using a training set and a validation set. During the training process, Optuna is used to optimize the hyperparameters of the hotline work order recognition model. Input the test set into the trained hotline work order recognition model to obtain the hotline work order recognition results.

[0006] Furthermore, the feature cleaning specifically includes: Basic text cleaning is performed to remove non-textual parts. Mapping rules from spoken to written language in the hotline work order domain are constructed. The jieba word segmentation library is used to integrate the proper nouns existing in the hotline work order domain.

[0007] Furthermore, the standardized data annotation specifically refers to: Labeling is performed on the feature-cleaned dataset based on industry business classification standards and existing labeled Chinese text datasets. During the labeling process, it is ensured that the labels are strongly correlated with the text content, the labeled content is pure Chinese, and the number of samples in each category is balanced.

[0008] Furthermore, the hotline work order recognition model includes a bert-base-chinese word segmenter and a pre-trained model. The bert-base-chinese word segmenter converts the input Chinese text into a tokenID acceptable to the hotline work order recognition model, and the pre-trained model extracts the semantic features of the input Chinese text through a Transformer structure.

[0009] Furthermore, the pre-trained model extracts semantic features of the input Chinese text using a Transformer structure as follows: Use a tokenizer to split the input Chinese text into subwords, and add [CLS] at the beginning and [SEP] at the end of the split Chinese text to unify the length of the Chinese text.

[0010] Furthermore, when using Optuna to optimize the hyperparameters of the hotline work order identification model, the range of hyperparameter values ​​and tuning logic are designed based on Optuna's search space.

[0011] Furthermore, the hyperparameters include learning rate, weight decay, batch size, gradient accumulation steps, number of frozen layers, and loss function. gamma Value and preheating step ratio.

[0012] Furthermore, the specific range of values ​​for the hyperparameters is as follows: The search range for the learning rate is loguniform(1e-6, 5e-5), where loguniform represents a log-uniform distribution. The search range for the weight decay is uniform(0.001, 0.1), where uniform represents a continuous uniform distribution; The search range for the batch size is categorical([8, 16, 32]), where categorical represents the categorical distribution; The search range for the gradient accumulation steps is categorical([1, 2, 4]). The search range for the number of frozen layers is intuniform(6, 10), where intuniform represents generating uniform integers; The loss function used is Focal Loss. gamma The search range for the value is uniform(1.0, 3.0); The search range for the preheating step ratio is uniform(0.05, 0.2).

[0013] Furthermore, the tuning logic for the hyperparameters is as follows: The optimization logic for the learning rate is that fine-tuning the learning rate to within the range of 1e-5 to 3e-5 is optimal, and a logarithmic distribution is used to cover smaller or larger values ​​of the learning rate to prevent omissions. The optimization logic for the weight decay is to control overfitting and use regularization intensity when covering the range. The optimization logic for batch size is to avoid non-integer values ​​when limited by video memory; The optimization logic for the gradient accumulation step count is to simulate a large batch when there is insufficient video memory. The optimization logic for the number of frozen layers is to explore a range of 6 to 10 frozen layers to adapt to the data volume. The Focal Loss gamma The optimization logic for the value is as follows gamma =2 is the default value; explore values ​​from 1 to 3. gamma The values ​​are selected to optimize the weights of hard-to-classify samples; The optimization logic for the preheating step ratio is that the learning rate preheating accounts for 5% to 20% of the total number of steps, in order to avoid oscillations caused by an excessively high initial learning rate.

[0014] This invention also provides a public service hotline work order identification system based on the BERT-base-chinese model, comprising: The data acquisition and preprocessing module acquires the hotline work order dataset and performs feature cleaning and standardized data labeling. The standardized data labeling dataset is then divided into training set, validation set, and test set. The module for building a hotline work order recognition model constructs a hotline work order recognition model based on the BERT-Base-Chinese model. The model training module uses the training set and validation set to train the hotline work order recognition model. During the training process, Optuna is used to optimize the hyperparameters of the hotline work order recognition model. The hotline work order recognition module takes the test set as input to the trained hotline work order recognition model and obtains the hotline work order recognition results.

[0015] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: This invention optimizes the hyperparameters of the bert-base-chinese model using Optuna. Based on this, after standardizing and labeling hotline work orders, the optimized bert-base-chinese model is used to identify hotline work orders. This effectively integrates contextual information and achieves accurate identification of various types of hotline work orders. Attached Figure Description

[0016] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart of a method in a preferred embodiment of the present invention.

[0017] Figure 2 This is a process diagram of the simulation experiment in a preferred embodiment of the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0019] Reference Figure 1 As shown, this invention discloses a method for identifying public service hotline work orders based on the BERT-base-chinese model, comprising the following steps: S1: Environment Configuration. Configure the Python environment (install libraries such as transformers and torch), prepare a labeled Chinese text dataset, and configure the bert-base-chinese word segmenter and pre-trained model.

[0020] S2: Acquire and preprocess hotline work order data. Acquire the hotline work order dataset and perform feature cleaning and standardized data labeling. Divide the standardized data labeling dataset into training set, validation set, and test set.

[0021] S2-1: Obtain the hotline work order dataset. In this embodiment, the dataset obtained is a one-year hotline work order dataset (approximately 300,000 entries) provided by a public service center in a certain city, covering three categories: information consultation, complaints and reports, and opinions and suggestions.

[0022] S2-2: Feature cleaning is performed to address the characteristics of hotline work orders, such as overly colloquial expressions, dense and specialized domain terminology, and fragmented information. Specifically, feature cleaning involves: S2-2-1: The first step is to perform basic text cleaning to remove non-text parts, including punctuation marks and numbers, which are not very meaningful for subsequent text mining.

[0023] S2-2-2: Optimize the standardization of colloquial expressions and construct mapping rules from colloquial to written language in the hotline work order field, such as "breakdown" corresponding to "vehicle malfunction" and "tow" corresponding to "towing service". Reduce colloquial interference through mapping rules.

[0024] S2-2-3: In terms of word segmentation, the jieba word segmentation library (which supports custom dictionaries and stop word lists) is used first to integrate proper nouns in the hotline work order field, thereby optimizing and improving word segmentation accuracy.

[0025] S2-3: The standardized data annotation specifically refers to: Based on industry business classification standards and existing labeled Chinese text datasets, the feature-cleaned dataset is labeled. During the labeling process, it is ensured that the labels are strongly related to the text content (e.g., the reason for labeling work orders for "complaints and reports" needs to be clearly stated), the labeled content is pure Chinese (or contains a small number of punctuation marks and numbers, without special symbols), and the number of samples in each category is as balanced as possible (e.g., avoid the proportion of "complaints and reports" being much higher than that of "information consultation" to reduce model bias).

[0026] Labeling based on industry business classification standards ensures more standardized data classification. In this embodiment, labels are applied according to the JT / T1019.4-2025 standard (e.g., label 0: congestion information, label 1: detour). The format of the labeled Chinese text dataset is [text, label] (text is the work order content, label is the label). Examples of label annotation (CSV format) are shown in Table 1.

[0027] Table 1. Example of Label Marking

[0028] S2-4: In this embodiment, the standardized and data-labeled dataset is divided into three parts by stratified random sampling: a training set (80%) for training model parameters, a validation set (10%) for evaluating the model parameters during each training, and a test set (10%) as an independent evaluation set that does not participate in training throughout the process.

[0029] S3: Construct a hotline work order recognition model based on the bert-base-chinese model.

[0030] bert-base-chinese is a BERT variant developed by Google for Chinese. It can learn Chinese grammar and context dependencies through pre-training on large-scale Chinese corpora (such as encyclopedias, news, and social media). Its bidirectional Transformer architecture (12-layer encoder, with about 110 million parameters) solves the problem of insufficient context modeling in unidirectional models and improves classification accuracy. However, there is no precedent for classifying work orders based on the bert-base-chinese model in current work, which limits the work order transfer efficiency and big data analysis capabilities. In this embodiment, a hotline work order recognition model is constructed based on the bert-base-chinese model, and on this basis, the model is fine-tuned using Optuna to adapt to the text classification task of hotline work orders.

[0031] The hotline work order recognition model includes a bert-base-chinese tokenizer and a pre-trained model. In this embodiment, the bert-base-chinese model is loaded using the transformers library in a Python environment. The bert-base-chinese tokenizer converts the input Chinese text into token IDs acceptable to the hotline work order recognition model (for example, converting "High-speed traffic jam" into [101, 800, 4919, 102], where 101 is [CLS] and 102 is [SEP]). The pre-trained model extracts the semantic features of the input Chinese text through the Transformer structure.

[0032] The pre-trained model extracts the semantic features of the input Chinese text through the Transformer structure specifically as follows: First step, tokenization, using the tokenizer to split the input Chinese text into subwords (for example, splitting "Apply for the qualification of a network car driver" into ["Apply", "Network car", "Driver qualification"]); Second step, adding special tokens, adding [CLS] (global representation for classification tasks) at the beginning and [SEP] (separating different sentences) at the end of the split Chinese text; Third step, truncating / padding, unifying the length of the Chinese text to a fixed value (such as 128 or 512, truncating if it exceeds and padding with [PAD] if it is insufficient) to ensure the consistency of batch processing.

[0033] S4: Train the hotline work order recognition model using the training set and validation set. During the training process, use Optuna to optimize the hyperparameters of the hotline work order recognition model.

[0034] Optuna is a lightweight and flexible automatic hyperparameter optimization framework. It dynamically explores the hyperparameter space through Bayesian optimization, finding the optimal parameter combination more efficiently than grid search or random search. To improve the efficiency and accuracy of hotline work order processing, and considering the characteristics of the hotline work order classification task, this paper optimizes the hyperparameters affecting model convergence and classification accuracy. When optimizing the hyperparameters of the hotline work order recognition model using Optuna, the range of hyperparameter values ​​and tuning logic are designed based on Optuna's search space. The hyperparameters include learning rate, weight decay, batch size, gradient accumulation steps, number of freeze layers, and loss function. gamma Values ​​and preheating step ratios. Table 2 shows the value range and optimization logic for each parameter.

[0035] Table 2. Value range and optimization logic for each parameter.

[0036] As shown in Table 2, the tuning logic for the range of hyperparameter values ​​is as follows: The search range for the learning rate is loguniform(1e-6, 5e-5), where loguniform represents a logarithmic uniform distribution. The optimization logic is to fine-tune the learning rate to be optimal between 1e-5 and 3e-5, and to use a logarithmic distribution to cover smaller or larger values ​​of the learning rate to prevent omissions.

[0037] The search range for the weight decay is uniform(0.001, 0.1), where uniform represents a continuous and uniform distribution; the optimization logic is to control overfitting, and the regularization strength is used when covering the range.

[0038] The search range for the batch size is categorical([8, 16, 32]), where categorical represents the category distribution; the optimization logic is to avoid non-integer values ​​when limited by video memory by using the category type.

[0039] The search range for the gradient accumulation steps is categorical([1, 2, 4]), and the optimization logic is to use gradient accumulation to simulate a large batch when there is insufficient video memory.

[0040] The search range for the number of frozen layers is intuniform(6, 10), where intuniform represents generating uniform integers. The optimization logic is to explore the range of 6 to 10 frozen layers to adapt to the amount of data.

[0041] The loss function used is Focal Loss. gamma The search range for the value is uniform(1.0, 3.0); the tuning logic is as follows: gamma =2 is the default value; explore values ​​from 1 to 3. gamma The values ​​are selected to optimize the weights of difficult-to-classify samples.

[0042] The search range for the preheating step ratio is uniform(0.05, 0.2), and the optimization logic is that the learning rate preheating accounts for 5% to 20% of the total number of steps, in order to avoid oscillation caused by an excessively high initial learning rate.

[0043] In this embodiment, num_labels=3. Training parameters: num_train_epochs=4, learning_rate=2e-5, batch_size=8. The validation set accuracy after training reaches 87.5%.

[0044] The model that performs best during training will be used as the hotline work order recognition model after training is completed.

[0045] S5: Input the test set into the trained hotline work order recognition model to obtain the hotline work order recognition result.

[0046] The test set is converted to the same model input format as during training. This is then input into the trained hotline work order recognition model to obtain its raw prediction results. The maximum value of the logits vector in the raw prediction results is taken as the predicted category of the text. The model predicts numeric IDs, which need to be mapped back to meaningful label names (such as "traffic jam" or "detour"). The mapping relationship uses the labels defined during training.

[0047] This invention also discloses a public service hotline work order identification system based on the BERT-base-chinese model, comprising: The data acquisition and preprocessing module acquires the hotline work order dataset and performs feature cleaning and standardized data labeling. The standardized data labeling dataset is then divided into training set, validation set, and test set. The module for building a hotline work order recognition model constructs a hotline work order recognition model based on the BERT-Base-Chinese model. The model training module uses the training set and validation set to train the hotline work order recognition model. During the training process, Optuna is used to optimize the hyperparameters of the hotline work order recognition model. The hotline work order recognition module takes the test set as input to the trained hotline work order recognition model and obtains the hotline work order recognition results.

[0048] The present invention also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a public service hotline work order identification method based on the bert-base-chinese model.

[0049] The present invention also discloses an apparatus 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 implement a public service hotline work order identification method based on the bert-base-chinese model.

[0050] Compared with the prior art, the advantages of the present invention are: 1. By optimizing the hyperparameters of the bert-base-chinese model using Optuna, and then standardizing and labeling hotline work orders, the optimized bert-base-chinese model can be used to identify hotline work orders. This effectively integrates contextual information, thereby achieving accurate identification of various types of hotline work orders.

[0051] 2. During the feature cleaning process, a standardized mapping rule for the labeling of spoken and written language in the hotline work order domain is constructed to map unstructured text (such as "highway accident caused congestion") to a unified label (label 0: traffic jam). Subsequently, standardized data annotation is performed on the hotline work orders. By using standardized labels, classification consistency is ensured, thereby improving the model's generalization ability when identifying different categories of hotline work orders.

[0052] 3. The bert-base-chinese model used has a bidirectional architecture, which can resolve contextual dependencies (such as the ambiguity of "detour" in different contexts).

[0053] 4. By freezing the parameters of the pre-trained layers to preserve language knowledge and only updating the newly added fully connected layers (classification heads) to preserve the pre-trained language knowledge, the model is more suitable for recognizing Chinese hotline work orders.

[0054] 5. Use a small learning rate (e.g., 2e-5) and weight decay (e.g., 0.01) to prevent overfitting and improve the training effect of the model.

[0055] 6. Stratified random sampling ensures class balance among the training, validation, and test sets, reducing model bias.

[0056] To further demonstrate the beneficial effects of the present invention, this embodiment uses the method of the present invention, Support Vector Machine (SVM), Long Short-Term Memory (LSTM) network, and TextCNN to conduct comparative simulation experiments on hotline work order recognition. The simulation process is as follows: Figure 2 As shown in Table 3, evaluation metrics on the test set were calculated using the evaluate library. The evaluation metrics included accuracy and F1 score. The probability of misclassification of high-frequency categories (e.g., "detour" being identified as "traffic jam") was analyzed using a confusion matrix. The recognition results of each method are shown in Table 3.

[0057] Table 3 Results of hotline work order identification using different methods

[0058] As can be seen from Table 3, the accuracy and F1 score of the present invention are higher than those of the existing methods, and the probability of misclassification of high-frequency categories is significantly reduced compared with the existing methods, proving that the present invention can achieve accurate identification of various hotline work orders.

[0059] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0060] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0062] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A public service hotline ticket identification method based on a bert-base-chinese model, characterized in that, The method comprises the following steps: obtaining a hot line ticket dataset and performing feature cleaning and standardizing data labeling, and dividing the standardized data labeled dataset into a training set, a validation set and a test set; constructing a hot line ticket recognition model based on a bert-base-chinese model, training the hot line ticket recognition model using the training set and the validation set, and optimizing the hyperparameters of the hot line ticket recognition model using Optuna during the training process; inputting the test set into the trained hot line ticket recognition model to obtain hot line ticket recognition results.

2. The public service hotline ticket identification method based on the bert-base-chinese model according to claim 1, characterized in that, The feature cleaning specifically comprises: performing basic text cleaning to remove non-text parts, constructing a mapping rule from hot line ticket domain colloquial to written language, and integrating existing special terms in the hot line ticket domain using the jieba word segmentation library.

3. The public service hotline ticket identification method based on the bert-base-chinese model according to claim 1, characterized in that, The standardization data labeling specifically comprises: labeling the feature cleaned dataset based on industry business classification standards and existing labeled Chinese text datasets, and ensuring that the label is strongly related to the text content, the labeled content is pure Chinese, and the number of samples in each category is balanced during the labeling process.

4. The public service hotline ticket identification method based on the bert-base-chinese model according to claim 1, characterized in that: The hot line ticket recognition model comprises a bert-base-chinese tokenizer and a pre-trained model, the bert-base-chinese tokenizer converts the input Chinese text into tokenID acceptable to the hot line ticket recognition model, and the pre-trained model extracts semantic features of the input Chinese text through a Transformer structure.

5. The public service hotline ticket identification method based on the bert-base-chinese model according to claim 4, characterized in that, The pre-trained model extracts semantic features of the input Chinese text through a Transformer structure specifically comprises: splitting the input Chinese text into subwords using tokenizer, adding [CLS] at the beginning of the split Chinese text and [SEP] at the end, and unifying the length of the Chinese text.

6. The public service hotline ticket identification method based on the bert-base-chinese model according to any one of claims 1-5, characterized in that: When optimizing the hyperparameters of the hot line ticket recognition model using Optuna, the value range and tuning logic of the hyperparameters are designed based on Optuna.

7. The public service hotline ticket identification method based on the bert-base-chinese model according to claim 6, characterized in that: The hyperparameters include learning rate, weight decay, batch size, gradient accumulation step number, frozen layer number, value of loss function, and preheat step number ratio. The value range of the hyperparameters specifically comprises: value of loss function, and preheat step number ratio.

8. The public service hotline ticket identification method based on the bert-base-chinese model according to claim 7, characterized in that, The search range of the learning rate is loguniform(1e-6, 5e-5), and loguniform represents a logarithmic uniform distribution; The search range of the weight decay is uniform(0.001, 0.1), and uniform represents a continuous uniform distribution; The search range of the batch size is categorical([8, 16, 32]), and categorical represents a categorical distribution; The search range of the gradient accumulation step is categorical([1, 2, 4]), The search range of the number of frozen layers is intuniform(6, 10), and intuniform represents a generated uniform integer; The search range of the warm-up step ratio is uniform(0.05, 0.2). The loss function uses Focal Loss, the value of Focal Loss is searched in the range of uniform(1.0,3.0) The tuning logic of the hyperparameters is: uniform(1.0,3.0); ​ 9. The public service hotline ticket identification method based on the bert-base-chinese model according to claim 8, characterized in that, ​ The learning rate optimization logic is that the fine-tuning learning rate is optimal between 1e-5 and 3e-5, and the smaller or larger values of the learning rate are covered with a logarithmic distribution to prevent omission; The weight decay optimization logic is that the regularization strength is used when the range is covered to control overfitting; The batch size optimization logic is that when the memory is limited, the non-integer value is avoided by using the classification type; The gradient accumulation step number optimization logic is that when the memory is insufficient, the gradient accumulation is used to simulate a large batch; The frozen layer number optimization logic is that the range of 6-10 frozen layers is explored to adapt to the data volume, The Focal Loss The preheating step number ratio optimization logic is that the learning rate preheating accounts for 5-20% of the total number of steps to avoid oscillation caused by an excessively high initial learning rate. The tuning logic for the value of Comprise: = 2 is the default value, and values of 1-3 are explored to optimize the weights of difficult-to-classify samples. A data acquisition and preprocessing module that acquires a hotline ticket dataset and performs feature cleaning and standardizes data labeling, divides the standardized data labeled dataset into a training set, a validation set and a test set; ​ A hotline ticket identification model construction module that constructs a hotline ticket identification model based on a bert-base-chinese model; 10. A public service hotline work order identification system based on a bert-base-chinese model, characterized in that, A model training module that trains the hotline ticket identification model using the training set and the validation set, and optimizes the hyperparameters of the hotline ticket identification model using Optuna during the training process; A hotline ticket identification module that inputs the test set into the trained hotline ticket identification model to obtain hotline ticket identification results. ​ ​ ​

Images