Public transaction proposal automatic classification method and system based on deep learning algorithm

Abstract

The invention discloses a public transaction proposal automatic classification method and system based on a deep learning algorithm, and relates to the technical field of deep learning. The method comprises the following steps: acquiring a public transaction proposal text, and after executing a text preliminary processing operation, constructing a classification label system and labeling a sample; bERT is adopted as a deep learning model of the architecture, and whether set constraint conditions are met or not is judged; if yes, triggering a corresponding model improvement mechanism to obtain a standard training model; otherwise, taking the current deep learning model as a standard training model; the technical key points are as follows: efficient and accurate classification of massive proposals is realized through complete data acquisition, model training and system deployment processes, and the method not only solves the problems of low efficiency and poor accuracy in a traditional classification mode, but also has good expansibility and adaptability, and can provide powerful support for intelligent management.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and specifically to a method and system for automatically classifying public affairs proposals based on a deep learning algorithm. Background Art

[0002] Deep learning is a machine learning method based on multi-layer neural networks that can automatically extract features from data and perform classification or prediction. Currently, relevant systems for public affairs proposal classification usually adopt a method that combines rule-based matching with traditional machine learning. This type of solution mainly relies on manual extraction of shallow features such as keywords and word frequency statistics (such as TF-IDF), and combines classifiers such as naive Bayes, support vector machines (SVM) or random forests for training and prediction. Its core process includes: text segmentation, feature extraction, vectorized representation, model training and classification output, but this type of method is only suitable for public affairs proposal processing scenarios with small data scale and relatively fixed classification structure.

[0003] Existing public affairs proposal classification schemes mostly rely on traditional machine learning or shallow semantic analysis technology, which has problems such as insufficient classification accuracy, poor generalization ability, and difficulty in adapting to long texts and complex semantic expressions. For example, when processing long texts such as "Suggest optimization of XX", traditional classification methods cannot effectively capture contextual semantic relationships and are easily misclassified as other categories. In addition, most solutions lack the ability to adapt to resource-constrained environments (such as edge devices), have high deployment costs, and lack continuous learning mechanisms, making it difficult to dynamically update the model to adapt to new proposal content. This leads to a decline in classification performance during long-term use of the system, affecting processing efficiency and intelligence level. Summary of the Invention

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions: The automatic classification method of public affairs proposals based on deep learning algorithm includes the following steps: Obtain public affairs proposal text, perform preliminary text processing operations, build a classification label system and annotate samples; Use BERT as the deep learning model architecture and determine whether it meets the set constraints; When the conditions are met, the corresponding model improvement mechanism is triggered to obtain the standard training model; otherwise, the current deep learning model is used as the standard training model; based on the standard training model, secondary pre-training is performed using public data, supervised fine-tuning is performed using labeled samples, the output layer structure is adjusted to match the number of classification labels, and a multi-task learning mechanism is introduced to obtain the standard model to be optimized; a cross-validation strategy is used to evaluate the standard model to be optimized, and a secondary evaluation model is constructed based on the evaluation indicators. The output comprehensive evaluation value is compared with the standard threshold, and the decision on whether to initiate the optimization action is made based on the comparison result, ultimately obtaining the standard optimization model; The standard optimization model is encapsulated as a RESTful API service, the API interface is integrated with the service platform, and a model operation log system is established to record each classification result and user feedback; an incremental learning mechanism is designed to enable regular model updates to adapt to new proposal type changes.

[0005] Furthermore, the sources of public affairs proposal texts include at least: Suggestion texts submitted to the service platform, hotline voice transcriptions, feedback forms, internally circulated opinion reports, and feedback information published in the media; Among them, the public affairs proposal text contains complete content fields and attached metadata information; The metadata information includes at least: timestamp and submitter's identity.

[0006] Furthermore, the content of the preliminary processing operation at least includes: Remove meaningless characters, duplicate content and invalid paragraphs, perform word segmentation on Chinese, build a stop word library, filter common function words and modal particles, and use named entity recognition technology to identify key entities and unify their expressions.

[0007] Furthermore, a classification labeling system is established based on management needs, including a two-level structure of main categories and subcategories; Among them, the main categories include at least: transportation and environmental protection; When the main category is transportation, the corresponding subcategories are road congestion management, bus route optimization, and subway construction; when the main category is environmental protection, the corresponding subcategories are noise pollution control, air quality monitoring, and garbage disposal; then, the expert labeling + semi-supervised learning method is used to construct the initial training dataset.

[0008] Furthermore, the constraints set include: limited deployment environment resources and text input length exceeding the limit; When the constraint condition is that the deployment environment resources are limited, the model improvement mechanism triggered is to adopt the improved lightweight model TinyBERT; when the constraint condition is that the text input length exceeds the standard, the model improvement mechanism triggered is to adopt the auxiliary model architecture: BiLSTM + Attention combined model, as an auxiliary feature extractor; when the constraint condition is that the deployment environment resources are limited + the text input length exceeds the standard, the model improvement mechanism triggered is to introduce the text slicing sub-mechanism and combine it with the hybrid model architecture.

[0009] Furthermore, the text slicing sub-mechanism: Filter public affairs proposal texts with more than M tokens and segment them using a sliding window method. The length of each segment is controlled within the maximum input length supported by the lightweight model; where M is greater than 0; The combined hybrid model architecture is as follows: Backbone model: TinyBERT as the base model; Auxiliary model: BiLSTM + Attention module, used to capture long-term dependencies across segments; Fusion strategy: The TinyBERT output vector of each segment is passed through the BiLSTM layer for sequence modeling, and the attention mechanism is used to weight the fusion of each segment information, and finally sent to the fully connected layer for classification.

[0010] Furthermore, supervised fine-tuning is performed using labeled samples. During the fine-tuning process, the probabilities of the main category and subcategory are output; the probability is the last layer of the model and is calculated through the activation function.

[0011] Furthermore, the evaluation indicators include at least: accuracy, precision, recall and F1 value; A secondary evaluation model is constructed based on the evaluation indicators, and each evaluation indicator is weighted and summed to obtain a comprehensive evaluation value; The process of deciding whether to start the optimization action based on the comparison results is as follows: When the comprehensive evaluation value does not exceed the standard threshold, the optimization action is initiated: backtracking analysis of the classification samples, supplementing samples and retraining, and introducing adversarial training; when the comprehensive evaluation value exceeds the standard threshold, no response action is taken.

[0012] Furthermore, the corresponding interface parameters in the API interface include at least: the original public affairs proposal text, the returned category label, and the confidence score.

[0013] An automatic classification system for public affairs proposals based on deep learning algorithms, including: Data processing module: Obtain public affairs proposal text, perform preliminary text processing operations, build a classification label system and annotate samples; Model training module: A deep learning model using BERT as the architecture is used to determine whether it meets the set constraints. If so, the corresponding model improvement mechanism is triggered to obtain a standard training model. Otherwise, the current deep learning model is used as the standard training model. Based on the standard training model, secondary pre-training is performed using public data, and supervised fine-tuning is performed using labeled samples. The output layer structure is adjusted to match the number of classification labels. After introducing a multi-task learning mechanism, a standard model to be optimized is obtained. A cross-validation strategy is used to evaluate the standard model to be optimized. A secondary evaluation model is constructed based on the evaluation indicators. The output comprehensive evaluation value is compared with the standard threshold. Based on the comparison result, it is decided whether to initiate the optimization action, and finally a standard optimized model is obtained. Deployment and operation module: Encapsulate the standard optimization model as a RESTful API service, integrate the API interface with the service platform, establish a model operation log system to record each classification result and user feedback; design an incremental learning mechanism to implement regular model updates to adapt to new proposal type changes.

[0014] The present invention provides a method and system for automatically classifying public affairs proposals based on a deep learning algorithm, which has the following beneficial effects: On the one hand, this solution can improve the original or basic functionality, taking into account resource constraints and long text processing capabilities. After quantization and runtime optimization, the inference speed meets the platform's real-time response requirements. On the other hand, it can promote the implementation of the model on edge devices, allowing the public affairs proposal classification system to be deployed on edge nodes, avoiding uploading raw data to the cloud, improving data privacy and security, and also enhancing the scalability of the system for subsequent use. It not only solves the problem that traditional lightweight models have difficulty processing long texts, but also takes into account the deployment feasibility in resource-constrained scenarios, with good practicality and promotion value. This solution significantly optimizes and improves the efficiency of the traditional proposal classification system. By leveraging the semantic understanding capabilities of an improved deep learning model, accuracy is also improved. The overall classification scheme can distinguish between main categories and break down into specific subcategories, facilitating refined management. Furthermore, through a continuous learning mechanism, the model can automatically adapt to emerging proposal types, ensuring the sustainable use of the classification system. This solution achieves efficient and accurate classification of massive public affairs proposals through a complete data collection, model training and system deployment process. This method not only solves the problems of low efficiency and poor accuracy in traditional classification methods, but also has good scalability and adaptability, and can provide strong support for intelligent governance. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1This is a schematic diagram of the overall process of the automatic classification method of public affairs proposals based on deep learning algorithm in the present invention. DETAILED DESCRIPTION

[0016] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0017] Example 1: See also Figure 1 This embodiment provides an automatic classification method for public affairs proposals based on deep learning algorithms, which includes three main stages: data collection and preprocessing, model training and optimization, and system deployment and operation. It aims to achieve efficient and accurate classification of opinion texts, and support a multi-level labeling system and continuous learning mechanism to improve the level of intelligent processing, thereby realizing and promoting the effective advancement of digitalization.

[0018] S1. Data collection and preprocessing: S1 is divided into the following sub-steps: S1.1. Collecting Public Affairs Proposal Texts: Obtain public affairs proposal texts from multiple sources or channels, and the sources of public affairs proposal texts include at least: The service platform can also collect data from other sources during the actual collection process, including suggestion texts submitted, hotline voice transcriptions, feedback forms, internally circulated opinion reports, and feedback information from the media, to ensure that the data in the classification system database is comprehensive and effective. Specifically, the collected public affairs proposal texts should contain complete content fields and be accompanied by metadata information; metadata information should at least include: timestamp, submitter identity, etc., for subsequent analysis; S1.2. Preliminary text processing (which may include text cleaning and standardization): Conduct preliminary processing on the collected public affairs proposal texts, and the content of the preliminary processing shall at least include: Remove meaningless characters (such as emoticons and special punctuation), remove duplicate content and invalid paragraphs, perform word segmentation on Chinese characters, build a stop word library, filter common function words and modal particles, and use named entity recognition technology to identify key entities and unify their expressions; Specifically, when performing word segmentation on Chinese, relevant tools are used, and the relevant tools are any one of Jieba, THULAC, or HanLP; when filtering common function words and modal particles, modal particles such as "ah,哎" in the text are filtered; when using named entity recognition technology to identify key entities.

[0019] S1.3. Construct a classification label system and annotate samples: Establish a classification label system according to management requirements, including a two-level structure of main categories and sub-categories; Among them, the main categories at least include: transportation and environmental protection; When the main category is: transportation, the corresponding sub-categories are: road congestion management, bus line optimization, and subway construction; when the main category is: environmental protection, the corresponding sub-categories are: noise pollution control, air quality monitoring, and waste treatment; Then, an initial training data set is constructed by using the method of expert annotation + semi-supervised learning; Specifically, it means: first, domain experts perform high-quality manual annotation on a small number of samples to form a seed data set with high credibility; then, use this seed set to train the initial model, and use the model to predict a large number of unannotated data, and select samples with higher confidence to join the training set; through an iterative method, perform multiple rounds of annotation and training to gradually expand the data scale; this method effectively reduces the manual annotation cost and improves the data utilization rate while ensuring the annotation quality, and is suitable for the actual application requirements of complex label systems, strong professionalism, and limited annotation resources in the scenario.

[0020] S2. Model training and optimization: In S2, it is also divided into the following several sub-steps: S2.1. Selection of deep learning model architecture: Use a deep learning model based on the Transformer structure, and this model includes: BERT and RoBERTa; in this embodiment, BERT is selected as the deep learning model architecture to obtain stronger Chinese semantic understanding ability; when the set constraint conditions are met, the corresponding model improvement mechanism is triggered to obtain the final standard training model; when the set constraint conditions are not met, the current deep learning model is used as the final standard training model; Among them, the set constraint conditions include: limited deployment environment resources and excessive text input length; Specifically, deployment environment resources are limited by: edge devices, low-profile servers, and embedded systems. Edge devices: These devices typically have lower computing power and memory, such as IoT devices and mobile devices. Low-profile servers: These may refer to servers equipped with fewer CPU cores, memory, and storage space, compared to high-performance computing clusters or cloud services. Embedded systems: These specifically refer to specialized computer systems integrated into other products. These often have strict requirements on cost, size, and energy consumption, resulting in relatively low hardware configurations. Exceeding the text input length limit occurs when the proportion of public affairs proposal texts exceeding 512 tokens exceeds 50%. Statistical analysis is used to obtain the average length, maximum length, and distribution curve of public affairs proposal texts. Based on the distribution characteristics and proportion, a conclusion is drawn as to whether the text input length exceeds the limit. When the deployment environment resources are limited, the model improvement mechanism is triggered: the improved lightweight model TinyBERT is used for subsequent processing. When the text input length exceeds the limit, the model improvement mechanism is triggered: the auxiliary model architecture: BiLSTM + Attention combined model is used as an auxiliary feature extractor. When the deployment environment resources are limited and the text input length exceeds the limit, the model improvement mechanism is triggered: the text slicing sub-mechanism is introduced and combined with the hybrid model architecture. Specifically, the text slicing sub-mechanism: Public affairs proposals with more than M tokens are filtered and segmented using a sliding window approach, with each segment length controlled within the maximum input length supported by the lightweight model (e.g., 512 tokens for TinyBERT). For example, a 2000-token proposal can be divided into multiple sub-segments, each 512 tokens long, with an overlap of 128 tokens. These sub-segments are then fed into the model for encoding. The overlap ensures contextual coherence. Combined with hybrid model architecture: Backbone model: TinyBERT is the basic model responsible for local semantic modeling; Auxiliary model: BiLSTM + Attention module, used to capture long-term dependencies across segments; Fusion strategy: The TinyBERT output vector of each segment is passed through the BiLSTM layer for sequence modeling, and the attention mechanism is used to weight the fusion of each segment information, and finally sent to the fully connected layer for classification; Results: This solution not only reduces the number of overall model parameters (meeting resource constraints) but also effectively models long-distance semantic associations (addressing the challenge of long text). For example, TinyBERT is used to extract vector representations for each of the three sub-segments described above. The BiLSTM-Attention module then integrates the semantics of the three segments, ultimately outputting a unified classification result. By adopting the above technical solutions, on the one hand, the original or basic functionality can be improved, taking into account resource constraints and long text processing capabilities: through the text slicing + hybrid model architecture, effective modeling of long texts under low-resource conditions is achieved; after experiments, it was found that: on the test set, the F1 value of the hybrid model reached 92.7%, slightly lower than the complete BERT model (94.8%), but much higher than the use of TinyBERT alone (89.5%), reflecting that the model or system maintains a high classification accuracy; after quantization and runtime optimization, the inference speed meets the real-time response requirements of the platform.

[0021] On the other hand, derivative benefits can be achieved by promoting the application of models on edge devices. This allows the public affairs proposal classification system to be deployed on edge nodes, avoiding uploading raw data to the cloud, improving data privacy and security, and also enhancing the scalability of the system. This hybrid architecture can be used for other long text classification tasks, such as the classification of petition materials, thereby forming a standardized, lightweight NLP processing flow and providing a reusable technical paradigm for subsequent AI systems. In summary, when the deployment environment is resource-constrained and long text processing needs coexist, this embodiment proposes a comprehensive solution based on a sliding window slicing + lightweight Transformer + BiLSTM-Attention hybrid architecture. This solution not only solves the problem that traditional lightweight models have difficulty processing long texts, but also takes into account the deployment feasibility in resource-constrained scenarios. It has good practicality and promotion value and is suitable for digital-related application scenarios.

[0022] S2.2. Fine-tune the model and perform transfer learning: Based on the standard training model, a second pre-training was performed using public data to improve the model's semantic understanding. Subsequently, supervised fine-tuning was performed using labeled proposal data (i.e., labeled public affairs proposal text or samples), adjusting the output layer structure to match the number of classification labels. A multi-task learning mechanism was also introduced to enable the model to simultaneously predict main category and subcategory labels, resulting in a standard model to be optimized. Public data includes: public corpus in the public domain; the example of S2.2 is: during the fine-tuning process, the corresponding text vector is input, and the probability of outputting the main category "traffic travel" is 0.96, and the probability of outputting the subcategory "road congestion management" is 0.92; Specifically, the probability given in the above example is calculated by the last layer of the model through an activation function (usually the Softmax function), which is used to indicate the degree of certainty of the model that the input text belongs to a specific category; The process is as follows: Feature extraction: First, the input text is converted into a form that the model can understand, that is, a text vector; this step usually includes word segmentation, word embedding and other processes to map the original text into a continuous vector space; Pass through the network: Next, this text vector is passed through a multi-layer neural network (such as the Transformer in BERT) Encoder layer) to capture the semantic information and contextual relationships in the text; each layer transforms the input data and gradually refines higher-level feature representations; Classifier layer: In the last layer, the model uses a fully connected layer to map the learned features to predefined categories. For multi-category classification problems, the Softmax activation function is usually applied after this layer; the Softmax function can convert all neuron outputs of the previous layer into a set of probability values. The sum of these probability values ​​is 1, which represents the likelihood that the input sample belongs to each category; Output probability: Ultimately, for a given input text, for example, the model may output two probability values: the probability of the main category "traffic" is 0.96, and the probability of the subcategory "road congestion management" is 0.92, which means that the model is very confident (96% probability) that the proposal mainly involves the field of "traffic", and in the sub-field, "road congestion management" is the most likely related topic (92% probability); S2.3. Evaluate and optimize model performance: A cross-validation strategy is used to evaluate the stability of the standard model to be optimized. The evaluation indicators include accuracy, precision, recall, and F1 value. A secondary evaluation model is constructed based on the evaluation indicators. The weighted sum of each evaluation indicator is used to obtain a comprehensive evaluation value. When the comprehensive evaluation value does not exceed the standard threshold, the optimization action is initiated: a retrospective analysis is performed on the classified samples (i.e., the samples formed after the public affairs proposal text is classified), samples are supplemented and retrained, and adversarial training is introduced to improve the robustness of the model; when the comprehensive evaluation value exceeds the standard threshold, no response action is taken; and finally a standard optimization model is obtained; a specific example of S2.3 is: after three iterative trainings, the comprehensive evaluation value of the model on the test set exceeds the standard threshold, meeting the actual deployment requirements.

[0023] S3. System deployment and operation: S3.1. Encapsulate the model as an API interface service: The standard optimization model obtained after optimization or training is encapsulated as a RESTful API service, providing an input and output interface in JSON format and supporting batch processing requests. The interface parameters include at least: the original public affairs proposal text, the returned category label, and the confidence score. S3.2. System integration with existing platforms: Integrate the API interface with the service platform (in addition, you can also choose to integrate with the petition management system according to actual needs), realize automatic classification of proposals and push them directly to the corresponding location, and set up a permission control module to ensure data security and access compliance (not detailed here); S3.3. Deploy monitoring and continuous learning mechanisms: Establish a model operation log system to record each classification result and user feedback; regularly sample new data for review and processing; design an incremental learning mechanism, and regularly update the model to adapt to new changes in proposal types; set up an anomaly detection module to identify samples with classification confidence below the set threshold and conduct a second review and processing; an example is: the system receives a new proposal, and through the incremental learning mechanism, the model successfully identifies that the proposal belongs to the "Transportation / Green Travel Management" category.

[0024] By adopting the above technical solutions, the traditional proposal classification system has been significantly optimized and its efficiency improved. The classification time for a single proposal has been reduced from an average of 3 minutes to less than 0.5 seconds, achieving millisecond-level response. By utilizing the semantic understanding capabilities of the improved deep learning model, the accuracy rate has reached over 95%, achieving the effect of improving classification accuracy. The overall system can not only distinguish between main categories but also refine them into specific subcategories, facilitating refined management. At the same time, through a continuous learning mechanism, the model can automatically adapt to emerging proposal types, ensuring the sustainable use of the classification system. From the service perspective: The application of this classification method promotes digital transformation: it provides a foundation for intelligent decision-making support and promotes process automation; it improves public satisfaction: proposals are processed faster and more accurately, and appeals are responded to more efficiently; it reduces labor costs: it reduces a large amount of repetitive work and frees up human resources for higher-value positions; it accumulates high-quality knowledge graphs: the classification results can be used to construct problem knowledge graphs, laying the foundation for future AI-assisted decision-making; and it enhances collaborative capabilities: the classification system can automatically identify and associate multiple proposals, promoting a joint processing mechanism. To sum up, the deep learning-based automatic classification method for public affairs proposals provided in this embodiment realizes efficient and accurate classification of massive proposals through a complete data collection, model training and system deployment process; this method not only solves the problems of low efficiency and poor accuracy in traditional classification methods, but also has good scalability and adaptability, and can provide strong support for intelligent governance.

[0025] Example 2: Based on Example 1, this embodiment also provides an automatic classification system for public affairs proposals based on a deep learning algorithm, including the following modules that run in sequence, specifically: Data processing module: Obtain public affairs proposal text, perform preliminary text processing operations, build a classification label system and annotate samples; Model training module: A deep learning model using BERT as the architecture is used to determine whether it meets the set constraints. If so, the corresponding model improvement mechanism is triggered to obtain a standard training model. Otherwise, the current deep learning model is used as the standard training model. Based on the standard training model, secondary pre-training is performed using public data, and supervised fine-tuning is performed using labeled samples. The output layer structure is adjusted to match the number of classification labels. After introducing a multi-task learning mechanism, a standard model to be optimized is obtained. A cross-validation strategy is used to evaluate the standard model to be optimized. A secondary evaluation model is constructed based on the evaluation indicators. The output comprehensive evaluation value is compared with the standard threshold. Based on the comparison result, it is decided whether to initiate the optimization action, and finally a standard optimized model is obtained. Deployment and operation module: Encapsulate the standard optimization model as a RESTful API service, integrate the API interface with the service platform, establish a model operation log system to record each classification result and user feedback; design an incremental learning mechanism to implement regular model updates to adapt to new proposal type changes.

[0026] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0027] 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, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0028] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. An automatic classification method for public affairs proposals based on a deep learning algorithm, characterized by: The steps include: Obtain public affairs proposal text, perform preliminary text processing operations, build a classification label system and annotate samples; Use BERT as the deep learning model architecture and determine whether it meets the set constraints; If the conditions are met, the corresponding model improvement mechanism is triggered to obtain the standard training model; otherwise, the current deep learning model is used as the standard training model; Based on the standard training model, we use public data for secondary pre-training, use labeled samples for supervised fine-tuning, adjust the output layer structure to match the number of classification labels, and introduce a multi-task learning mechanism to obtain the standard optimization model. We use a cross-validation strategy to evaluate the standard optimization model, and construct a secondary evaluation model based on the evaluation indicators. The output comprehensive evaluation value is compared with the standard threshold. The comparison result determines whether to initiate the optimization action, and finally the standard optimization model is obtained. The standard optimization model is encapsulated as a RESTful API service, the API interface is integrated with the service platform, and a model operation log system is established to record each classification result and user feedback; an incremental learning mechanism is designed to enable regular model updates to adapt to new proposal type changes.

2. The method for automatically classifying public affairs proposals based on a deep learning algorithm according to claim 1 is characterized by: The sources of public affairs proposal texts include at least: Suggestion texts submitted to the service platform, hotline voice transcriptions, feedback forms, internally circulated opinion reports, and feedback information published in the media; Among them, the public affairs proposal text contains complete content fields and attached metadata information; The metadata information includes at least: timestamp and submitter's identity.

3. The method for automatically classifying public affairs proposals based on a deep learning algorithm according to claim 1 is characterized by: The content of the preliminary processing operation shall at least include: Remove meaningless characters, duplicate content and invalid paragraphs, perform word segmentation on Chinese, build a stop word library, filter common function words and modal particles, and use named entity recognition technology to identify key entities and unify their expressions.

4. The method for automatically classifying public affairs proposals based on a deep learning algorithm according to claim 1, characterized in that: Establish a classification labeling system based on management needs, including a two-level structure of main categories and subcategories; Among them, the main categories include at least: transportation and environmental protection; When the main category is transportation, the corresponding subcategories are road congestion management, bus route optimization, and subway construction; when the main category is environmental protection, the corresponding subcategories are noise pollution control, air quality monitoring, and garbage disposal; then, the expert labeling + semi-supervised learning method is used to construct the initial training dataset.

5. The method for automatically classifying public affairs proposals based on a deep learning algorithm according to claim 1 is characterized by: The constraints set include: limited deployment environment resources and text input length exceeding the limit; When the constraint condition is that the deployment environment resources are limited, the model improvement mechanism triggered is to adopt the improved lightweight model TinyBERT; when the constraint condition is that the text input length exceeds the standard, the model improvement mechanism triggered is to adopt the auxiliary model architecture: BiLSTM + Attention combined model, as an auxiliary feature extractor; when the constraint condition is that the deployment environment resources are limited + the text input length exceeds the standard, the model improvement mechanism triggered is to introduce the text slicing sub-mechanism and combine it with the hybrid model architecture.

6. The method for automatically classifying public affairs proposals based on a deep learning algorithm according to claim 5, characterized in that: Text slicing sub-mechanism: Filter public affairs proposal texts with more than M tokens and segment them using a sliding window approach. The length of each segment is controlled within the maximum input length supported by the lightweight model; where M is greater than 0. The combined hybrid model architecture is as follows: Backbone model: TinyBERT as the base model; Auxiliary model: BiLSTM + Attention module, used to capture long-term dependencies across segments; Fusion strategy: The TinyBERT output vector of each segment is passed through the BiLSTM layer for sequence modeling, and the attention mechanism is used to weight the fusion of each segment information, and finally sent to the fully connected layer for classification.

7. The method for automatically classifying public affairs proposals based on a deep learning algorithm according to claim 1, characterized in that: Use labeled samples for supervised fine-tuning. During the fine-tuning process, the probabilities of the main category and subcategory are output; the probabilities are the last layer of the model and are calculated through the activation function.

8. The method for automatically classifying public affairs proposals based on a deep learning algorithm according to claim 1, characterized in that: Evaluation indicators include at least: accuracy, precision, recall and F1 value; A secondary evaluation model is constructed based on the evaluation indicators, and each evaluation indicator is weighted and summed to obtain a comprehensive evaluation value; The process of deciding whether to start the optimization action based on the comparison results is as follows: When the comprehensive evaluation value does not exceed the standard threshold, the optimization action is initiated: backtracking analysis of the classification samples, supplementing samples and retraining, and introducing adversarial training; when the comprehensive evaluation value exceeds the standard threshold, no response action is taken.

9. The method for automatically classifying public affairs proposals based on a deep learning algorithm according to claim 1, characterized in that: The corresponding interface parameters in the API interface include at least: the original public affairs proposal text, the returned category label, and the confidence score.

10. An automatic classification system for public affairs proposals based on deep learning algorithms, characterized by: include: Data processing module: Obtain public affairs proposal text, perform preliminary text processing operations, build a classification label system and annotate samples; Model training module: uses BERT as the deep learning model architecture and determines whether it meets the set constraints. If it does, it triggers the corresponding model improvement mechanism to obtain the standard training model. Otherwise, the current deep learning model is used as the standard training model. Based on the standard training model, we use public data for secondary pre-training, use labeled samples for supervised fine-tuning, adjust the output layer structure to match the number of classification labels, and introduce a multi-task learning mechanism to obtain the standard optimization model. We use a cross-validation strategy to evaluate the standard optimization model, and construct a secondary evaluation model based on the evaluation indicators. The output comprehensive evaluation value is compared with the standard threshold. The comparison result determines whether to initiate the optimization action, and finally the standard optimization model is obtained. Deployment and operation module: Encapsulate the standard optimization model as a RESTful API service, integrate the API interface with the service platform, establish a model operation log system to record each classification result and user feedback; design an incremental learning mechanism to implement regular model updates to adapt to new proposal type changes.

Images