Online public opinion research, judgment, analysis, classification and grading early warning method and system

Abstract

The invention discloses an online public opinion research, judgment, analysis, classification and grading early warning method and system, and the method comprises the steps: constructing a malicious public opinion data detection and judgment model, and carrying out the iterative training of a pre-training neural network through large-scale online public opinion historical data, so as to achieve the fine adjustment and optimization of the model; after malicious public opinion data are detected, fine-grained classification of malicious public opinions is realized through deep semantic comprehension and multi-dimensional text structure analysis by adopting BERT and ERNIE deep learning models and combining semantic role labeling and dependency syntactic analysis technologies; a malicious public opinion data grading and distinguishing model is constructed, deep correlation between public opinion events is captured through multi-layer convolution operation, and the recognition capacity for public opinion sensitivity and urgency degree is improved. According to the invention, malicious public opinion data related to a specific industry field can be efficiently and accurately identified and classified, and the malicious public opinion data is subjected to severity grading, so that the accuracy and the real-time performance of network public opinion management are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent judgment and early warning technology for network public opinion, and in particular to a method and system for classifying and grading early warning of network public opinion judgment and analysis. Background Art

[0002] With the sustained and rapid development of my country's economy and the accelerating process of social informatization, the online public opinion environment has become increasingly diversified and complex. Especially when it comes to key industries related to the national economy and people's livelihood, such as the power industry, the opinions on the Internet are even more complicated, and there is no lack of malicious public opinion, which poses a severe challenge to the reputation management of related companies and the construction of public trust.

[0003] At present, although some online public opinion monitoring systems have been applied in actual work, most of these systems rely on traditional keyword retrieval methods. This method may be competent when dealing with simple and clear public opinion information, but it is powerless when facing the complex and changeable online language environment, especially malicious public opinion containing complex semantics such as metaphors and sarcasm. The system cannot deeply understand the true intention and potential impact of this information, resulting in slow response speed, low recognition accuracy, and frequent occurrence of a large number of situations requiring manual review, which not only increases the audit cost, but also seriously affects the processing efficiency.

[0004] Existing online public opinion monitoring systems generally have low recognition accuracy, high false positive and false negative rates when identifying malicious public opinion data. This is not only due to technical limitations, but also due to the lack of effective methods to accurately distinguish malicious public opinion data from ordinary public opinion information. Malicious public opinion is often concealed and misleading. They may be hidden in seemingly normal online discussions and are difficult to be effectively captured and identified by traditional methods.

[0005] Therefore, how to efficiently and accurately identify and classify malicious public opinion data related to specific industry fields (such as power grids), and grade their severity to improve the accuracy and real-time nature of online public opinion governance has become an urgent problem to be solved in the current field of online public opinion governance. Summary of the invention

[0006] In view of this, the present invention proposes a network public opinion assessment and analysis classification and grading early warning method and system to solve the problems existing in the prior art of low recognition accuracy, high false alarm and missed alarm rates, and inability to effectively distinguish between malicious public opinion and ordinary public opinion in the network public opinion monitoring system.

[0007] The specific technical solutions of the present invention are as follows:

[0008] A network public opinion research and analysis classification and grading early warning method, comprising:

[0009] Construct malicious public opinion data detection and discrimination model: Build a detection and identification model for malicious public opinion based on the Transformer pre-trained neural network, and use large-scale online public opinion historical data to iteratively train the pre-trained neural network to achieve model fine-tuning and optimization;

[0010] Construct a fine-grained classification and discrimination model for malicious public opinion data: After detecting malicious public opinion data, the deep learning models of BERT and ERNIE are used, combined with semantic role labeling and dependency syntax analysis technology, to achieve fine-grained classification of malicious public opinion through deep semantic understanding and multi-dimensional text structure analysis;

[0011] Construct a hierarchical discrimination model for malicious public opinion data: Construct a public opinion knowledge graph, in which the basic information of public opinion events, such as time, location, people involved, and event type, is included. At the same time, the semantic information of sentiment analysis, topic classification, and keyword extraction, as well as the correlation between public opinion events, including causal relationships and similar events, are introduced to provide rich feature inputs for the graph neural network; a dynamic update mechanism for the public opinion knowledge graph is established to incrementally update the graph by collecting new public opinion data in real time, and the incremental learning ability of the graph neural network is combined to ensure that the model can continuously adapt to the new public opinion environment; a hierarchical discrimination model is constructed based on the graph neural network, and a multi-layer graph convolutional network GCN is used to extract features from the public opinion knowledge graph. The deep correlation between public opinion events is captured through multi-layer convolution operations to improve the recognition ability of public opinion sensitivity and urgency.

[0012] Furthermore, the malicious public opinion data detection and discrimination model is constructed, which also includes:

[0013] The detection and recognition model integrates the attention mechanism in deep learning to capture key information and contextual dependencies in the text and improve sensitivity to malicious speech features; the attention mechanism includes variants of multi-head self-attention and position encoding enhanced self-attention mechanism;

[0014] Incorporate auxiliary technologies and strategies for adversarial sample training into the detection and recognition model to enhance the robustness of the model, reduce misjudgments caused by data noise or adversarial attacks, and thereby reduce missed detection and false alarm rates.

[0015] Furthermore, the malicious public opinion data detection and discrimination model adopts data enhancement technology such as synonym replacement and sentence transformation to simulate the diversity of malicious public opinion, continuously enrich training samples, and improve the generalization ability of the model.

[0016] Furthermore, constructing a fine-grained classification and discrimination model for malicious public opinion data also includes: when analyzing text, in addition to vocabulary and syntactic structure, introducing multi-dimensional features of the text's sentiment tendency, topic distribution, and paragraph structure to perform comprehensive text representation to improve classification results.

[0017] Furthermore, constructing a fine-grained classification and discrimination model for malicious public opinion data also includes: utilizing the few-sample sequence labeling technology of cross-domain transfer learning to migrate the labeled data in related fields to the malicious public opinion classification task, thereby effectively labeling key entities and events in the public opinion text.

[0018] Furthermore, constructing a fine-grained classification and discrimination model for malicious public opinion data also includes: using few-sample learning technologies including prototype networks and meta-learning, combined with LSTM-CRF and BERT-CRF sequence labeling models, to accurately extract key information while identifying malicious public opinion, and to accurately label key entities and events with limited labeled data, and then subdivide malicious public opinion into types such as false information, offensive speech, rumors, and misleading propaganda.

[0019] Furthermore, the malicious public opinion data classification and discrimination model adopts an adaptive classification algorithm to dynamically adjust the classification strategy according to the characteristics of the input text, improve the adaptability to different types of malicious public opinion, and continuously collect new malicious public opinion data for updating and optimization through incremental learning mechanism and user feedback mechanism. At the same time, a user feedback mechanism is established to collect user feedback on classification results for further adjustment and improvement of the model.

[0020] Furthermore, constructing a hierarchical identification model for malicious public opinion data also includes: introducing an attention mechanism into the hierarchical identification model, weighting the important nodes and key connections of public opinion events, increasing attention to key information, analyzing the development context of public opinion events, and grading the sensitivity and urgency of events in combination with the public opinion knowledge graph, and dividing public opinion events into general public opinion, sensitive public opinion, and important public opinion.

[0021] Furthermore, the method also includes: a public opinion text structured processing process based on adaptive sequence annotation; the process includes:

[0022] Sequence labeling of public opinion events, including entity labeling of public opinion texts through a Transformer-based pre-trained neural network sequence labeling model, and accurate extraction of key entities of people, organizations, and locations in the event;

[0023] Identification of entity relationships, including using entity relationship modeling technology to build a knowledge graph of public opinion events and generate structured public opinion event descriptions;

[0024] The structured expression of public opinion events includes storing structured information in a graph database.

[0025] A network public opinion research and analysis classification and grading early warning system, using the above network public opinion research and analysis classification and grading early warning method, the system includes:

[0026] The Internet data collection module uses web crawler technology to obtain public opinion data on specific fields from the Internet. After cleaning and filtering, it uses semi-automatic annotation tools and expert review to ensure the high quality and diversity of the data set;

[0027] The semantic embedding processing module is used to clean, segment, convert and deeply embed the collected data; it converts the collected raw public opinion data into more valuable information through rules and semantic analysis;

[0028] Malicious public opinion detection and identification module, which accurately identifies potential malicious public opinions based on deep learning models;

[0029] Classification and discrimination module, used to further refine the categories of malicious public opinion data;

[0030] The classification and discrimination module is used to provide graded warnings based on the sensitivity and urgency of public opinion events;

[0031] The public opinion warning module is used to receive the results of the core analysis module and make the final public opinion warning.

[0032] The beneficial effects of the present invention are:

[0033] 1. Significantly improve recognition accuracy and efficiency: Through deep learning technology, the recognition accuracy rate is increased to more than 88%, effectively reducing the false alarm rate and missed alarm rate, ensuring high efficiency and accuracy when processing large-scale public opinion data.

[0034] 2. Significantly reduce the cost of manual review: The automated public opinion monitoring and classification system reduces the workload of manual review and reduces the labor costs of enterprises in public opinion management.

[0035] 3. Enhance the real-time response to public opinion: monitor and grade malicious public opinion in real time, promptly capture the development trend of public opinion events, provide a rapid response mechanism, and ensure a rapid response to major public opinion events.

[0036] 4. The system is highly robust and scalable: The malicious public opinion detection, classification and grading models can adapt to a variety of complex public opinion environments and can be updated to respond to emerging malicious public opinion data.

[0037] 5. Adapt to public opinion events of different scales and complexities: It has strong response capabilities during the rapid fermentation stage of events, can accurately identify key nodes of public opinion dissemination, and provide a visual evolution path.

[0038] 6. High-precision analysis and fine-grained classification: Through time series models and enhanced intelligence technology, accurate perception and dynamic prediction of the spread of public opinion can be achieved, and visual display can help managers intuitively grasp the evolution trend of public opinion.

[0039] 7. Provide multi-dimensional intelligent early warning function: provide strong support for decision makers, ensure timely and effective response to potential crisis events, and significantly improve the efficiency and effectiveness of public opinion management and response. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0041] Figure 1 This is a schematic diagram of the structural framework of the network public opinion research and analysis classification and grading early warning system of the present invention;

[0042] Figure 2 A schematic diagram of the process of sequence annotation of public opinion events of the present invention;

[0043] Figure 3 It is a structural diagram of the detection and recognition model of network public opinion data of the present invention;

[0044] Figure 4 It is a flow chart of the network public opinion research and analysis classification and grading early warning method of the present invention;

[0045] Figure 5 It is a flow chart of the network public opinion classification and grading processing model of the present invention. DETAILED DESCRIPTION

[0046] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0047] The present invention proposes a network public opinion research and analysis classification and grading early warning method and system.

[0048] First of all, the method of the present invention is committed to building a set of efficient malicious public opinion data detection and discrimination models, which are based on Transformer pre-trained neural networks (such as BERT, RoBERTa, T5 and its variants) for model fine-tuning and optimization. These models have demonstrated excellent context understanding and language generation capabilities in the field of natural language processing. By fine-tuning these pre-trained models, they are more suitable for malicious public opinion detection tasks. The present invention uses large-scale network public opinion historical data for iterative training to build a high-precision detection and recognition model for malicious public opinion. This model not only integrates the attention mechanism in deep learning, such as introducing variants of the self-attention mechanism (including multi-head self-attention and position encoding enhancement, etc.) to more finely capture the key information and contextual dependencies in the text, and improve the model's sensitivity to malicious speech features, but also incorporates a variety of auxiliary technologies and strategies, such as adversarial sample training, to enhance the robustness of the model, reduce misjudgments caused by data noise or adversarial attacks, thereby further reducing the missed rate and false alarm rate, and improving the overall performance.

[0049] Secondly, the present invention constructs an accurate fine-grained classification and discrimination model for malicious public opinion data. After successfully detecting malicious public opinion data, the model realizes fine-grained classification of malicious public opinion through deep semantic understanding and multi-dimensional text structure analysis. The present invention adopts advanced deep learning models (such as BERT, ERNIE, etc.), and combines semantic role labeling, dependency syntax analysis and other technologies to deeply understand the complex semantic relationships in the text and improve the accuracy of key information extraction. At the same time, the present invention not only analyzes the vocabulary and syntactic structure of the text, but also introduces multi-dimensional features such as the sentiment tendency, topic distribution, and chapter structure of the text, and uses these features to perform a more comprehensive text representation to improve the classification effect. In addition, the present invention uses the few-sample sequence labeling technology of cross-domain transfer learning to migrate the labeled data in related fields to the malicious public opinion classification task, and effectively label the key entities and events in the public opinion text. Through transfer learning, the present invention can utilize existing knowledge, quickly adapt to new malicious public opinion classification tasks, and improve classification efficiency and accuracy; by adopting few-sample learning technology (such as prototype networks, meta-learning, etc.) and combining sequence annotation models (such as LSTM-CRF, BERT-CRF, etc.), the present invention can accurately extract key information while identifying malicious public opinion; it can accurately annotate key entities and events with limited annotated data, provide strong support for fine-grained classification, and further subdivide malicious public opinion into multiple types such as false information, offensive speech, rumors, misleading propaganda, etc., providing more accurate information support for the security governance of cyberspace.

[0050] Finally, the present invention constructs a malicious public opinion data classification and discrimination model based on the public opinion knowledge graph and the graph neural network. When constructing the public opinion knowledge graph, the present invention incorporates the basic information of the public opinion event (such as time, place, people involved, event type, etc.), and introduces semantic information such as sentiment analysis, topic classification, keyword extraction, and the correlation between public opinion events (such as causal relationships, similar events, etc.), providing rich feature inputs for the graph neural network. The present invention establishes a dynamic update mechanism for the public opinion knowledge graph, collects new public opinion data in real time, and performs incremental updates of the graph. At the same time, combined with the incremental learning ability of the graph neural network, the present invention ensures that the model can continue to adapt to the new public opinion environment. The model deeply analyzes the development context of public opinion events, accurately distinguishes the sensitivity and urgency of public opinion, and realizes the intelligent classification of general public opinion, sensitive public opinion and important public opinion. This not only helps enterprises to respond quickly, but also effectively avoids the spread of negative public opinion and maintains the brand image and social reputation of the enterprise.

[0051] Figure 1 The overall framework design of the network public opinion research and analysis classification and grading warning system is demonstrated. The network public opinion research and analysis classification and grading warning system of the present invention mainly includes an Internet data collection module, a semantic embedding processing module, a malicious public opinion detection and identification module, a classification and discrimination module, and a grading and discrimination module. The Internet data collection module is responsible for obtaining public opinion data on a specific field from the network, such as the power grid; the semantic embedding processing module cleans, segments, converts, and deeply semantically embeds the collected data; the public opinion research and judgment module identifies malicious public opinion data based on a deep learning model; the public opinion classification module subdivides malicious public opinion data by category; the public opinion grading module performs graded warnings on the sensitivity and urgency of public opinion events; the public opinion warning module receives the results of the core analysis module and performs the final public opinion warning. Specifically, the Internet data collection module focuses on extensively and deeply collecting public opinion data on specific fields (such as the power grid) from major social platforms, news websites and other network channels. These data are strictly cleaned and filtered, and the double guarantee of semi-automatic annotation tools and expert review is used to ensure the high quality and diversity of the data set. Subsequently, the semantic embedding processing module will perform in-depth processing on these data, including word segmentation conversion, deep semantic embedding representation, etc., to provide a solid semantic foundation for subsequent modules. The malicious public opinion detection and identification module accurately identifies potential malicious public opinions based on the deep learning model. The classification and discrimination module will further refine the categories of these malicious public opinion data, and the graded discrimination module will provide graded warnings based on the sensitivity and urgency of public opinion events, providing decision makers with timely and accurate public opinion information.

[0052] The process of structuring public opinion text based on adaptive sequence annotation includes sequence annotation of public opinion events, identification of entity relationships and structured expression of public opinion events. Figure 2The sequence labeling of public opinion events is demonstrated: First, a pre-trained neural network language model based on Transformer is constructed to predict the probability of the next word based on the context characters. On this basis, the characters and words in the public opinion text are encoded and embedded, and then the bidirectional LSTM+CRF layer is used to predict the optimal tag sequence, implement entity labeling, and accurately extract key entities such as people, organizations, and places in the event. Identification of entity relationships: Then, using entity relationship modeling technology, a knowledge graph of public opinion events is constructed, and a structured description of public opinion events is generated. Structured expression of public opinion events: Finally, this structured information is stored in a graph database, which greatly facilitates subsequent analysis and mining. This process aims to better understand and analyze the information contained in the public opinion text through structured processing, thereby providing more valuable data support for decision makers.

[0053] Figure 3 The detection and recognition model of online public opinion data is demonstrated. The model uses data enhancement technology (including synonym replacement, word segmentation conversion, sentence transformation, semantic embedding representation, etc.) to simulate the diversity of malicious public opinion, continuously enrich training samples, significantly improve the generalization ability of the model, and support semantic relationship recognition, syntactic structure recognition, and topic clustering analysis. Secondly, the Transformer Encoder architecture is used to fine-tune and optimize the pre-trained neural network. Combined with multi-head self-attention, position encoding enhancement, and adversarial sample generation, the features of public opinion text are extracted through sample adversarial training to achieve accurate semantic understanding of malicious public opinion. Then, by constructing a network public opinion knowledge graph, entity relationship extraction, semantic disambiguation / entity linking, and knowledge graph joint embedding technology are used to achieve cross-domain transfer few-sample learning, graph neural network reinforcement learning, and prototype learning / meta-learning. The model has been trained with a large amount of field-related historical public opinion data, and its attention mechanism can accurately capture keywords and sentence structures in public opinion data, thereby achieving high-precision malicious public opinion detection, false information / rumor recognition, and offensive speech recognition.

[0054] Figure 4The flow chart of the network public opinion research and analysis classification and grading early warning method is shown. The main functions of the system include data collection, semantic processing, public opinion research and analysis, and public opinion early warning. Data collection: The system grabs data from various sources on the Internet, including Weibo, WeChat public accounts, blogs, forums, and news comments. The data collection process involves web crawler technology to obtain a large amount of original public opinion data in an automated manner. Semantic processing: Process the collected original public opinion data, including content extraction and data mining. Through rules and semantic analysis, the data is converted into more valuable information. Public opinion research and analysis: The public opinion research and analysis module is responsible for public opinion identification, sentiment analysis, and classification and grading. According to the characteristics of public opinion data, it is deeply analyzed to determine its nature and severity. Public opinion early warning: According to the results of public opinion research and analysis, the system can predict the trend and severity of public opinion development. The early warning results can be displayed to users in the front-end display interface of the network public opinion research and analysis classification and grading early warning system in the form of charts, etc., to help them understand the current public opinion situation. Through these steps, the entire system realizes the effective management and monitoring of network public opinion and provides users with timely and accurate public opinion information.

[0055] like Figure 5 As shown in the figure, the network public opinion classification and grading processing model mainly analyzes network public opinion texts to identify industry relevance, public opinion categories, and classify public opinion data. Specifically, based on the graph neural network, a multi-layer graph convolutional network (GCN) is used to extract features from the public opinion knowledge graph. Through multi-layer convolution operations, the deep correlation between public opinion events is gradually captured, and the model's ability to identify the sensitivity and urgency of public opinion is improved. By analyzing the semantics and context of the public opinion text, the types of malicious public opinion, such as power distribution, power transmission, power consumption, power outages, excessively high electricity prices, and unstable voltage, are accurately identified, which improves the classification accuracy of the model in long texts. The model also introduces an adaptive classification algorithm to dynamically adjust the classification strategy according to the characteristics of the input text, thereby improving the adaptability to different types of malicious public opinion. In addition, through the incremental learning mechanism and user feedback mechanism, the model can continuously collect new malicious public opinion data for updating and optimization, ensuring the accuracy and reliability of the classification results. At the same time, a user feedback mechanism is established to collect user feedback on the classification results for further adjustment and improvement of the model. The attention mechanism is introduced to perform weighted processing on important nodes and key connections of public opinion events, so as to increase the model's attention to key information and improve the accuracy of classification and discrimination. By analyzing the development context of public opinion events and combining the public opinion knowledge graph, the sensitivity and urgency of the events are classified. The system automatically divides public opinion events into non-public opinion, general public opinion and serious public opinion according to the relevance and impact range of different public opinion events, helping decision makers to take timely response measures.

[0056] The present invention proposes an innovative solution to the significant technical defects existing in the existing network public opinion monitoring system, namely, low recognition accuracy, high false alarm rate and missed alarm rate, and the inability to effectively distinguish between malicious public opinion data and ordinary public opinion - a network public opinion judgment analysis classification and grading early warning method and system. The current network public opinion monitoring system mainly relies on traditional keyword retrieval methods, which may be effective when processing simple texts, but its limitations are fully revealed when faced with network public opinion information containing complex semantics and contextual relationships. Especially when processing malicious public opinion data, the existing systems often lack deep semantic understanding capabilities, resulting in slow response speeds and low recognition accuracy, which in turn triggers a large number of manual review requirements, which not only increases costs, but also seriously affects processing efficiency.

[0057] In order to overcome these challenges, the present invention aims to achieve a deep understanding and intelligent analysis of online public opinion information by introducing cognitive enhancement intelligent technology. The present invention can efficiently and accurately identify and classify malicious public opinion data that is closely related to specific industry fields (such as power grids). Through advanced algorithm models, the present invention can deeply understand the semantics and emotional tendencies behind the text, thereby effectively distinguishing malicious public opinion from ordinary public opinion. In addition, the present invention also has the ability to grade the severity of the identified malicious public opinion data, and conduct a comprehensive assessment based on multiple dimensions such as the scope of information dissemination, degree of influence, potential hazards, etc., to provide a scientific basis for subsequent governance measures.

[0058] To sum up, the network public opinion analysis, classification and grading early warning method and system proposed in the present invention not only solves the problems of existing systems in terms of recognition accuracy and false alarm rate, but also significantly improves the accuracy and real-time performance of network public opinion management through deep semantic understanding and severity grading functions, providing strong technical support for network public opinion management in related industry fields.

[0059] Specifically, the beneficial effects of the present invention are:

[0060] 1. Improve recognition accuracy and efficiency: This invention uses deep learning technology to conduct intelligent research and analysis on online public opinion, effectively reducing the false alarm rate and missed alarm rate, and improving the recognition accuracy to more than 88%. This significant improvement ensures the efficiency and accuracy of the system when processing large-scale public opinion data, thereby improving the accuracy of online public opinion governance.

[0061] 2. Reduce manual review costs: Through the automated public opinion monitoring and classification system, the present invention greatly reduces the workload of manual review, thereby reducing the labor costs of enterprises in public opinion management and improving work efficiency.

[0062] 3. Enhance the real-time response to public opinion: The system has the ability to monitor and handle malicious public opinion in real time and in a graded manner. It can promptly capture the development trend of public opinion events, provide enterprises with a rapid response mechanism, and ensure rapid response and effective handling of major public opinion events.

[0063] 4. Robustness and scalability of the system: The malicious public opinion detection, classification and grading model designed by the present invention has good robustness and scalability, can adapt to a variety of complex public opinion environments, and can update the model at any time to cope with emerging malicious public opinion data, ensuring the continuous effectiveness and adaptability of the system.

[0064] 5. Adapt to public opinion events of different scales and complexities: The public opinion analysis method of the present invention can cope with public opinion events of different scales and complexities, especially has a strong response ability in the rapid fermentation stage of the event, which helps enterprises to manage public opinion in a timely and effective manner.

[0065] 6. Accurately identify key nodes of public opinion dissemination: The system can accurately identify key nodes of public opinion dissemination and provide a detailed visual evolution path, so that managers can more clearly understand the dissemination trends and key nodes of public opinion, and enable managers to more intuitively grasp the evolution trend of public opinion.

[0066] 7. High-precision analysis and fine-grained classification capabilities: The present invention has significant high-precision analysis and fine-grained classification capabilities, which can effectively cope with the complex and changeable public opinion environment and improve the pertinence and effectiveness of public opinion management.

[0067] 8. Realize accurate perception and dynamic prediction of the public opinion dissemination trend: Through time series model and enhanced intelligent technology, the present invention realizes accurate perception and dynamic prediction of the public opinion dissemination trend, provides managers with more comprehensive public opinion information, and helps to formulate more effective response strategies.

[0068] 9. Provide multi-dimensional public opinion intelligent early warning function: The present invention provides a multi-dimensional public opinion intelligent early warning function, which provides strong support for decision makers, ensures that potential crisis events can be responded to in a timely and effective manner, and further improves the efficiency and effectiveness of public opinion management and response.

[0069] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A network public opinion analysis, classification and grading early warning method, characterized in that: include: Construct malicious public opinion data detection and discrimination model: Based on the Transformer pre-trained neural network, a detection and identification model for malicious public opinion is constructed. The pre-trained neural network is iteratively trained using large-scale online public opinion historical data to achieve model fine-tuning and optimization; Construct a fine-grained classification and discrimination model for malicious public opinion data: After detecting malicious public opinion data, the deep learning models of BERT and ERNIE are used, combined with semantic role labeling and dependency syntax analysis technology, to achieve fine-grained classification of malicious public opinion through deep semantic understanding and multi-dimensional text structure analysis; Construct a hierarchical discrimination model for malicious public opinion data: Construct a public opinion knowledge graph, in which the basic information of public opinion events, such as time, location, people involved, and event type, is included. At the same time, sentiment analysis, topic classification, semantic information extracted from keywords, and the correlation between public opinion events, including causal relationships and similar events, are introduced to provide rich feature input for the graph neural network. A dynamic update mechanism for the public opinion knowledge graph is established, and the graph is incrementally updated by collecting new public opinion data in real time. Combined with the incremental learning ability of the graph neural network, this ensures that the model can continuously adapt to the new public opinion environment. The hierarchical discrimination model is constructed based on the graph neural network, and the multi-layer graph convolutional network GCN is used to extract features from the public opinion knowledge graph. The deep correlation between public opinion events is captured through multi-layer convolution operations, thereby improving the recognition ability of the sensitivity and urgency of public opinion.

2. The network public opinion research and analysis classification and grading early warning method according to claim 1 is characterized in that: Building a malicious public opinion data detection and discrimination model also includes: The detection and recognition model is integrated with the attention mechanism in deep learning to capture key information and contextual dependencies in the text and improve sensitivity to malicious speech features; the attention mechanism includes variants of multi-head self-attention and position encoding enhanced self-attention mechanisms; The detection and recognition model is integrated with auxiliary techniques and strategies for adversarial sample training to enhance the robustness of the model, reduce misjudgments caused by data noise or adversarial attacks, and thereby reduce missed detection and false alarm rates.

3. The network public opinion analysis, classification and grading early warning method according to claim 1 is characterized in that: The malicious public opinion data detection and discrimination model adopts data enhancement technology synonym replacement and sentence transformation to simulate the diversity of malicious public opinion, continuously enrich training samples, and improve the generalization ability of the model.

4. The network public opinion research and analysis classification and grading early warning method according to claim 1 is characterized in that: Constructing a fine-grained classification and discrimination model for malicious public opinion data also includes: when analyzing text, in addition to vocabulary and syntactic structure, introducing multi-dimensional features of the text's sentiment tendency, topic distribution, and paragraph structure for comprehensive text representation to improve classification results.

5. The network public opinion research and analysis classification and grading early warning method according to claim 1 is characterized in that: Constructing a fine-grained classification and discrimination model for malicious public opinion data also includes: using the few-sample sequence labeling technology of cross-domain transfer learning to migrate the labeled data in related fields to the malicious public opinion classification task, thereby effectively labeling the key entities and events in the public opinion text.

6. The network public opinion research and analysis classification and grading early warning method according to claim 1 is characterized in that: Building a fine-grained classification and discrimination model for malicious public opinion data also includes: using few-sample learning technologies including prototype networks and meta-learning, combined with LSTM-CRF and BERT-CRF sequence labeling models, to accurately extract key information while identifying malicious public opinion, and accurately label key entities and events with limited labeled data, and then subdivide malicious public opinion into types such as false information, offensive speech, rumors, and misleading propaganda.

7. The network public opinion research and analysis classification and grading early warning method according to claim 1 is characterized in that: The malicious public opinion data classification and discrimination model adopts an adaptive classification algorithm, dynamically adjusts the classification strategy according to the characteristics of the input text, improves the adaptability to different types of malicious public opinions, and continuously collects new malicious public opinion data for updating and optimization through an incremental learning mechanism and a user feedback mechanism. At the same time, a user feedback mechanism is established to collect user feedback on classification results for further adjustment and improvement of the model.

8. The network public opinion research and analysis classification and grading early warning method according to claim 1 is characterized in that: Constructing a malicious public opinion data grading and identification model also includes: introducing an attention mechanism into the grading and identification model, weighting the important nodes and key connections of public opinion events, increasing attention to key information, analyzing the development context of public opinion events, and grading the sensitivity and urgency of events in combination with the public opinion knowledge graph, and dividing public opinion events into general public opinion, sensitive public opinion, and important public opinion.

9. The network public opinion research and analysis classification and grading early warning method according to claim 1 is characterized in that: The method further includes: a public opinion text structured processing flow based on adaptive sequence annotation; the flow includes: Sequence labeling of public opinion events, including entity labeling of public opinion texts through a Transformer-based pre-trained neural network sequence labeling model, and accurate extraction of key entities of people, organizations, and locations in the event; Identification of entity relationships, including using entity relationship modeling technology to build a knowledge graph of public opinion events and generate structured public opinion event descriptions; The structured expression of public opinion events includes storing structured information in a graph database.

10. A network public opinion research and analysis classification and grading warning system, using the network public opinion research and analysis classification and grading warning method as claimed in claim 1, characterized in that: The system comprises: The Internet data collection module uses web crawler technology to obtain public opinion data on specific fields from the Internet. After cleaning and filtering, it uses semi-automatic annotation tools and expert review to ensure the high quality and diversity of the data set; The semantic embedding processing module is used to clean, segment, convert and deeply embed the collected data; it converts the collected raw public opinion data into more valuable information through rules and semantic analysis; Malicious public opinion detection and identification module, which accurately identifies potential malicious public opinions based on deep learning models; Classification and discrimination module, used to further refine the categories of malicious public opinion data; The classification and discrimination module is used to provide graded warnings based on the sensitivity and urgency of public opinion events; The public opinion warning module is used to receive the results of the core analysis module and make the final public opinion warning.

Images