Identification and processing methods of abnormal network public opinion

Abstract

The present invention relates to a method for identifying and processing abnormalities in network public opinion, and belongs to the field of artificial intelligence. The present invention uses innovative points such as intelligent knowledge extraction technology of network public opinion intelligence, network public opinion abnormality identification technology based on social communication structure, expert cognitive process and strategy modeling technology for network public opinion event analysis, and key element screening technology based on time relationship graph attention mechanism to improve the accuracy of abnormal identification of network public opinion and the timeliness of abnormal identification of network public opinion. This solution can provide great help for early warning of large-scale public opinion outbreaks, thereby providing a reference for efficient crisis response of relevant departments.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, and specifically relates to a method for identifying and processing abnormal network public opinion. Background Art

[0002] The processing of massive, multi-modal, multi-source heterogeneous data is a global problem. In the field of public opinion anomaly identification research, which is highly professional and has high quality requirements, it is particularly difficult to achieve the transformation of data into knowledge.

[0003] At present, from the knowledge space construction of network public opinion data to knowledge mining, network public opinion prediction and identification, all countries have adopted multiple technical routes to carry out some technical research in their respective fields, but have not yet formed a complete network public opinion anomaly identification system. In terms of the current status and development trends of related technologies at home and abroad, it still faces a series of technical challenges, mainly reflected in the lack of network public opinion knowledge acquisition and integration capabilities and the lack of network public opinion quantitative prediction and identification methods. Aiming at the current problem of network public opinion anomaly identification in my country, combined with the current new theoretical and technical methods, the present invention forms a complete network public opinion anomaly identification solution. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] The technical problem to be solved by the present invention is how to provide a method for identifying and processing abnormalities in online public opinion, so as to solve the current problems in my country of insufficient ability to acquire and integrate online public opinion knowledge and insufficient means for quantitative prediction and identification of online public opinion.

[0006] (II) Technical solution

[0007] In order to solve the above technical problems, the present invention proposes a method for identifying and processing abnormal network public opinion, which comprises the following steps:

[0008] S1. Constructing a network public opinion data resource library: Obtaining the original network public opinion data and constructing a data resource library. Obtaining the original network public opinion data includes: tracking and sorting out the data source, intelligent data collection, and building a data source system; constructing a data resource library includes: processing the obtained multi-heterogeneous original data and cataloging the processed data;

[0009] S2. Constructing a network public opinion graph and analyzing the path of public opinion propagation: This includes a public opinion graph construction module and a network public opinion propagation path analysis module. The public opinion graph construction module processes the structured and unstructured data in the public opinion data resource library to form a knowledge graph containing entities, relationships, and causal events, and constructs a graph based on the knowledge graph. The network public opinion propagation path analysis module vectorizes the graph and identifies the path of public opinion propagation through a graph search algorithm and similarity comparison.

[0010] S3. Identification of abnormal online public opinion based on social communication structure: It includes three steps:

[0011] The first step is to represent the event word vector based on TF-IDF or word2vec, and calculate the event distance;

[0012] The second step is to construct an abstract graph of network public opinion based on social communication. This graph is constructed by analyzing similar events based on homogeneity theory and event heat analysis based on message dissemination.

[0013] The third step is to identify network public opinion anomalies based on the abstract event graph. First, classify network public opinion anomalies, then predict the popularity based on the LDA model, and classify network public opinion anomalies according to the popularity. Finally, identify the anomaly type based on the network public opinion anomaly recognition technology of the scalable graph neural network model;

[0014] S4. Strategies for responding to abnormal network public opinion events based on the mechanism of event propagation and evolution: Based on the expert thinking in the field of public opinion research and the causal reasoning mechanism of the causal graph, the verification and review of the abnormal identification of network public opinion can be realized; based on the screening model technology of the time series graph attention mechanism, the key factors that play a decisive role in the identification of abnormal network public opinion can be screened; according to the results of the abnormal identification of network public opinion, targeted response strategies can be formulated.

[0015] (III) Beneficial effects

[0016] The present invention proposes a method for identifying and processing abnormal network public opinion. The present invention proposes a set of solutions for identifying and responding to abnormal public opinion in my country. It uses the knowledge intelligent extraction technology of network public opinion intelligence, the network public opinion abnormality identification technology based on social communication structure, the expert cognitive process and strategy modeling technology of network public opinion event analysis, and the key element screening technology based on the time relationship graph attention mechanism. The innovation points improve the accuracy of abnormal network public opinion identification and the timeliness of abnormal network public opinion identification. This solution can provide a good help for the early warning of large-scale public opinion outbreaks, thereby providing a reference for the efficient crisis response of relevant departments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Schematic diagram of the network public opinion anomaly identification and handling method of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, content and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below in conjunction with the drawings and examples.

[0019] The present invention combines artificial intelligence methods and social communication theories to construct an accurate and efficient method for identifying public opinion anomalies, and based on the above theoretical methods, forms a complete set of solutions for identifying public opinion anomalies.

[0020] The present invention provides a complete set of network public opinion anomaly identification and handling methods, which are used to solve the current problems in my country in terms of insufficient network public opinion knowledge acquisition and integration capabilities and insufficient network public opinion quantitative prediction and identification means.

[0021] In order to realize the recognition and handling of abnormal network public opinion, the present invention is carried out in the following aspects, and the specific methods and steps are as follows:

[0022] S1. Constructing a network public opinion data resource library: Obtaining the original network public opinion data and constructing a data resource library. Obtaining the original network public opinion data includes: tracking and sorting out the data source, intelligent data collection, and building a data source system; constructing a data resource library includes: processing the obtained multi-heterogeneous original data and cataloging the processed data;

[0023] S2. Constructing a network public opinion graph and analyzing the path of public opinion propagation: This includes a public opinion graph construction module and a network public opinion propagation path analysis module. The public opinion graph construction module processes the structured and unstructured data in the public opinion data resource library to form a knowledge graph containing entities, relationships, and causal events, and constructs a graph based on the knowledge graph. The network public opinion propagation path analysis module vectorizes the graph and identifies the path of public opinion propagation through a graph search algorithm and similarity comparison.

[0024] S3. Identification of abnormal online public opinion based on social communication structure: It includes three steps:

[0025] The first step is to represent the event word vector based on TF-IDF or word2vec, and calculate the event distance;

[0026] The second step is to construct an abstract graph of network public opinion based on social communication. This graph is constructed by analyzing similar events based on homogeneity theory and event heat analysis based on message dissemination.

[0027] The third step is to identify network public opinion anomalies based on the abstract event graph. First, classify network public opinion anomalies, then predict the popularity based on the LDA model, and classify network public opinion anomalies according to the popularity. Finally, identify the anomaly type based on the network public opinion anomaly recognition technology of the scalable graph neural network model;

[0028] S4. Strategies for responding to abnormal network public opinion events based on the mechanism of event propagation and evolution: Based on the expert thinking in the field of public opinion research and the causal reasoning mechanism of the causal graph, the verification and review of the abnormal identification of network public opinion can be realized; based on the screening model technology of the time series graph attention mechanism, the key factors that play a decisive role in the identification of abnormal network public opinion can be screened; according to the results of the abnormal identification of network public opinion, targeted response strategies can be formulated.

[0029] The network public opinion anomaly identification and processing method proposed in the present invention is composed of a network public opinion data resource library construction module, a network public opinion event map construction and public opinion propagation path analysis module, a network public opinion anomaly identification module based on social propagation structure, and a network public opinion anomaly event response strategy generation module based on event propagation evolution mechanism. The specific implementation steps are as follows:

[0030] The generation of response strategies for abnormal network public opinion events based on the mechanism of event propagation and evolution first integrates the thinking of experts in the public opinion field, the display form of abnormal network public opinion, and the causal reasoning mechanism of the causal graph to realize the verification and review of the identification of abnormal network public opinion and form the knowledge of responding to abnormal network public opinion. Then, based on the screening model of the time series graph attention mechanism and other technologies, the key factors that play a decisive role in the identification of abnormal network public opinion are screened. Finally, based on the results of abnormal network public opinion identification, combined with the verification and review experience of abnormal network public opinion identification and the key factors that play a decisive role in abnormal network public opinion, the generation of response strategies for abnormal events is realized.

[0031] The following is a detailed introduction to the methods for identifying and handling anomalies in online public opinion.

[0032] (1) Network public opinion data resource library construction module

[0033] This module mainly includes the following two aspects:

[0034] 1) Obtaining the original data of online public opinion, mainly including tracking and sorting out data sources, conducting intelligent data collection based on topic sniffing technology, and building a data source system.

[0035] 2) Constructing a data resource library, which mainly includes processing the acquired multivariate heterogeneous raw data and cataloging the processed data. This module finally outputs an organized network public opinion data resource library for use by subsequent modules. The following is a detailed introduction to the relevant content of the above functional points involved in this module.

[0036] When obtaining the original data of online public opinion, the data sources of online public opinion events involved mainly include world-renowned international affairs research institutions, relevant journals, foreign government websites and other institutions. The preliminary sorting of the data sources of online public opinion events facilitates the rapid development of data intelligent collection, and then the theme sniffing and data automatic tracking and collection work are carried out. While the theme sniffing and data automatic tracking and collection work are carried out, the evaluation and feedback of the data source will be formed, thereby promoting the construction and dynamic improvement of the data source system of online public opinion events. The theme sniffing technology analyzes and mines the given data source through technical means to parse out potential data sources. On the one hand, this technology verifies the existing data source to ensure its information output stability and information quality, and on the other hand, it discovers potential new information sources from these information sources. According to the characteristics of Internet data sources and sniffing forms, the information acquisition agents are divided into three categories, namely sniffing agents, perception agents, and autonomous agents, which respectively realize the tracking and sniffing of information sites, search engines and interactive media information.

[0037] When building a data source system, the main focus is on systematically planning, selecting, acquiring, organizing and managing data sources, and establishing the entire process of a resource system with specific functions.

[0038] In the data resource library construction part, multi-source data processing mainly integrates network public opinion data from different sources and types to form a unified data resource library, providing data support for the subsequent network public opinion event map construction. In order to improve the utilization rate of data resources and the efficiency of information retrieval, the present invention adopts keyword-based multi-source data fusion technology and data content-based network public opinion fusion processing technology.

[0039] The cataloging of online public opinion event data mainly includes: metadata design, format conversion and metadata integration. The metadata design is completed manually by domain experts to build the ontology system, and the metadata database is built by engineers. The format and metadata integration work is completed by engineers who convert the attributes and relationship information of the ontology content into the table structure in the database to complete the metadata format conversion and data integration. The constructed online public opinion data resource library will be used by subsequent modules.

[0040] (2) Module for constructing the logic of online public opinion and analyzing the path of public opinion dissemination

[0041] The network public opinion reasoning map construction and public opinion propagation path module mainly include two modules, the public opinion reasoning map construction module and the network public opinion propagation path analysis module.

[0042] Among them, the public opinion and reasoning graph construction module is mainly responsible for processing the structured and unstructured data in the public opinion data resource library to form a knowledge graph including entities, relationships and causal events, and further construct a reasoning graph based on the knowledge graph;

[0043] The network public opinion propagation path analysis module vectorizes the event graph produced in the previous step and identifies the public opinion propagation path through graph search algorithm and similarity comparison. Finally, the module obtains the public opinion event graph and the corresponding public opinion propagation path analysis results, which provide support for subsequent anomaly identification and strategy formulation.

[0044] Specifically,

[0045] In the public opinion graph construction module, the public opinion text data is first preprocessed through methods such as word segmentation, part-of-speech tagging, grammatical analysis and relationship rule analysis, and then the knowledge graph is constructed through methods such as entity recognition, relationship extraction, event extraction, knowledge fusion, etc. Finally, the graph is constructed through methods such as event framework construction, event type identification, event element identification and event causal relationship extraction. At this point, the construction of the graph is completed.

[0046] The network public opinion propagation path analysis module mainly completes the propagation path analysis of network public opinion events, which requires causal event extraction, causal relationship extraction rule setting, causal relationship structure classification and causal relationship extraction. It mainly includes two steps: first, for a new network public opinion event, calculate its similarity with each node in the event graph, and find the most similar node of the new event; second, compare the similarity with the threshold to determine whether to continue propagation. Specifically, given a new public opinion event, traverse the nodes in the network public opinion event graph, calculate the similarity with the new event, and finally compare the maximum similarity with the given threshold. If it is less than the preset threshold, it means that there is no corresponding node of the target event in the graph, and it cannot be propagated. Otherwise, the possible events in reality are inferred based on the subsequent nodes in the graph. If a node has multiple subsequent events, the probability of occurrence is calculated based on the edge weight coefficient, and then this process is recursively executed until there are no subsequent events, so that one or more possible complete propagation paths can be obtained.

[0047] (3) Network public opinion anomaly identification module based on social communication structure

[0048] The network public opinion anomaly identification module based on social communication structure is used to analyze the spread of network public opinion and identify the anomalies of network public opinion. It mainly includes the following three steps:

[0049] The first step is to represent the event word vector, which can be based on the event word vector representation of TF-IDF or word2vec, and calculate the event distance.

[0050] The second step is to construct an abstract map of network public opinion based on social communication. Through similar event analysis based on homogeneity theory and event heat analysis based on message dissemination, an abstract map of network public opinion can be constructed.

[0051] The third step is to identify network public opinion anomalies based on abstract event graphs. First, classify the network public opinion anomalies, then predict the popularity based on the LDA model, and grade the network public opinion anomalies according to the popularity. Finally, identify the anomaly type based on the network public opinion anomaly recognition technology based on the scalable graph neural network model.

[0052] Here, the network public opinion anomaly recognition technology mainly applies entity granularity representation learning technology, temporal relationship quintuple granularity representation learning technology, and event granularity representation learning technology. The following introduces these three technologies respectively.

[0053] In the entity granularity representation learning technology, an entity representation learning model that integrates entity types is proposed. This model attempts to integrate information to enrich the semantics in knowledge representation. The model represents entities, relations, and entity types as low-dimensional dense vectors, and designs a mapping function based on entity types to construct the representation of entities under specific types. In a triple containing a specific relationship, the representation of the head and tail entities is first mapped to the representation under the type restricted by the relationship, and then the translation idea is used to learn the semantics expressed by the triple. The mapping function adopts the form of circular convolution, so that the type vector and the entity vector can fully communicate, and the feature association between the corresponding dimension and the non-corresponding dimension is obtained, which improves the quality of knowledge embedding. In negative sampling, the model restricts the type of the sampled entity to be consistent with the type of the replaced entity with a certain probability, which increases the difference between entities of the same type and takes into account some similarities between them. This method enriches the semantic information contained in the knowledge representation and solves the problem of complex relationships to a certain extent.

[0054] Granular representation learning technology for temporal relationship quintuples: The knowledge representation algorithm for contextual information of temporal and spatial relationship quintuples in knowledge graphs is proposed to address the shortcomings of traditional knowledge graphs in terms of temporal and spatial representation models. While modeling the static structural features of entities in knowledge graphs, the algorithm considers the temporal sequence characteristics, time interval characteristics, and spatial association constraints of the evolution of relationships between entities, and obtains vector representations of entities and relationships with temporal and spatial associations based on learning. Specifically, this method uses the static structural context information and dynamic structural context information of quintuples in knowledge graphs for representation learning, where the dynamic structural context information injects information about the temporal and spatial evolution of relationships.

[0055] Event granularity representation learning technology proposes discrete event representation, event representation based on distributed semantics, and event representation based on inter-event information, which makes up for the shortcomings of previous separate representations.

[0056] The second step of this module is to construct an abstract graph of network public opinion based on social communication. The abstract graph is to generalize events and then find more general and common causal relationships of event evolution in a certain network public opinion field. In the present invention, events with similar distances are generalized by clustering, and multiple specific events are represented by a more representative event, so as to construct an abstract graph of network public opinion. The clustering samples are events, that is, text samples. The original kmeans clustering method uses TF-IDF vectors in the process of converting text into numerical vectors. The vectors formed lack semantic elements. The present invention improves the clustering algorithm by using word2vec word vectors and distance calculation methods to achieve the best generalization results.

[0057] The third step is to identify the anomaly of network public opinion based on the abstract event graph. The model further proposes a scalable graph neural network model (SGNN), which can perform network representation learning on the graph structure of large-scale propagation event graphs. By borrowing the idea of ​​divide and conquer during the training process, SGNN only calculates on a small-scale related subgraph at a time to expand GGNN to a large-scale directed and cyclic weighted graph instead of calculating on the entire graph. The model can also be applied to other network representation learning tasks after modification. By calculating the similarity between the representation of the event in the context and the representation of the candidate event, the model can identify the correct type of post-network public opinion anomaly.

[0058] (4) Strategies for dealing with abnormal online public opinion events based on the evolution mechanism of event propagation

[0059] This module mainly includes: (1) Based on the expert thinking in the field of public opinion research, the different forms of abnormal display of online public opinion, and the causal reasoning mechanism of the event graph, the verification and review of the abnormal identification of online public opinion is realized; (2) The key factors that play a decisive role in the abnormal identification of online public opinion are screened by using technologies such as the screening model based on the time series graph attention mechanism; (3) According to the results of abnormal identification of online public opinion, targeted response strategies are formulated. The following is a detailed introduction to the relevant contents of these three aspects.

[0060] The network public opinion anomaly identification and review technology is mainly completed through three parts: expert thinking acquisition and construction, network public opinion anomaly multi-form display and interpretable causal reasoning enhanced by the reasoning map. 1) Expert thinking acquisition and construction is mainly based on the relevant theories of problem solving and cognitive psychology. The network public opinion anomaly in the real situation is taken as the research object. First, relevant information is summarized from public documents, policies, and think tanks to determine the analysis angle, and then a questionnaire is formed. After that, semi-structured interviews are conducted with experts to obtain the process information of experts analyzing network public opinion anomalies, form an integrated event analysis process, analyze the cognitive process and processing characteristics of experts analyzing network public opinion anomalies, and provide a theoretical basis for constructing an expert thinking model for network public opinion anomaly analysis. 2) In the analysis of key element intelligence of network public opinion anomaly events, multi-form display can effectively and clearly display the reasoning process of complex public opinion anomaly events. The present invention notes that in the process of causal reasoning, there may be a series of evidence events between the cause event and the result event. These evidence events can serve as additional clues to help the model understand the causal mechanism between cause and effect, and enhance the interpretability and robustness of the results. 3) In order to make full use of the role of evidence events in causal reasoning in the part of identifying anomalies in online public opinion, an event graph knowledge enhanced causal reasoning framework (ExCAR) is used. Given a causal pair, this framework can obtain a series of causal evidence from a pre-built causal knowledge base.

[0061] The screening part of key elements for identifying abnormal network public opinion mainly includes the screening model based on the time-series graph attention mechanism and the key elements for identifying abnormal network public opinion. 1) In the screening model based on the time-series graph attention mechanism, in the process of knowledge graph construction, the URI address is used to record the original source of each knowledge fragment, so that in the deductive analysis, for a given knowledge fragment, the original data source can be traced according to its corresponding URI, thereby providing fine-grained data basis for decision-making. In the knowledge graph constructed this time, the structure of the graph is no longer static, because entities and the links between them will evolve over time and establish connections in the form of events, so the present invention is also called an event knowledge graph. Here, the present invention abstracts the event knowledge graph as G inf In G inf In the present invention, the predicted entity (i.e., knowledge fragment) and the edge between the entire reasoning graph have been obtained. If the original data of the knowledge fragment is to be traced back, it is necessary to calculate the event with the largest attention score on the path related to it. At the same time, the key elements of the knowledge fragment are obtained according to the high and low arrangement of the attention scores. 2) The key elements of the abnormal identification of network public opinion mainly involve the use of attention mechanism and event relationship graph attention layer to judge the impact of key elements after screening.

[0062] The part of the strategy for dealing with abnormal network public opinion mainly includes abnormal network public opinion identification, multi-dimensional guidance of network public opinion, and evaluation and response of network public opinion. 1) In the abnormal identification of network public opinion, network public opinion has the characteristics of freedom, interactivity, diversity, deviation, and suddenness. It can influence the emotions and judgments of the people, promote and change the development and trend of events, and is easily used by opposition elements. It has become an important factor affecting social stability. Based on the method of constructing a reasoning map and extracting information, the present invention identifies 5 types of anomalies that are concerned in network public opinion analysis, combined with the background of the subject it relies on and the needs of network public opinion analysis. In the present invention, the abnormal types of network public opinion are divided into 5 categories, namely, abnormal network public opinion events caused by ideology, abnormal network public opinion events caused by interest demands, abnormal network public opinion events caused by emotional catharsis, abnormal network public opinion events caused by concept advocacy, and abnormal network public opinion events caused by social security. 2) Pay attention to the diversification of social participation mechanisms, and let entities at all levels participate in the emergency management of online public opinion. Different management methods should be set up for different personnel. Positive news can be spread through mainstream online media and other channels to guide people to develop positive online public opinion, and use diversified management and education methods to play the role of universal participation. Guidance can be carried out from the following aspects: Strengthen the disclosure of online public opinion information in emergencies, and adhere to the transparency given by the concept of good governance. Strive to achieve the rule of law in the online public opinion space, and adhere to the rule of law given by the concept of good governance. Establish a scientific management consciousness and adhere to the effectiveness given by the concept of good governance. 3) In terms of online public opinion evaluation and response, by studying the prediction of online public opinion dissemination trends under emergencies, formulate an emergency mechanism that can control the correct direction of online public opinion based on the prediction results, in order to accurately evaluate the dynamics of online public opinion, achieve accurate resolution of social crises caused by online public opinion, avoid unnecessary impacts of online public opinion on society, and look forward to the coordinated dissemination of online public opinion in a positive direction.

[0063] This invention proposes a set of solutions for the needs of identifying and responding to public opinion anomalies in my country. It uses the knowledge intelligent extraction technology of network public opinion intelligence, network public opinion anomaly identification technology based on social communication structure, expert cognitive process and strategy modeling technology for network public opinion event analysis, and key element screening technology based on time relationship graph attention mechanism. It improves the accuracy of network public opinion anomaly identification and the timeliness of network public opinion anomaly identification. This solution can provide great help for the early warning of large-scale public opinion outbreaks, thereby providing a reference for efficient crisis response of relevant departments.

[0064] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for identifying and processing abnormal network public opinion, characterized in that: The method comprises the following steps: S1. Constructing a network public opinion data resource library: Obtaining the original network public opinion data and constructing a data resource library. Obtaining the original network public opinion data includes: tracking and sorting out the data source, intelligent data collection, and building a data source system; constructing a data resource library includes: processing the obtained multi-heterogeneous original data and cataloging the processed data; S2. Constructing a network public opinion graph and analyzing the path of public opinion propagation: This includes a public opinion graph construction module and a network public opinion propagation path analysis module. The public opinion graph construction module processes the structured and unstructured data in the public opinion data resource library to form a knowledge graph containing entities, relationships, and causal events, and constructs a graph based on the knowledge graph. The network public opinion propagation path analysis module vectorizes the graph and identifies the path of public opinion propagation through a graph search algorithm and similarity comparison. S3. Identification of abnormal online public opinion based on social communication structure: It includes three steps: The first step is to represent the event word vector based on TF-IDF or word2vec, and calculate the event distance; The second step is to construct an abstract graph of network public opinion based on social communication. This graph is constructed by analyzing similar events based on homogeneity theory and event heat analysis based on message dissemination. The third step is to identify network public opinion anomalies based on the abstract event graph. First, classify network public opinion anomalies, then predict the popularity based on the LDA model, and classify network public opinion anomalies according to the popularity. Finally, identify the anomaly type based on the network public opinion anomaly recognition technology of the scalable graph neural network model; S4. Strategies for responding to abnormal network public opinion events based on the mechanism of event propagation and evolution: Based on the expert thinking in the field of public opinion research and the causal reasoning mechanism of the causal graph, the verification and review of the abnormal identification of network public opinion can be realized; based on the screening model technology of the time series graph attention mechanism, the key factors that play a decisive role in the identification of abnormal network public opinion can be screened; according to the results of the abnormal identification of network public opinion, targeted response strategies can be formulated.

2. The method for identifying and processing abnormal network public opinion according to claim 1, characterized in that: When obtaining the original data of online public opinion, the data sources of online public opinion events include world-renowned international affairs research institutions, relevant journals and foreign government websites. The data sources of online public opinion events are preliminarily sorted out, and then subject sniffing and automatic data tracking and collection are carried out. While the subject sniffing and automatic data tracking and collection are carried out, an evaluation and feedback on the data source will be formed, thereby promoting the construction and dynamic improvement of the data source system of online public opinion events; among them, the subject sniffing technology analyzes and mines the given data source, parses out the potential data source, and divides the information acquisition agents into three categories according to the characteristics of the Internet data source and the sniffing form, namely sniffing agents, perception agents and autonomous agents, which respectively realize the tracking and sniffing of information sites, search engines and interactive media information; when constructing the data source system, the whole process of systematically planning, selecting, acquiring, organizing and managing data sources and establishing a resource system with specific functions is carried out.

3. The method for identifying and processing abnormal network public opinion according to claim 2, characterized in that: When constructing a data resource library, multi-source institutional data processing will integrate network public opinion data from different sources and of different types to form a unified data resource library, providing data support for the subsequent construction of a network public opinion event map; including: keyword-based multi-source data fusion technology and data content-based network public opinion fusion processing technology; network public opinion event data cataloging includes: metadata design, format conversion and metadata integration, among which metadata design is manually completed by domain experts to build the ontology system, and the metadata database is built by engineers. The format and metadata integration work is completed by engineers who convert the attributes and relationship information of the ontology content into the table structure in the database, complete the metadata format conversion and data integration, and the constructed network public opinion data resource library will be used by subsequent modules.

4. The method for identifying and processing abnormal network public opinion according to claim 3, characterized in that: In the public opinion graph construction module, the public opinion text data is first preprocessed through word segmentation, part-of-speech tagging, grammatical analysis and relationship rule analysis methods, and then the knowledge graph is constructed through entity recognition, relationship extraction, event extraction, and knowledge fusion methods. Finally, the graph is constructed through event framework construction, event type identification, event element identification and event causal relationship extraction methods. At this point, the construction of the graph is completed.

5. The method for identifying and processing abnormal network public opinion according to claim 4, characterized in that: The network public opinion propagation path analysis module completes the propagation path analysis of network public opinion events, which requires causal event extraction, causal relationship extraction rule setting, causal relationship structure classification and causal relationship extraction; The method includes two steps: first, for a new online public opinion event, calculate its similarity with each node in the event graph and find the most similar node to the new event; second, compare the similarity with the threshold to determine whether to continue propagation; specifically, given a new public opinion event, traverse the nodes in the online public opinion event graph, calculate the similarity with the new event, and finally compare the maximum similarity with the given threshold. If it is less than the preset threshold, it means that there is no node corresponding to the target event in the graph and it cannot be propagated. Otherwise, infer the possible events in reality based on the subsequent nodes in the graph. If a node has multiple subsequent events, calculate the probability of occurrence based on the edge weight coefficient, and then recursively execute this process until there are no subsequent events. In this way, one or more possible complete propagation paths are obtained.

6. The method for identifying and processing abnormal network public opinion according to claim 5, characterized in that: The network public opinion anomaly recognition based on social communication structure uses entity granularity representation learning technology, temporal relationship quintuple granularity representation learning technology and event granularity representation learning technology; In the entity granularity representation learning technology, an entity representation learning model that integrates entity types is proposed. This model attempts to integrate information to enrich the semantics in knowledge representation. This model represents entities, relations, and entity types as low-dimensional dense vectors, and designs a mapping function based on entity types to construct the representation of entities under specific types. In a triple containing a specific relationship, the representation of the head and tail entities is first mapped to the representation under the type restricted by the relationship, and then the translation idea is used to learn the semantics expressed by the triple. The mapping function adopts the form of circular convolution, so that the type vector and the entity vector can fully communicate with each other, and the feature association between corresponding dimensions and non-corresponding dimensions is obtained, which improves the quality of knowledge embedding. In negative sampling, the model restricts the type of the sampled entity to be consistent with the type of the replaced entity with a certain probability, which increases the difference between entities of the same type and takes into account some similarities between them. Time-series relationship quintuple granular representation learning technology: The static structural context information and dynamic structural context information of the quintuple in the knowledge graph are used for representation learning. The dynamic structural context information injects the information of the temporal and spatial evolution of the relationship. Event granularity representation learning technology proposes discrete event representation, event representation based on distributed semantics, and event representation based on inter-event information, which makes up for the shortcomings of previous separate representations.

7. The method for identifying and processing abnormal network public opinion according to claim 6, characterized in that: The second step of the network public opinion anomaly identification based on social communication structure is to construct an abstract graph of network public opinion based on social communication. The abstract graph is to generalize events and then find more general and common causal relationships of event evolution in a certain network public opinion field; generalize events with similar distances by clustering, and use a more representative event to represent multiple specific events, so as to construct an abstract graph of network public opinion; cluster samples are events, that is, text samples, and the clustering algorithm is improved by using word2vec word vectors and distance calculation methods to make the generalization result optimal.

8. The method for identifying and processing abnormal network public opinion according to claim 6, characterized in that: The third step of the network public opinion anomaly identification based on social communication structure uses a network public opinion anomaly identification model based on an abstract event graph. The model further proposes a scalable graph neural network model (Scaled Graph Neural Network, SGNN) to perform network representation learning on the large-scale communication event graph structure; by borrowing the divide and conquer idea during the training process, SGNN only calculates on a small-scale related subgraph at a time to expand GGNN to a large-scale directed cyclic weighted graph instead of calculating on the entire graph; by calculating the similarity between the representation of the event in the context and the representation of the candidate event, the model identifies the correct post-network public opinion anomaly type.

9. The method for identifying and processing abnormal network public opinion according to any one of claims 6 to 8, characterized in that: The verification and review of the identification of abnormal online public opinion is completed through the acquisition and construction of expert thinking, the multi-form display of abnormal online public opinion, and the explainable causal reasoning enhanced by the event map. The acquisition and construction of expert thinking is based on the relevant theories of problem solving and cognitive psychology. It takes the abnormal network public opinion in real situations as the research object. First, relevant information is summarized from public documents, policies, and think tanks to determine the analysis angle, and then a questionnaire is formed. Then, through semi-structured interviews with experts, the process information of experts' analysis of abnormal network public opinion is obtained to form an integrated event analysis process, analyze the cognitive process and processing characteristics of experts' analysis of abnormal network public opinion, and provide a theoretical basis for constructing an expert thinking model for abnormal network public opinion analysis. Multi-modal display of abnormal online public opinion: Multi-modal display can effectively and clearly show the reasoning process of complex abnormal public opinion events. In the causal reasoning process, there may be a series of evidence events between the cause event and the result event. These evidence events can serve as additional clues to help the model understand the causal mechanism between the cause and the result, and enhance the interpretability and robustness of the results. In the explainable causal reasoning part enhanced by event graph, in order to make full use of the role of evidential events in causal reasoning, a causal reasoning framework enhanced by event graph knowledge (EventGraph KnowledgeEnhancedExplainableCausalReasoning, ExCAR) is used. Given a causal pair, this framework can obtain a series of causal evidence from the pre-built causal knowledge base.

10. The method for identifying and processing abnormal network public opinion according to claim 9, characterized in that: When screening the key factors that play a decisive role in identifying abnormal online public opinion, a screening model based on the time-series graph attention mechanism and key factors for identifying abnormal online public opinion are included; In the screening model based on the time-series graph attention mechanism, the original source of each knowledge fragment is recorded using the URI address during the knowledge graph construction process. Therefore, in the deductive analysis, for a given knowledge fragment, the original data source can be traced according to its corresponding URI, thereby providing fine-grained data basis for decision-making. In the knowledge graph constructed this time, the structure of the graph is no longer static, because entities and the links between them will evolve over time and establish connections in the form of events, so it is also called an event knowledge graph. Abstract the event knowledge graph as G inf , in G inf In the above example, we get the predicted entity, i.e. the knowledge piece, and the edge between the whole reasoning graph. If we want to trace the source of the original data of the knowledge piece, we need to calculate the event with the largest attention score on the path related to it, and arrange the key elements of the knowledge piece according to the high and low attention scores; Key elements for identifying abnormal online public opinion: Use the attention mechanism and event relationship graph attention layer to determine the impact of key elements after screening; The strategy for dealing with abnormal online public opinion includes: identifying abnormal online public opinion, guiding online public opinion in multiple ways, and evaluating and dealing with online public opinion; In the identification of abnormal network public opinion, based on the method of constructing the event graph and extracting information, five types of abnormalities in network public opinion analysis are identified, namely, abnormal network public opinion events caused by ideology, abnormal network public opinion events caused by interest demands, abnormal network public opinion events caused by emotional catharsis, abnormal network public opinion events caused by the promotion of ideas, and abnormal network public opinion events caused by social security; We should attach importance to the diversification of social participation mechanisms, let entities at all levels jointly participate in the emergency management of online public opinion, set up different management methods for different personnel, spread positive news through mainstream online media, guide people to develop positive online public opinion, and use diversified management and education methods to give full play to the role of national participation; In terms of online public opinion assessment and response, by studying the prediction of online public opinion dissemination trends during emergencies, an emergency mechanism is formulated based on the prediction results to control the direction of online public opinion.

Images