Risk public opinion analysis method based on natural language processing

Abstract

The invention discloses a natural language processing-based risk public opinion analysis method, which comprises the following steps of: capturing a public opinion text in real time, and desensitizing a user identity field by adopting a Hash pseudo identifier to obtain an original text stream; constructing a double-classification weak supervision framework, calculating a risk confidence coefficient, and when the risk confidence coefficient is greater than or equal to a first threshold value, writing a corresponding text into a labeling pool to form a risk corpus; mapping the risk corpus into a dynamic heterogeneous risk knowledge graph according to an entity-relationship-time slice meta-path rule; applying a multi-head space-time diffusion Transform to a sub-graph intersecting with the current window in the graph to obtain a risk propagation vector of an event node; and inputting the risk propagation vector and the historical baseline vector into a causal comparative analysis model, outputting a risk early warning index, and when the early warning index is greater than or equal to a second threshold value, pushing a risk alarm including an event path to a background management and control center. According to the invention, the automation degree, accuracy and reliability of public opinion risk monitoring and early warning are effectively improved.

Description

Technical Field

[0001] This invention relates to the technical field of natural language processing and online public opinion risk monitoring, and in particular to a risk public opinion analysis method based on natural language processing. Background Technology

[0002] With the rapid development of internet technology, a large amount of textual public opinion information has been generated on various social media, industry forums, and government information disclosure platforms, forming an increasingly rich data source and providing an important foundation for risk public opinion monitoring and analysis. In response to the massive and diverse public opinion text data, intelligent analysis methods represented by Natural Language Processing (NLP) have gradually become an important research direction in the field of public opinion risk assessment. Most existing public opinion analysis technologies rely on basic text feature extraction methods and shallow machine learning classification models, which are difficult to efficiently handle the heterogeneity, timeliness, and cross-language nature of the data. Moreover, most of them only stay at the static analysis level of single-language texts and lack in-depth mining and analysis of the dynamic event propagation patterns.

[0003] CN111538888A discloses a network public opinion intensity evolution analysis system based on an active monitoring engine and big data, which can realize active data collection and sentiment analysis, but still has limitations in public opinion risk identification and propagation path prediction, and lacks effective analysis and inference of the dynamic evolution process and causal relationship of risk events.

[0004] CN117290493A discloses a method for monitoring corporate risk public opinion, involving steps such as data preprocessing, analysis and processing, and risk identification. Although it can effectively avoid the tediousness and subjectivity of manual statistical analysis and improve the accuracy of risk assessment, it still has shortcomings in its implementation, such as a relatively simple risk identification method, lack of dynamic modeling of risk propagation time sequence, and insufficient causal relationship mining. In addition, it fails to effectively address the cross-language characteristics of public opinion data and the issue of sensitive information protection, and it also fails to deeply explore the complex propagation paths and inherent causal relationships between public opinion risk events. Therefore, there is still room for improvement in the early warning and precise deployment of intervention measures for risks.

[0005] In summary, existing public opinion analysis technologies based on natural language processing generally suffer from insufficient capabilities in processing heterogeneous data, inadequate mining of risk event propagation paths and causal relationships, and deficiencies in sensitive data protection and cross-language public opinion alignment. Summary of the Invention

[0006] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0007] In view of the aforementioned existing problems, the present invention is proposed.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: real-time capture of public opinion texts from social media, industry forums and government information disclosure platforms based on a unified timestamp, and desensitization of user identity fields using hash pseudo-identifiers to obtain the original text stream;

[0009] A dual-classification weakly supervised framework is constructed based on a text classification model and a semantic clustering model. The risk confidence score is calculated for the original text stream. When the risk confidence score is greater than or equal to a first threshold, the corresponding text is written into the annotation pool to form a risk corpus.

[0010] Based on the entity-relationship-time slice meta-path rule, the risk corpus is mapped into a dynamic heterogeneous risk knowledge graph, and the low-resource language nodes are aligned to a unified semantic space through a dual encoder.

[0011] Apply a multi-head spatiotemporal diffusion Transformer to the subgraph intersecting with the current window in the dynamic heterogeneous risk knowledge graph to obtain the risk propagation vector of the event node;

[0012] The risk propagation vector and the historical baseline vector are input into the causal comparison analysis model to output the early risk warning index. When the early risk warning index is greater than or equal to the second threshold, a risk alarm containing the event path is pushed to the background control center.

[0013] As a preferred embodiment of the risk sentiment analysis method based on natural language processing described in this invention, obtaining the original text stream includes:

[0014] The system calls social media application interfaces, industry forum crawlers, and government information disclosure platform subscription services at minute-level granularity to retrieve public opinion texts in real time.

[0015] The retrieved public opinion texts are sorted according to the source identifier-unified timestamp rule and written into the cache queue;

[0016] The cosine approximation algorithm is used to remove duplicates from texts whose similarity reaches a preset threshold, retaining only the first text.

[0017] After desensitizing and replacing the user identity fields of Uniform Resource Locator, telephone number, and ID card number using irreversible hash mapping, the original text stream is output.

[0018] As a preferred embodiment of the risk sentiment analysis method based on natural language processing described in this invention, a dual-classification weakly supervised framework is constructed based on a text classification model and a semantic clustering model to calculate the risk confidence level of the original text stream, including:

[0019] Call the text classification algorithm trained on large-scale public opinion annotation data to output the risk probability value of each text;

[0020] A semantic clustering algorithm based on sentence vector density is invoked to calculate consistency scores for the same topic cluster;

[0021] The risk confidence level is obtained by linearly combining the risk probability value and the consistency score according to the weighting coefficient.

[0022] As a preferred embodiment of the risk sentiment analysis method based on natural language processing described in this invention, the dual-classification weakly supervised framework includes:

[0023] The text classification algorithm module is used to determine the risk category of text based on a multi-layer attention network.

[0024] The semantic clustering algorithm module is used to cluster sentence vectors based on adaptive density thresholds to generate topic consistency labels;

[0025] The confidence fusion module is used to receive the output results of the text classification algorithm module and the semantic clustering algorithm module and generate risk confidence.

[0026] The threshold adaptive module is used to dynamically adjust the first threshold based on the recall and precision of the previous sliding window.

[0027] As a preferred embodiment of the risk sentiment analysis method based on natural language processing described in this invention, when the risk confidence level is ≥ a first threshold, the corresponding text is written into a labeling pool to form a risk corpus, including:

[0028] When the risk confidence level is greater than or equal to the first threshold, the corresponding text will be marked as risk text;

[0029] Texts that meet the first threshold condition, along with their source, timestamp, risk confidence level, and language type, are written into the risk corpus pool.

[0030] A time decay priority queue is used to manage the capacity of the risk corpus pool. When the capacity reaches the upper limit, the risk corpus with the longest time and the lowest confidence is eliminated first.

[0031] As a preferred embodiment of the risk sentiment analysis method based on natural language processing described in this invention, the risk corpus is mapped into a dynamic heterogeneous risk knowledge graph, including:

[0032] Based on the meta-path rule of subject-risk factor-attribute-time, entity extraction, relation extraction and attribute standardization are performed on the risk corpus;

[0033] The extracted triples are written into the graph database according to fixed ten-minute time slices, and subgraphs corresponding to the time slices are generated.

[0034] The newly added nodes and historical nodes in the current subgraph are matched based on semantic vector similarity. When the similarity exceeds a preset threshold, the newly added nodes and historical nodes in the current subgraph are merged.

[0035] After projecting nodes from languages ​​with limited resources into a unified semantic space through cross-language alignment operations, they are written into the dynamic heterogeneous risk knowledge graph.

[0036] As a preferred embodiment of the risk sentiment analysis method based on natural language processing described in this invention, obtaining the risk propagation vector of an event node includes:

[0037] Extract the subgraph from the dynamic heterogeneous risk knowledge graph that intersects with the current sliding time window;

[0038] Perform a random walk of up to three hops on this subgraph to expand the upstream and downstream nodes;

[0039] The expanded subgraph is input into a multi-head spatiotemporal diffusion Transformer, and each attention head captures the interaction features of entity-relationship, relation-time, and entity-time, respectively.

[0040] The multi-head feature vectors output by each event node are averaged and aggregated to generate the corresponding risk propagation vector.

[0041] As a preferred embodiment of the risk public opinion analysis method based on natural language processing described in this invention, the output risk early warning index includes:

[0042] Retrieve baseline vectors of similar historical events based on event type and geographic tag;

[0043] Input the risk propagation vector and the baseline vector into the causal comparison analysis model to calculate the difference score;

[0044] The risk early warning index is obtained by multiplying the difference score by the event audience size factor.

[0045] As a preferred embodiment of the risk public opinion analysis method based on natural language processing described in this invention, when the early risk warning index is ≥ the second threshold, a risk alarm containing the event path is pushed to the backend control center, specifically including:

[0046] An alarm is triggered when the early warning index of the risk is greater than or equal to the second threshold.

[0047] Generate structured alarm messages that include event identifiers, early risk warning indices, triggering factors, time-regional propagation paths, and confidence levels;

[0048] Alarm messages are pushed to the backend control center through a high-priority message channel.

[0049] The beneficial effects of this invention are as follows: This invention solves the problems of secure collection of heterogeneous multi-source data, efficient risk text recognition, cross-language semantic fusion, dynamic risk propagation analysis, and early causal warning. It effectively improves the automation, accuracy, and reliability of public opinion risk monitoring and early warning, and has high security, universality, and analytical precision. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0051] Figure 1 This is a flowchart illustrating the risk sentiment analysis method based on natural language processing as described in this invention. Detailed Implementation

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0053] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0054] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0055] According to an embodiment of the present invention, in combination Figure 1 The flowchart shown illustrates a risk sentiment analysis method based on natural language processing, which specifically includes the following steps:

[0056] S1. Capture public opinion texts from social media, industry forums, and government information disclosure platforms in real time based on a unified timestamp. Use hash pseudo-identifiers to de-identify user identity fields, obtaining the original text stream. Note that the following points should be noted in this step:

[0057] The system calls social media application interfaces, industry forum crawlers, and government information disclosure platform subscription services at minute-level granularity to retrieve public opinion texts in real time.

[0058] The retrieved public opinion texts are sorted according to the source identifier-unified timestamp rule and written into the cache queue;

[0059] The cosine approximation algorithm is used to remove duplicates from texts with a similarity of 88% or higher, keeping only the first text.

[0060] After desensitizing and replacing the user identity fields of Uniform Resource Locator, telephone number, and ID card number using irreversible hash mapping, the original text stream is output.

[0061] As an example, the deduplication operation specifically includes: calculating sentence vector features for newly added text in the cache queue and retrieving neighboring texts in the local sensitive hash index; if the cosine similarity is higher than 88% and the publication time difference is no more than five minutes, it is considered a duplicate, and only the earlier one is retained.

[0062] In an optional implementation, a method for de-identifying user identity fields using hash pseudo-identifiers specifically includes:

[0063] Concatenate the sensitive field with the random value of the day;

[0064] The digest is calculated using a one-way hash algorithm that conforms to national commercial cryptography standards;

[0065] The summary result is Base64 encoded, and the first sixteen characters are extracted.

[0066] Replace the original sensitive fields with the extracted string to achieve irreversible desensitization.

[0067] It should be noted that the random value is automatically rotated daily and is stored only in the hardware security module to prevent reverse engineering.

[0068] Add content type tags to the deduplicated and anonymized text. Tags include, but are not limited to, short posts, long articles, comments, bullet comments, government replies, and announcement text.

[0069] The final generated structured record includes source identifier, timestamp, content type, de-identified text, and deduplicated text, and is written to a streaming message queue for subsequent risk confidence calculation.

[0070] For example, public opinion texts include at least the main text and comments of Weibo posts, articles and messages on WeChat official accounts, topic posts and replies on industry forums, messages on government information disclosure, and policy announcements.

[0071] Preferably, this step uses a hash pseudo-identifier to de-identify the user identity field to obtain the original text stream, which realizes the unified and standardized collection of heterogeneous and multi-source text data and the effective protection of sensitive information, ensuring the security and consistency of subsequent data processing, avoiding privacy and security risks caused by information leakage, and thus improving the data security and credibility of the entire public opinion analysis system.

[0072] S2. Construct a dual-classification weakly supervised framework based on a text classification model and a semantic clustering model. Calculate the risk confidence score for the original text stream. When the risk confidence score is greater than or equal to a first threshold, write the corresponding text into the annotation pool to form a risk corpus. Note that the following should be noted in this step:

[0073] Call the text classification algorithm trained on large-scale public opinion annotation data to output the risk probability value of each text;

[0074] A semantic clustering algorithm based on sentence vector density is invoked to calculate consistency scores for the same topic cluster;

[0075] The risk confidence level is obtained by linearly combining the risk probability value and the consistency score based on the weighted coefficients.

[0076] Furthermore, the two-class weakly supervised framework includes:

[0077] The text classification algorithm module is used to determine the risk category of text based on a multi-layer attention network.

[0078] The semantic clustering algorithm module is used to cluster sentence vectors based on adaptive density thresholds to generate topic consistency labels;

[0079] The confidence fusion module is used to receive the output results of the text classification algorithm module and the semantic clustering algorithm module and generate risk confidence scores.

[0080] The threshold adaptive module is used to dynamically adjust the first threshold based on the recall and precision of the previous sliding window.

[0081] In an optional implementation, when the risk confidence level is greater than or equal to a first threshold, the corresponding text is written into the annotation pool to form a risk corpus, including:

[0082] When the risk confidence level is greater than or equal to the first threshold, the corresponding text will be marked as risk text;

[0083] The first threshold is set according to the following rules: the median of the risk confidence distribution of the most recent N windows is used as the base value. When the recall rate is lower than the target value, it is reduced by 5% and when the false alarm rate is higher than the target value, it is increased by 5%. The value range of the first threshold is limited to 0.40 to 0.85.

[0084] Texts that meet the first threshold condition, along with their source, timestamp, risk confidence level, and language type, are written into the risk corpus pool.

[0085] A time decay priority queue is used to manage the capacity of the risk corpus pool. When the capacity reaches the upper limit, the risk corpus with the longest time and the lowest confidence is eliminated first.

[0086] Preferably, by automating the identification and preliminary filtering of risky texts, the cost of manual annotation can be reduced while efficiently capturing public opinion information that truly poses a risk. This significantly improves the efficiency and accuracy of risk identification and effectively reduces the subjectivity and uncertainty of manual intervention.

[0087] S3. Based on the entity-relationship-time slice meta-path rule, map the risk corpus into a dynamic heterogeneous risk knowledge graph, and align the low-resource language nodes to a unified semantic space using a dual encoder. Note that the following points should be noted in this step:

[0088] Based on the meta-path rule of subject-risk factor-attribute-time, entity extraction, relation extraction and attribute standardization are performed on the risk corpus;

[0089] The extracted triples are written into the graph database according to fixed ten-minute time slices, and subgraphs corresponding to the time slices are generated.

[0090] The new node in the current subgraph is matched with the historical node based on semantic vector similarity. When the similarity exceeds a preset threshold (e.g., 85%), the new node in the current subgraph is merged with the historical node; otherwise, it is retained as a new node.

[0091] After projecting nodes from languages ​​with limited resources into a unified semantic space through cross-language alignment operations, they are written into a dynamic heterogeneous risk knowledge graph.

[0092] As an example, the meta-path rules for subject-risk factor-attribute-time include:

[0093] Subject: refers to the enterprise, institution, public figure, or geographical area that generates or is affected;

[0094] Risk factors: These refer to keywords or event tags that trigger negative impacts, such as "production stoppage" or "broken capital chain."

[0095] Attributes: These refer to descriptive information such as risk level, sentiment polarity, and confidence level.

[0096] Time: refers to the time of publication of the text or the time expressed explicitly in the text.

[0097] In an optional implementation, nodes of low-resource languages ​​(such as French and Hindi) are aligned to a unified semantic space using a dual encoder. This embodiment employs dual-channel semantic coding, specifically including:

[0098] First channel: The low-resource language encoder converts the original text into a semantic vector;

[0099] The second channel: First, the original text is automatically transcribed into Chinese by learning machine translation, and then a reference vector is generated by a high-resource language encoder;

[0100] By minimizing the distance between two channel vectors, the low-resource vector is aligned to the Chinese-English shared space;

[0101] After alignment, with a similarity of no less than 80%, low-resource nodes are mapped to corresponding high-resource nodes.

[0102] If the threshold (80%) is not reached, create a cross-language link edge;

[0103] The merged and aligned subgraphs are written to the graph database, and the time slice number and data batch are recorded in the graph header.

[0104] It should be noted that this step, through the fusion of multi-dimensional information and the effective mapping of cross-language nodes, can intuitively and dynamically reveal the complex relationship network in risk public opinion, which facilitates the tracking and analysis of subsequent risk dissemination. Thus, while improving the accuracy of information fusion, it solves the language barrier problem in multilingual public opinion analysis.

[0105] S4. Apply a multi-head spatiotemporal diffusion Transformer to the subgraphs intersecting with the current window in the dynamic heterogeneous risk knowledge graph to obtain the risk propagation vectors of event nodes. Note that the following points should be noted in this step:

[0106] Extract the subgraph from the dynamic heterogeneous risk knowledge graph that intersects with the current sliding time window;

[0107] Perform a random walk of up to three hops on this subgraph to expand the upstream and downstream nodes;

[0108] The expanded subgraph is input into a multi-head spatiotemporal diffusion Transformer, and each attention head captures the interaction features of entity-relationship, relation-time, and entity-time, respectively.

[0109] The multi-head feature vectors output by each event node are averaged and aggregated to generate the corresponding risk propagation vector.

[0110] Preferably, this step, by effectively simulating and capturing the spread and diffusion trends of risk public opinion in the time dimension and relationship network, is conducive to the timely deployment of risk warning and disposal measures, and significantly improves the accuracy of risk prediction and tracking analysis.

[0111] S5. Input the risk propagation vector and historical baseline vector into the causal comparison analysis model, and output the early risk warning index. When the early risk warning index is greater than or equal to the second threshold, push a risk alarm containing the event path to the back-end control center. Note that the following should be noted in this step:

[0112] Based on the topic and location tags of the current event, retrieve the risk propagation vector set of similar events from the historical event database, and calculate the weighted average of the samples from the most recent three months based on the generation period to obtain the historical baseline vector;

[0113] Input the risk propagation vector in the current window and the historical baseline vector into the causal comparison analysis model. Through structured causal inference, compare the degree of deviation between the two in terms of entity relationship, time diffusion speed and emotional polarity, and output the difference score.

[0114] The risk early warning index is obtained by linearly combining the difference score with the event audience size factor;

[0115] An alarm is triggered when the early risk warning index is greater than or equal to the second threshold.

[0116] The rule for setting the second threshold is as follows: take the 90th digit of the distribution of the warning index of the most recent M windows as the initial value. If two consecutive windows have missed reports, the threshold will be reduced by 3%. If the false alarm rate exceeds the upper limit, the threshold will be increased by 3%. The second threshold is limited to the range of 60 to 90.

[0117] Generate structured alarm messages that include event identifiers, early risk warning indices, triggering factors, time-regional propagation paths, and confidence levels;

[0118] Alarm messages are pushed to the backend control center through a high-priority message channel.

[0119] As an example, the causal comparison analysis model constructed in this embodiment is used to calculate the early warning index of risk. The mathematical expression formula of this model is as follows:

[0120]

[0121] Where ε is the early warning index for risk, ranging from [0, 100], Θ is the time span within the current sliding window, Ω is the total number of dimensions contained in the risk propagation vector, and P k (t) represents the k-th dimension of risk propagation at time point t, B k(t) represents the historical baseline component of the k-th dimension at time point t, Λ represents the degree of difference, Ψ represents the smoothing constant, Γ represents the emotional fluctuation adjustment coefficient, ∑(t) represents the current window emotional polarity variance, Φ represents the event audience size factor, and Δ represents the audience size index.

[0122] When ε belongs to the [0,40] interval, it indicates that the public opinion risk is low and can be recorded and observed.

[0123] When ε belongs to the interval [40, 70], it indicates that public opinion has entered an early warning state and requires manual tracking.

[0124] When ε falls within the range of [70, 100], it indicates that the public opinion event is of high risk, and an alert will be sent immediately.

[0125] It should also be noted that this embodiment, by introducing a causal comparison analysis method, achieves a differentiated assessment between risk propagation factors and historical risk scenarios, strengthens the scientific and objective nature of early warning decisions, and thus provides decision-makers with a more effective basis for risk response.

[0126] The aforementioned text classification algorithm trained on large-scale public opinion annotation data and semantic clustering algorithm based on sentence vector density clustering can be implemented using existing technologies, and will not be elaborated on in this example.

[0127] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A risk sentiment analysis method based on natural language processing, characterized in that, include: Based on a unified timestamp, the public opinion texts from social media, industry forums, and government information disclosure platforms are captured in real time. The user identity fields are desensitized using hash pseudo-identifiers to obtain the original text stream. A dual-classification weakly supervised framework is constructed based on a text classification model and a semantic clustering model. The risk confidence score is calculated for the original text stream. When the risk confidence score is greater than or equal to a first threshold, the corresponding text is written into the annotation pool to form a risk corpus. Based on the entity-relationship-time slice meta-path rule, the risk corpus is mapped into a dynamic heterogeneous risk knowledge graph, and the low-resource language nodes are aligned to a unified semantic space through a dual encoder. Apply a multi-head spatiotemporal diffusion Transformer to the subgraph intersecting with the current window in the dynamic heterogeneous risk knowledge graph to obtain the risk propagation vector of the event node; The risk propagation vector and the historical baseline vector are input into the causal comparison analysis model, and the early risk warning index is output. When the early risk warning index is greater than or equal to the second threshold, a risk alarm containing the event path is pushed to the background control center.

2. The risk sentiment analysis method based on natural language processing according to claim 1, characterized in that, The process of obtaining the original text stream includes: The system calls social media application interfaces, industry forum crawlers, and government information disclosure platform subscription services at minute-level granularity to retrieve public opinion texts in real time. The retrieved public opinion texts are sorted according to the source identifier-unified timestamp rule and written into the cache queue; The cosine approximation algorithm is used to remove duplicates from texts whose similarity reaches a preset threshold, retaining only the first text. After desensitizing and replacing the user identity fields of Uniform Resource Locator, telephone number, and ID card number using irreversible hash mapping, the original text stream is output.

3. The risk sentiment analysis method based on natural language processing according to claim 1 or 2, characterized in that, A dual-classification weakly supervised framework is constructed based on a text classification model and a semantic clustering model to calculate the risk confidence of the original text stream, including: Call the text classification algorithm trained on large-scale public opinion annotation data to output the risk probability value of each text; A semantic clustering algorithm based on sentence vector density is invoked to calculate consistency scores for the same topic cluster; The risk confidence level is obtained by linearly combining the risk probability value and the consistency score according to the weighting coefficient.

4. The risk sentiment analysis method based on natural language processing according to claim 3, characterized in that, The dual-class weak supervision framework includes: The text classification algorithm module is used to determine the risk category of text based on a multi-layer attention network. The semantic clustering algorithm module is used to cluster sentence vectors based on adaptive density thresholds to generate topic consistency labels; The confidence fusion module is used to receive the output results of the text classification algorithm module and the semantic clustering algorithm module and generate risk confidence. The threshold adaptive module is used to dynamically adjust the first threshold based on the recall and precision of the previous sliding window.

5. The risk sentiment analysis method based on natural language processing according to claim 3, characterized in that, When the risk confidence level is greater than or equal to the first threshold, the corresponding text is written into the annotation pool to form a risk corpus, including: When the risk confidence level is greater than or equal to the first threshold, the corresponding text will be marked as risk text; Texts that meet the first threshold condition, along with their source, timestamp, risk confidence level, and language type, are written into the risk corpus pool. A time decay priority queue is used to manage the capacity of the risk corpus pool. When the capacity reaches the upper limit, the risk corpus with the longest time and the lowest confidence is eliminated first.

6. The risk sentiment analysis method based on natural language processing according to claim 5, characterized in that, Mapping the aforementioned risk corpus into a dynamic heterogeneous risk knowledge graph includes: Based on the meta-path rule of subject-risk factor-attribute-time, entity extraction, relation extraction and attribute standardization are performed on the risk corpus; The extracted triples are written into the graph database according to fixed ten-minute time slices, and subgraphs corresponding to the time slices are generated. The newly added nodes and historical nodes in the current subgraph are matched based on semantic vector similarity. When the similarity exceeds a preset threshold, the newly added nodes and historical nodes in the current subgraph are merged. After projecting nodes from languages ​​with limited resources into a unified semantic space through cross-language alignment operations, they are written into the dynamic heterogeneous risk knowledge graph.

7. The risk sentiment analysis method based on natural language processing according to claim 6, characterized in that, The obtained risk propagation vector of the event node includes: Extract the subgraph from the dynamic heterogeneous risk knowledge graph that intersects with the current sliding time window; Perform a random walk of up to three hops on this subgraph to expand the upstream and downstream nodes; The expanded subgraph is input into a multi-head spatiotemporal diffusion Transformer, and each attention head captures the interaction features of entity-relationship, relation-time, and entity-time, respectively. The multi-head feature vectors output by each event node are averaged and aggregated to generate the corresponding risk propagation vector.

8. The risk sentiment analysis method based on natural language processing according to claim 7, characterized in that, The output risk early warning index includes: Retrieve baseline vectors of similar historical events based on event type and geographic tag; Input the risk propagation vector and the baseline vector into the causal comparison analysis model to calculate the difference score; The risk early warning index is obtained by multiplying the difference score by the event audience size factor.

9. The risk sentiment analysis method based on natural language processing according to claim 8, characterized in that, When the early risk warning index is greater than or equal to the second threshold, a risk alarm containing the event path is pushed to the backend control center, specifically including: An alarm is triggered when the early warning index of the risk is greater than or equal to the second threshold. Generate structured alarm messages that include event identifiers, early risk warning indices, triggering factors, time-regional propagation paths, and confidence levels; Alarm messages are pushed to the backend control center through a high-priority message channel.

Images