Negative public opinion situation awareness risk identification method, device and equipment

Abstract

The invention discloses a negative public opinion situation awareness risk identification method, device and equipment, and the method comprises the steps: obtaining news public opinion data through responding to an identification request sent by a request end; predicting the news public opinion data by adopting a risk emotion prediction model, and determining a corresponding risk type probability and emotion information; weighting according to the risk type probability and the emotion information, and constructing a negative public opinion index; performing regression analysis on the negative public opinion indexes, and determining an emotion risk index and an information risk index; and according to the negative public opinion index, the emotion risk index and the information risk index, public opinion risk description information corresponding to the request end is determined, so that public opinion risk prediction and description are quickly and accurately carried out, and a user is helped to carry out risk early warning and risk defense on a company where the request end is located.

Description

Technical Field

[0001] The present invention relates to the field of risk identification technology, and in particular to a method, device and equipment for identifying risks of negative public opinion situation awareness. Background Art

[0002] In recent years, the rise of Internet media and self-media has changed the previous pattern of financial news being reported by only a few mainstream media, and the financial media ecosystem has undergone tremendous changes. Internet media has fast transmission, high publication frequency and strong uncontrollability. In the era of information explosion, its massive information has both advantages and disadvantages on the securities market. On the one hand, it can increase the proportion of informed traders and alleviate information asymmetry to improve market efficiency; on the other hand, information chaos will interfere with investors' judgment, and the subjective emotions of the media can easily trigger irrational behavior of investors, leading to violent fluctuations in the stock market.

[0003] Internet media is less constrained, objectivity is easily lost, and subjective emotions are more obvious. In the Internet era, retail investors mostly get news through the Internet, and media emotions can easily lead to changes in investor emotions and be reflected in stock prices. Negative news has a greater impact on stock prices, and negative media emotions are more likely to cause stock price fluctuations.

[0004] The existing stock sentiment aggregation method based on media news data provides information assistance to investors and quantitative fund companies through steps such as crawling news, calculating heat, preprocessing, comprehensive sentiment and theme analysis, etc. However, in the Internet era, media reprints quickly. In addition to gathering public opinion news, how to provide investors with accurate and readable risk information to help them quickly understand and make decisions, explore negative news risks, and prevent negative public opinion from causing systemic financial risks has become the key. Summary of the invention

[0005] The present invention provides a method, device and equipment for identifying risks of negative public opinion situation awareness, which solves the technical problem of how to provide investors with accurate and readable risk information to help them quickly understand and make decisions.

[0006] The first aspect of the present invention provides a method for identifying negative public opinion situation awareness risks, comprising:

[0007] Respond to the identification request sent by the requester and obtain news and public opinion data;

[0008] Using a risk sentiment prediction model to predict the news public opinion data, and determine the corresponding risk type probability and sentiment information;

[0009] Constructing a negative public opinion index by weighting the risk type probability and the sentiment information;

[0010] Performing regression analysis on the negative public opinion indicators to determine emotional risk indicators and information risk indicators;

[0011] According to the negative public opinion index, the emotional risk index and the information risk index, the public opinion risk description information corresponding to the requesting end is determined.

[0012] Optionally, the method further comprises:

[0013] Obtain historical news and public opinion data;

[0014] Preprocessing the historical news public opinion data to obtain multiple key words and form a bag-of-words model;

[0015] One-hot encoding is performed on the bag-of-words model to generate a document vector and annotate risk labels and sentiment labels;

[0016] The document vector is used to train multiple preset machine learning models and respectively calculate prediction effect indicators, and a risk sentiment prediction model is selected according to the prediction effect indicators.

[0017] Optionally, the adopting of the document vector to train multiple preset machine learning models and respectively calculating prediction effect indicators, and selecting a risk sentiment prediction model according to the prediction effect indicators, includes:

[0018] Inputting the document vectors into a plurality of preset machine learning models respectively to obtain prediction results;

[0019] Compare the prediction results with the risk label and the emotion label respectively, and calculate the loss function value;

[0020] Optimize the parameters of each of the machine learning models according to the loss function value until the loss function value is less than a preset loss threshold, thereby obtaining a plurality of intermediate machine learning models;

[0021] Calculate the prediction effect indicators corresponding to each of the intermediate machine learning models, and select the intermediate machine learning model whose prediction effect indicators meet the selection conditions as the risk sentiment prediction model.

[0022] Optionally, the method further comprises:

[0023] When multiple negative public opinion indicators are obtained, the requesting ends are sorted according to the negative public opinion indicators to obtain a requesting end sequence;

[0024] Divide the request end sequence according to a preset number of groups to obtain multiple request end combinations;

[0025] A resource configuration strategy is created to allocate resources to each of the request end combinations according to preset weights.

[0026] Optionally, the adopting a risk sentiment prediction model to predict the news public opinion data to determine the corresponding risk type probability and sentiment information includes:

[0027] Convert the news and public opinion data into a vector to be input;

[0028] Inputting the vector to be input into the risk sentiment prediction model, and determining the risk type probability corresponding to the vector to be input through the risk prediction branch of the risk sentiment prediction model;

[0029] The emotion probability corresponding to the input vector is determined as emotion information through the emotion prediction branch of the risk emotion prediction model.

[0030] Optionally, weighting the risk type probability and the sentiment information to construct a negative public opinion index includes:

[0031] The risk weight corresponding to the probability of each risk type is determined by using the analytic hierarchy process;

[0032] Constructing a negative public opinion index according to the risk weight, the risk type probability and the sentiment information;

[0033] Negative public opinion indicators are:

[0034]

[0035] Among them, risk k,t,m is the probability value of risk type m belonging to requester k on day t; n is the total number of news and public opinion data of requester k on day t; is the sentiment information of the a-th news public opinion data of the requesting end k on the t-th day; is the risk type probability of the a-th news public opinion data of requester k on day t belonging to risk type m; θ m is the risk weight of risk type m; b is the total number of risk types; MRI k,t is the total public opinion risk value of the requesting end k on the tth day.

[0036] Optionally, performing regression analysis on the negative public opinion index to determine the emotional risk index and the information risk index includes:

[0037] Using the negative public opinion index as an independent variable and the number of daily comments as a dependent variable, a regression analysis is performed to construct a regression analysis formula;

[0038] Transforming the regression analysis formula to determine the emotional risk index and the information risk index;

[0039] The regression analysis formula is:

[0040] MRI k,t =a k +b k ×Tpostk,t +e k,t

[0041] The sentiment risk indicators are:

[0042]

[0043] The information risk indicator MRSI k,t for:

[0044] MRSI k,t =MRI k,t -MRII k,t

[0045] Among them, MRII k,t is the information risk index of requester k on day t, MRSI k,t is the sentiment risk index of requester k on day t, a k and b k are the regression coefficients of the request end k, Tpost k,t is the number of daily comments related to requester k on day t, e k,t is the error term of requester k on day t, and is the estimated value of the regression coefficient.

[0046] Optionally, determining the public opinion risk description information corresponding to the requesting end according to the negative public opinion index, the emotional risk index and the information risk index includes:

[0047] Calculate the first lagged rate of return according to the negative public opinion index and the preset request end parameters;

[0048] Calculate a second lagged rate of return according to the sentiment risk indicator and the information risk indicator in combination with the requesting end parameter;

[0049] The first lagged rate of return and the second lagged rate of return are used to determine the public opinion risk description information corresponding to the requesting end.

[0050] A second aspect of the present invention provides a negative public opinion situation awareness risk identification device, comprising:

[0051] The data acquisition module is used to respond to the identification request sent by the request end and obtain news and public opinion data;

[0052] A risk sentiment determination module is used to predict the news public opinion data using a risk sentiment prediction model to determine the corresponding risk type probability and sentiment information;

[0053] A negative public opinion index determination module, used to construct a negative public opinion index by weighting the risk type probability and the emotional information;

[0054] An indicator regression analysis module is used to perform regression analysis on the negative public opinion indicators to determine emotional risk indicators and information risk indicators;

[0055] The risk identification module is used to determine the public opinion risk description information corresponding to the requesting end according to the negative public opinion index, the emotional risk index and the information risk index.

[0056] The third aspect of the present invention provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of the negative public opinion situation awareness risk identification method as described in any one of the first aspect of the present invention.

[0057] It can be seen from the above technical solutions that the present invention has the following advantages:

[0058] The present invention obtains news public opinion data by responding to the identification request sent by the requesting end; adopts the risk emotion prediction model to predict the news public opinion data to determine the corresponding risk type probability and emotion information; weights the risk type probability and emotion information to construct a negative public opinion index; performs regression analysis on the negative public opinion index to determine the emotion risk index and the information risk index; determines the public opinion risk description information corresponding to the requesting end based on the negative public opinion index, the emotion risk index and the information risk index, thereby quickly and accurately predicting the public opinion risk and describing it, which helps users to carry out risk warning and risk defense for the company where the requesting end is located. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, 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 invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0060] Figure 1 A flowchart of the steps of a method for identifying the risk of negative public opinion situation awareness provided by an embodiment of the present invention;

[0061] Figure 2 A ROC curve diagram drawn for the Naive Bayes method in an embodiment of the present invention;

[0062] Figure 3 The average monthly MRI and the discounted rate of return graph provided by the embodiment of the present invention;

[0063] Figure 4 A monthly MRI and yield line chart provided by an embodiment of the present invention;

[0064] Figure 5 A schematic diagram of the impact of MRI on stock prices in subsequent days provided by an embodiment of the present invention;

[0065] Figure 6 A schematic diagram of an autocorrelation coefficient of MRI provided by an embodiment of the present invention;

[0066] Figure 7 A structural block diagram of a negative public opinion situation awareness risk identification device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0067] The prior art discloses a method for gathering sentiments of individual stocks based on media news data, which includes the following steps: first crawling news information, forming news documents and storing them in a document storage database, then calculating the popularity of each article and removing duplicate documents; preprocessing the content items in the news documents to form a text set, performing sentiment analysis and topic analysis on each text set to form a set of two-tuples, performing text topic clustering and grouping, integrating all relevant financial news, forming a set of three-tuples based on individual stocks, and finally, gathering the above results with individual stocks as the core, and displaying the results to users. The scheme can provide investors in the financial market with accurate and highly readable concise topic sentiment information, help investors spend less time to understand and make better investment decisions, and provide important auxiliary information for prediction models for quantitative fund companies. According to the agenda setting theory of communication, the media increases the importance of the information among the public by reporting news data with a biased repetitiveness. In the Internet era, the speed of reprinting between various media is very fast. Therefore, in addition to aggregating and classifying public opinion news, it is also crucial to provide investors with more accurate and readable risk information, help investors understand and make investment decisions in a shorter time, explore the risks hidden in negative news, and prevent the spread of negative public opinion risks from causing systemic financial risks.

[0068] To this end, the embodiments of the present invention provide a method, device and equipment for identifying the risk of negative public opinion situation awareness, which is used to solve the technical problem of how to provide investors with accurate and readable risk information to help them quickly understand and make decisions. The risks contained in negative public opinion are strictly monitored to provide investors with more accurate and readable risk information, helping investors to understand and make investment decisions in a shorter time, and preventing the spread of negative public opinion risks from causing systemic financial risks.

[0069] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0070] See also Figure 1 , Figure 1 A flowchart of the steps of a method for identifying negative public opinion situation awareness risks provided by an embodiment of the present invention.

[0071] The present invention provides a method for identifying negative public opinion situation awareness risks, comprising:

[0072] Step 101, responding to the identification request sent by the requesting end, and obtaining news public opinion data;

[0073] The requesting end refers to a client that has a demand for risk identification and has a user identifier, which can be used to identify an enterprise or user.

[0074] News and public opinion data refers to relevant data involving the enterprise or user, including but not limited to news title, full text of the news, news release time, news release source, and corporate stock price. Among them, the news release source is not limited to a few news platforms, but should cover the entire network, including official reports, financial media platforms, news public accounts, etc.

[0075] In this embodiment, after receiving the identification request sent by the requesting end, the news and public opinion data of the day or the historical time period is obtained to obtain the data basis for subsequent risk identification.

[0076] In one example of the present invention, the method further includes S11-S14:

[0077] S11. Obtain historical news and public opinion data;

[0078] S12, pre-processing the historical news public opinion data to obtain multiple key words and form a bag-of-words model;

[0079] In this embodiment, the data type of historical news public opinion data is basically the same as that of news public opinion data, except that its data source is a historical time period. After obtaining the historical news public opinion data related to the request end, it is preprocessed, such as data cleaning, text segmentation, part-of-speech tagging, and removal of stop words to obtain multiple key segmentations. Taking each article in the historical news public opinion data as a unit, its related multiple key segmentations are combined into a bag-of-words model, waiting for subsequent processing.

[0080] In the preprocessing operation of text segmentation, in order to improve the accuracy of segmentation, relevant proper nouns, including company names and financial terms, are imported. In the preprocessing operation of removing stop words, articles and conjunctions that have no information content in the text are removed. Removing stop words can further improve the accuracy of text mining.

[0081] For example, after classifying the risk of historical news public opinion data, the TF-IDF natural language processing method can be used to extract the top 20 keywords for each type of risk, as shown in Table 1 below:

[0082] Table 1

[0083]

[0084] Although there are duplicate keywords for very few risks, the keywords for each type of risk can reflect the risks involved. For operating risks, we can see from "product", "project", "business" and so on that they mainly refer to problems with the company's product projects during the operation process; "repayment", "overdue", "debtor", "investor" and so on indicate problems with the company's financing and investment business; and broader words such as "finance", "supply chain", "equity", "funds" are also classified as operating risks. For accounting and financial risks, "%" indicates that comparisons with previous years often appear in news reports, such as net profit, assets, and liabilities growth; "liabilities", "asset-liability ratio", "assets", "net profit", "goodwill", and "revenue" are all accounting items and accounting indicators in the financial statements, which can also directly explain the accounting and financial risks contained in the company; at the same time, it is worth noting that the words "taxation", "deferral", "income tax", "taxes", and "tax bureau" in accounting and financial risks all indicate that taxation is also a major financial issue that is easy to attract attention. Regarding market risks, media reports mainly focus on equity risks, including "shareholders'" "holding increase" and "holding reduction", "equity" "repurchase" and "transfer", stock "pledge", etc.; it also includes the company's mergers and acquisitions, such as "shell" listing, "acquisition", "control", "restructuring" and other words. Regarding legal risks, media reports focus on "judgment" results, "litigation" and "cases" and other words that clearly represent legal aspects; and these dispute cases mainly include: "violations" category, "contract disputes" category, and bond debt "repayment" category.

[0085] S13, perform one-hot encoding on the bag-of-words model, generate document vectors and annotate risk labels and sentiment labels;

[0086] After obtaining the bag-of-words model corresponding to each article, one-hot encoding is used to generate a document vector.

[0087] In addition, the encoding method of the document vector can also refer to the natural language processing method, such as the term frequency-inverse document frequency method TF-IDF and the word2vec method.

[0088] At the same time, the document vectors are annotated with risk labels and sentiment labels.

[0089] Specifically, when labeling public opinion news data with risk tags, the risks involved in the public opinion news data can be divided into: business risk, accounting and financial risk, stock market risk, legal and policy risk, other risks and no risk. The definition of each type of risk is shown in Table 2 below:

[0090] Table 2

[0091]

[0092] As for the annotation of emotional tags, it involves the positive and negative emotional judgment of public opinion news data. This embodiment uses the title of the document for modeling. This is because the title of the document is usually a one-sentence summary of the news report, which can condense the most important information of public opinion and usually has more obvious emotional characteristics. In order to extract the key information of the title, the term frequency-inverse document frequency method TF-IDF is used to extract the first 10 keywords of each title (which can be less than ten words), and the extracted parts of speech include four categories: place names, nouns, gerunds, and verbs. All keywords are combined into a bag of words model, and one-hot encoding is used to represent the spatial vector of each article.

[0093] For the identification of public opinion risk classification, this embodiment uses article content as the data source, because the article content will report and analyze the company's news events in detail, so it contains the most information. In order to better summarize the meaning of each article, the TF-IDF method is used to obtain the first 50 keywords of each article, and all keywords are combined into a bag-of-words model, and the space vector of each article is represented by one-hot encoding.

[0094] S14. Use document vectors to train multiple preset machine learning models and calculate prediction effect indicators respectively, and select risk sentiment prediction models according to the prediction effect indicators.

[0095] Further, S14 may include the following sub-steps:

[0096] Input document vectors into multiple preset machine learning models to obtain prediction results;

[0097] Compare the prediction results with the risk label and sentiment label respectively, and calculate the loss function value;

[0098] Optimize the parameters of each machine learning model according to the loss function value until the loss function value is less than a preset loss threshold, and obtain multiple intermediate machine learning models;

[0099] Calculate the prediction effect indicators corresponding to each intermediate machine learning model, and select the intermediate machine learning model whose prediction effect indicators meet the selection conditions as the risk sentiment prediction model.

[0100] Machine learning models include but are not limited to naive Bayes model, decision tree, random forest, Adaboost and support vector machine, etc. Prediction effect indicators include but are not limited to precision, recall, F1 score, ROC curve and AUC value, etc.

[0101] In this embodiment, after completing the initialization of the machine learning models of various model architectures, the labeled document vectors are input into multiple machine learning models to obtain prediction results under different model architectures, including risk prediction results and sentiment prediction results.

[0102] After obtaining the prediction results, the risk prediction results are compared with the annotated risk labels, and the emotion prediction results are compared with the emotion labels, and the loss function value is calculated according to the comparison results. If the loss function value is greater than or equal to the preset loss threshold, the gradient descent method or other parameter adjustment methods are used to tune the process parameters of each machine learning model until the loss function value is less than the preset loss threshold, at which time multiple intermediate machine learning models are obtained.

[0103] Furthermore, according to the above-mentioned prediction effect indicator types, the prediction effect indicators of each intermediate machine learning model are calculated respectively, and the intermediate machine learning model whose prediction effect indicator reflects the model performance best is selected as the risk sentiment prediction model.

[0104] For example, the following table 3 shows an example of using precision, recall, and F1 score as prediction effect indicators:

[0105] Table 3

[0106] Accuracy Accuracy Recall F1-score Naive Bayes 0.8527 0.8180 0.8482 0.8256 Decision Tree 0.8409 0.8017 0.8205 0.8090 Random Forest 0.8802 0.8624 0.8745 0.8675 Adaboost 0.7466 0.7862 0.7081 0.7153 Support Vector Machine 0.6189 0.8264 0.4806 0.5315

[0107] Among all machine learning models, Naive Bayes, Decision Tree, Random Forest, and Adaboost performed better. Among them, Random Forest had the highest accuracy, reaching 88.02%, and the worst performance was the Support Vector Machine model, which was only 61.89%. The prediction effects of Adaboost and Support Vector Machine were poor, especially the recall rate of Support Vector Machine was only 48.06%, which means that the Support Vector Machine model predicted more false negatives.

[0108] This example also needs to predict the sentiment value. The precision, recall, F1 score, and AUC value are used to evaluate the prediction effectiveness of the machine learning model. The prediction results of the sentiment are shown in Table 4. Figure 2 Plot the ROC curve for the Naive Bayes method.

[0109] Table 4

[0110] Accuracy Accuracy Recall F1-score AUC Naive Bayes 0.9273 0.9073 0.9278 0.9164 0.9750 Decision Tree 0.8821 0.8880 0.8282 0.8502 0.9313 Random Forest 0.9037 0.8993 0.8691 0.8821 0.9575 Adaboost 0.8389 0.8106 0.8479 0.8220 0.9244 Support Vector Machine 0.7878 0.8647 0.6552 0.6714 0.9275

[0111] The Naive Bayes model has the best prediction effect among all machine learning models, with all indicators exceeding 90%, an accuracy of 92.73%, and an AUC value of 97.5%, which also shows the practicality of the Naive Bayes model in text analysis. The support vector machine model has the worst effect, with an accuracy of only 78.78%, which may be because the support vector machine is not suitable for text classification problems with sparse vectors. The decision tree, as the most basic machine learning algorithm, also presents very robust results, with an accuracy of 88.21%. Random forest, as a classic ensemble learning algorithm based on bagging, has better overall running results than decision trees, with an accuracy of 90.37%, because bootstrap sampling voting of multiple decision trees is used here. Adaboost, as another classic ensemble learning algorithm based on boosting, has a less effective running result than decision trees, with an accuracy of only 83.89%, which may be because Adaboost pursues the accuracy of the training set too much and overfits on the test set.

[0112] Based on the above results, the Naive Bayes model can be selected as the emotion prediction branch in the risk emotion prediction model, and the Random Forest model can be selected as the risk prediction branch in the risk emotion prediction model.

[0113] Step 102, using the risk sentiment prediction model to predict the news public opinion data, and determine the corresponding risk type probability and sentiment information;

[0114] In one example of the present invention, step 102 may include the following sub-steps:

[0115] Convert news and public opinion data into vectors to be input;

[0116] Input the input vector to the risk sentiment prediction model, and determine the risk type probability corresponding to the input vector through the risk prediction branch of the risk sentiment prediction model;

[0117] The emotion prediction branch of the risk emotion prediction model is used to determine the emotion probability corresponding to the input vector as emotion information.

[0118] In this embodiment, the news public opinion data is preprocessed and converted into a vector to be input in vector form. The specific process can be seen in S12. The vector to be input is input into the risk sentiment prediction model, and the risk prediction branch of the risk sentiment prediction model determines the risk type probability corresponding to the vector to be input based on the above-mentioned learned parameters and patterns, such as the market risk probability of 0.4, the reputation risk probability of 0.3, and the other risk probability of 0.3. At the same time, the emotion prediction branch also processes the input vector and outputs the corresponding emotion probability as emotion information, such as the probability of positive emotion 0.2, the probability of negative emotion 0.7, and the probability of neutral emotion 0.1.

[0119] Step 103, weighting the risk type probability and sentiment information to construct a negative public opinion index;

[0120] In an example of the present invention, step 103 may include the following sub-steps:

[0121] The analytic hierarchy process is used to determine the risk weight corresponding to the probability of each risk type;

[0122] Construct negative public opinion indicators according to risk weight, risk type probability and sentiment information;

[0123] Negative public opinion indicators are:

[0124]

[0125] Among them, risk k,t,m is the probability value of risk type m belonging to requester k on day t; n is the total number of news and public opinion data of requester k on day t; is the sentiment information of the a-th news public opinion data of the requesting end k on the t-th day; is the risk type probability of the a-th news public opinion data of requester k on day t belonging to risk type m; θ m is the risk weight of risk type m; b is the total number of risk types; MRI k,t is the total public opinion risk value of the requesting end k on the tth day.

[0126] In this embodiment, the risk weight corresponding to the probability of each risk type can be determined by the hierarchical analysis method. Specifically, on the basis of classifying risks, the impact caused by each type of risk is different. For example, accounting and financial risks or judicial policy risks have a greater impact on the company, and these indicators should have a greater weight, while other risks should have a smaller weight. In order to scientifically allocate the weight of each risk category, the hierarchical analysis method is used for weighting. Experts in the field are used to weight different risk types. After the judgment matrix passes the consistency test, the specific weight of each risk is obtained, and the average is used to obtain the overall risk weight. The specific weights are shown in Table 5 below.

[0127] Table 5

[0128] Risk Type Risk Weight Financial Market Risks: 0.235046237210725 Accounting and Financial Risks: 0.310145369396954 Operational business risks: 0.160831663858335 Legal compliance risks: 0.24377521732951 Other risks: 0.0502015122044757

[0129] After obtaining the risk type probability, sentiment information and risk weight of each article, a negative public opinion index is constructed based on this data.

[0130] Step 104, performing regression analysis on the negative public opinion index to determine the emotional risk index and the information risk index;

[0131] In one example of the present invention, step 104 may include the following sub-steps:

[0132] Using negative public opinion indicators as independent variables and the number of daily comments as dependent variables, regression analysis was performed to construct a regression analysis formula;

[0133] Transform the regression analysis formula to determine the emotional risk index and information risk index.

[0134] MRI (Media Risk Index) can be used to measure the media's reaction and influence on specific events and subjects (such as companies, topics, etc.) in public opinion analysis, communication studies and other related research or business scenarios. For example, the index is constructed by analyzing the number of media reports, the scope of coverage, the emotional tendency of reports and other factors, and presents the comprehensive response of the media in a numerical form, which helps to judge the public opinion risks and social attention faced by the corresponding events or subjects. It is an indicator that combines the information theory and emotion theory of the media in a unified framework. The separation of information and emotion can be achieved through investor attention. The information in MRI refers to all negative information related to the company on that day, including the degree of negativity of news reports, the weight of information related to risk types, and even the importance of information brought by repeated reports. The emotion in MRI refers to the overreaction and irrational behavior of investors caused by media hype and repeated reports. The media's hot discussion will bring about hot discussions on social media and high attention from investors, and investors' attention will drive their overreaction and irrational emotions. Investor emotions are divided into careful emotions and focused emotions through investor attention. In this embodiment, MRI can be divided into emotional risk indicators (Media-Risk Sentiment Indicator, MRSI) that reflect emotions and information risk indicators (Media-Risk Information Index, MRII) that reflect information through investor attention.

[0135] In actual implementation, investors' attention is usually measured through search indexes, such as Google Index, Baidu Index, etc. However, although such measurements can reflect the search popularity of specific keywords over a period of time, they do not intuitively represent the actual behavior of investors.

[0136] For this reason, in this embodiment, the number of daily comments in comment communities such as Tieba is usually used to measure the attention of investors. Taking stock investment as an example, its comment community is a social media focused on stock investment, and the comments are written by actual investors, which directly reflects the interest and discussion level for a specific stock or market, and can reduce the noise caused by non-related searches. At the same time, the comment community has a community effect, and a network of information dissemination and opinion exchange will be formed between users. Comments are usually published in real time. When a topic or stock is discussed by more people, it may attract more potential investors to join the discussion, further amplifying the exposure and attention of the stock.

[0137] In this embodiment, for each stock associated with the request end, MRI (ie, the above MRI k,t ) performed a regression analysis on the number of daily comments in the stock review community and obtained the following analysis results:

[0138] The regression analysis formula is:

[0139] MRI k,t =a k +b k ×Tpost k,t +e k,t

[0140] The sentiment risk indicators are:

[0141]

[0142] Information Risk Index MRI k,t for:

[0143] MRI k,t =MRI k,t -MRII k,t

[0144] Among them, MRII k,t is the information risk index of requester k on day t, MRSI k,t is the sentiment risk index of requester k on day t, a k and b k are the regression coefficients of the request end k, Tpost k,t is the number of daily comments related to requester k on day t, e k,t is the error term of requester k on day t, and is the estimated value of the regression coefficient

[0145] For example, news and public opinion data can cover the period from year A, month B, day C to year D, month E, day F. For example, 3,285,120 media reports collected through 453 media platforms involve 250 listed companies, all of which may face the risk of stock price collapse due to the outbreak of negative news. News sources include all well-known self-media public account platforms and major financial media platforms. The data contains the article title, text content, publication date and corresponding platform information. For each news report, the MRI construction calculation shown in step 103 above is performed to obtain the daily frequency MRI index of these 250 companies, totaling 795,750 sample data. In order to separate the information and emotions in MRI, data from the stock comment community on a specific website is collected at the same time, including the number of daily comments (Tpostnum), including the number of positive comments (Pospostnum) and the number of negative comments (Negpostnum). By using the number of daily comments to represent investor attention (Dong et al.), the above formula regression analysis formula is used to split MRI into MRSI and MRII.

[0146] The explained variable is the company's daily return rate. The daily stock opening prices are collected through the CSMAR database, and the stock's daily return rate (r) is calculated by calculating the opening prices of two consecutive days. In order to intuitively show the relationship between the rate of return and MRI, Figure 3 The average monthly MRI (including MRSI and MRII) and yield discount chart of 250 stocks are displayed. At the same time, the companies with the most news among the 250 stocks are found. Figure 4 Draw its monthly MRI and yield line chart. By comparison, it can be found that the yield and MRI of a single stock show stronger volatility and extreme values ​​compared with the market average, especially between 2016 and 2017, when this stock was suspended for rectification for a period of time, and there were large fluctuations in media risk and yield before and after the period. There is a relatively obvious negative correlation between MRI and yield, and a relatively obvious positive correlation between MRI and MRII and MRSI. The correlation table of its yield r, MRI, MRII, MRSI and investor attention (IA) is shown in Table 6:

[0147] Table 6

[0148] r MRI MRII MRSI I A r 1.00 MRI -0.03 1.00 MRII -0.02 0.76 1.00 MRSI -0.02 0.65 0.00 1.00 I A -0.03 0.33 0.00 0.51 1.00

[0149] As shown in Table 6, the correlation between yield and investor attention is significantly negative, which means that compared with rising stock prices, falling stock prices are more likely to attract investors' strong attention. The correlation between MRI and yield is significantly negative, and the correlation between MRI and IA is significantly positive, which means that falling stock prices will cause negative media reports, and negative reports are likely to attract investors' attention. In addition, since MRII is the part of MRI that cannot be explained by IA, the correlation between MRII and IA and MRSI is 0.

[0150] Step 105, determining the public opinion risk description information corresponding to the requesting end according to the negative public opinion index, the emotional risk index and the information risk index.

[0151] In an example of the present application, step 105 may include the following sub-steps:

[0152] Calculate the first lagged rate of return according to the negative public opinion index and the preset request end parameters;

[0153] According to the sentiment risk index and the information risk index, combined with the request side parameters, the second lagged rate of return is calculated;

[0154] The first lagged rate of return and the second lagged rate of return are used to determine the public opinion risk description information corresponding to the requesting end.

[0155] In this embodiment, the percentile values ​​and maximum values ​​corresponding to the negative public opinion indicators over a period of time can be calculated. The public opinion characteristic types are matched according to the percentile values ​​and the maximum values. For example, as shown in Table 7 below, the 25th percentile and the 50th percentile are both 0, indicating that negative news has the characteristics of concentrated outbreaks and rapid spread. And because MRI is composed of MRII and MRSI, and MRII is the residual term after MRI regresses media coverage intensity, the average value of MRII is 0, and the mean value of MRSI is equal to the mean value of MRI. At the 25th, 50th and 75th percentiles, each media risk is 0, which shows that various types of media risks are relatively sparse. As shown in Table 7 below:

[0156] Table 7

[0157] Obs. Mean Std.Dev. Min Max 25％ 50％ 75％ r(％) 795,750 0.025 0.037 -80.732 47.058 -1.615 0.040 1.567 MRI 795,750 0.294 1.784 0 204.789 0 0 0.014 MRII 795,750 0.000 1.362 -123.806 202.366 -0.156 -0.044 0 MRSI 795,750 0.294 1.151 -2.705 180.978 0.016 0.076 0.232 Operational risk 795,750 0.030 0.152 0 19.040 0 0 0 Accounting risk 795,750 0.075 0.341 0 64.473 0 0 0 Stock market risk 795,750 0.116 0.425 0 48.667 0 0 0 Legal policy risk 795,750 0.125 0.476 0 57.128 0 0 0 Other risks 795,750 0.041 0.194 0 23.14 0 0 0 MKT_RF(%) 2,148 0.040 0.016 -9.478 9.165 -0.626 0.118 0.802 SMB (%) 2,148 0.028 0.008 -5.848 4.051 -0.365 0.091 0.482 HML(%) 2,148 0.011 0.007 -4.267 3.615 -0.461 -0.019 0.420 BIG4 1,972 0.048 0.214 0 0 0 0 0 SOE 1,972 0.254 0.436 0 1 0 0 1 TOP5(%) 1,972 49.006 15.892 6.907 97.456 37.519 48.633 59.369 POI (%) 1,972 40.846 23.034 0.001 98.125 22.685 40.098 57.745 SIZE 1,972 22.104 1.533 16.649 29.218 21.183 21.880 22.783 BM 1,972 0.391 0.646 -2.582 14.022 0.159 0.295 0.498 ROA 1,972 -0.069 1.302 -48.316 1.408 -0.012 0.017 0.045 LEV 1,972 2.155 9.116 -56.445 275.340 0.386 0.883 2.020 Analyst 1,972 0.951 1.147 0 4 0 0 2 Tpost 795,750 42.837 109.922 1 18,855 6 17 43 Post 795,750 10.989 26.532 0 3,919 2 5 11 Negpost 795,750 9.596 27.210 0 4,839 1 3 10

[0158] As shown in Table 4 above, in addition to MRI, it is also necessary to control market factors and company-level factors. Therefore, the control variables are divided into two categories: the first category contains daily time series data, such as the Fama-French three-factor model; the second category is annual frequency company panel data, including whether it is audited by the Big Four audit firms (BIG4), whether it is a specific type of enterprise (SOE), the shareholding ratio of the top five shareholders (TOP5), the shareholding ratio of institutional investors (POI), the logarithm of total assets (SIZE), the price-to-book ratio (MTB), the return on assets (ROA), the book leverage ratio (LEV) and the analyst attention (Analyst). The minimum value of the company's stock return is -80.732% and the maximum value is 47.058%, which indicates that the punished listed companies are more likely to experience extreme market conditions. The maximum value of MRI (Media Reaction Index) is 204.789, and its 25th and 50th percentiles are both 0, which indicates that negative news has the characteristics of concentrated outbreak and rapid spread. Since MRI is composed of MRII and MRSI, and MRII is the residual term after MRI is regressed on media coverage intensity, the mean of MRII is 0, and the mean of MRSI is equal to the mean of MRI. At the 25th, 50th, and 75th percentiles, each media risk is 0, which shows that each type of media risk is relatively sparse.

[0159] The definition of overreaction and underreaction is the correlation between a company's initial daily return after a news event and the company's daily return after the event. A positive correlation indicates an underreaction, while a negative correlation indicates an overreaction. Existing technologies mainly focus on the impact of individual news events released by a few authoritative media on stock prices. Therefore, event study methods are usually used to evaluate the impact of these specific events. However, in actual implementation, a stock may be affected by multiple related news and multiple events in one day. In this case, it is difficult for the event study method to capture the overall impact of all these concurrent events. In addition, the event study method often relies on a predetermined time window, which is not flexible enough for frequently updated information flows, and it is impossible to adjust the model appropriately as needed. MRI is a daily frequency indicator obtained by summing up the risks of all relevant news every day. It can capture the overall trend of media reports and their long-term impact on the company's stock price. This supports the use of a more flexible linear regression to examine the impact of media risk on stock prices. The first lag return is:

[0160]

[0161] The second lagged rate of return is:

[0162]

[0163] Among them, r is the rate of return of the request end k at the current time t and lagged by w days, α is the preset parameter, is the request side parameter. Among them, factor is the Fama-French three factors, which controls the impact of market factors. Control is other control variables at the company level, including BIG4, SOE, TOP5, POI, SIZE, BM, ROA, LEV and Analyst. At the same time, the year time fixed effect (Year) is controlled to prevent the endogeneity problem caused by omitted variables. The above formula can determine the impact of MRI, MRII and MRSI on stock prices, and the following Table 8 is obtained:

[0164] Table 8

[0165]

[0166] The t-statistics are in brackets. *: p<0.10, **: p<0.05, ***: p<0.01

[0167] Note: This table describes the impact of MRI (Media Reaction Index), MRII (Media Risk Information Index) and MRSI (Media Risk Sentiment Index) on stock returns. Fama-French factors and control variables are also included in the explanatory variables. Columns (1) and (2) test the impact of MRI on the stock price of the day, where the dependent variable is rk,t Columns (3) and (4) investigate the effect of MRI on the next day’s stock price, where the dependent variable is r k,t+1 e-2 means 10 of the reported number -2 , for example, -0.044 in column (1) means -0.044%.

[0168] As shown in Table 8, first, the stock price does not respond enough to MRI: the first column shows that the impact of MRI on the stock return rate on the same day is significantly negative (coefficient = -0.044, t-stat = -26.68, and the third column shows that the impact of MRI on the stock return rate on the next day is also significantly negative (coefficient = -0.021, t-stat = -12.00), and the signs of the two are consistent, so the stock price does not respond enough to MRI. In addition, for the public opinion risk description information, the emotional effect and the information effect can be further analyzed through the emotional risk index and the information risk index. For example, the impact of emotion on stock price is greater than that of information: the second column shows that in the media reports on the same day, MRSI has a significant negative impact on stock price. The effect is -0.053 (t-stat = -20.78), which is greater than the effect coefficient of MRII on stock price -0.038 (t-stat = -17.38). Finally, the impact of emotions lasts longer, while information is absorbed faster: in the fourth column, the impact of MRSI on the stock price on the next day decreases from -0.053 to -0.040 (t-stat = -14.67), and the impact of MRII on the stock price on the next day decreases from -0.038 to -0.007 (t-stat = -3.32), which shows that the impact of the emotional effect (MRSI) lasts until the next day, while the impact of the information effect (MRII) is significantly weakened on the next day, indicating that information is absorbed by the stock price faster.

[0169] To better demonstrate the impact of MRI on stock prices in subsequent days, Figure 5 As shown, this figure shows the impact coefficients of MRI (Media Reaction Index), MRSI (Media Risk Sentiment Index) and MRII (Media Risk Information Index) on stock returns at a lag of 20 trading days, as well as their 95% confidence intervals. The x-axis represents the number of lagged days, corresponding to w in the above formula. Each point in the figure represents the coefficient of MRI / MRSI / MRII obtained through regression analysis, and the vertical line at each point represents the 95% confidence interval. Figure 5The MRI, MRII and MRSI coefficients (95% confidence interval) for the next 20 trading days are shown. This figure shows that the stock price does under-react to MRI. As time goes by, the impact of MRI on subsequent stock prices gradually converges to 0, while the information part of MRI has already converged to 0 in the two days after the news report. Even if the coefficient is greater than 0 after t+10, the confidence intervals show that the hypothesis that the coefficient is equal to 0 cannot be rejected. Compared with MRII, MRSI will have a continuous impact on stock prices. The emotional effect brought about by repeated media reports continues to ferment, and the impact on stock price returns after t+10 is still -0.02.

[0170] like Figure 6 As shown in Figure 1, this figure shows the autocorrelation coefficients of MRI (Media Reaction Index), MRSI (Media Risk Sentiment Index), and MRII (Media Risk Information Index). The x-axis represents the number of lagged days, corresponding to w in the above formula. The y-axis shows the correlation coefficient between the current indicator and the same indicator lagged w days, from Figure 6 It can be seen that the autocorrelation coefficient of MRSI remains high at a longer time lag, especially after 10 periods of lag, the autocorrelation coefficient is still significantly higher than 0.5. In contrast, the autocorrelation coefficients of MRI and MRII drop rapidly at a shorter time lag, especially MRII, which almost drops below 0.2 after 4 periods of lag, which further indicates that most of the information in the media is independent and novel, while the emotions in the media are contagious and last longer.

[0171] Taking into account the autocorrelation of the media, a panel vector autoregression model is used to further determine the impact of MRI on stock prices. Compared with the direct linear regression model, PVAR can not only handle the dynamic relationship between multiple variables, but also effectively capture the lag effect of time series, control cross-sectional heterogeneity, and identify causal relationships, thereby providing more accurate and comprehensive results. Figure 3 and Figure 4 It can be found that MRI (including MRSI and MRII) after a lag of 4 days have entered a relatively stable state. This may be because the market is only open on weekdays, and the lag of four days happens to include a work week. Therefore, a fourth-order lag operator L4 can be defined, that is, L4(x_t) = [x_(t-1), x_(t-2), x_(t-3), x_(t-4)]. The regression formula is shown in the above regression analysis formula, and Table 7 reports the regression results. In addition, it should be noted that since the PVAR model deals with the dynamic interaction relationship between multiple endogenous variables, the existence of the lag term makes the traditional R2 concept no longer applicable. Therefore, the chi-square statistic is used to evaluate the overall goodness of fit and significance of the model and identify causal relationships:

[0172] rk,t =α1+β1L4(r k,t )+β2L4(MRI k,t )+ε 1,t

[0173] MRI k,t =α1+β1L4(r k,t )+β2L4(MRI k,t )+ε 2,t

[0174] The calculation results are shown in Table 9:

[0175] Table 9

[0176]

[0177]

[0178] T-statistics are in parentheses. *p<0.10, **p<0.05, ***p<0.01

[0179] Note: This table describes a vector autoregression model of four-day lagged returns and MRI (Media Reaction Index). It takes the current day return and MRI as dependent variables, and the previous four days' return and MRI as independent variables, respectively, where e-2 represents numbers reported in units of 10-2, for example, MRI of -0.016 in column (2) means -0.016%.

[0180] It can be seen that in the regression formula, the first lag variable of MRI is significantly negative (coeffcient = -0.016, t-stat = -4.70), while the second, third and fourth lag variables are not significant. Considering that its sign is consistent with the impact of MRI on the stock price on the same day in Table 3, it is further verified that the stock price has a low response to MRI, and this under-reaction only occurs in the media reports of the previous day. In the MRI regression formula, the first lag variable of the return rate is significantly negative at the first order (coeffcient = -0.196, t-stat = -2.02). This finding shows that although MRI has an under-reaction to stock prices, the news media also reported the risk of stock price decline on the previous day. From the chi-square statistic, MRI (especially MRIt-1) has a significant Granger causal relationship with stock price returns, while stock price returns (mainly rt-1) form a relatively weak causal relationship with MRI.

[0181] Table 10

[0182]

[0183]

[0184] This table describes the vector autoregression model of four-day lagged returns and MRI (Media Reaction Index) / MRSI (Media Risk Sentiment Index). For information effects, MRII (Media Risk Information Index) and PVAR (Vector Autoregression Model) of returns are used. For sentiment effects, MRSI and PVAR of returns are used, where e-2 indicates numbers reported in units of 10-2.

[0185] In addition, the information effect (MRII) and the emotional effect (MRSI) were investigated in MRI, and the results are shown in Table 10. For the information effect, the vector autoregression model of MRII and return rate is used; for the emotional effect, it is replaced by MRSI. In the regression of return rate, the first lag variable of MRII is negative (coefficient = -0.006, t-stat = -1.96), which is consistent with the sign in Table 6, indicating that stock prices underreact to the information component of media risk, and this underreaction is small and only lagged for one day, indicating that the information is absorbed by the stock price quickly. On the other hand, the first lag variable of MRSI is negative (coefficient = -0.046, t-stat = -4.61), while the fourth lag variable is positive (coefficient = 0.015, t-stat = 1.96), indicating that stock prices initially underreact to the emotional component of media risk and then overreact. In the regression of MRII, the first lagged variable of return rate is -0.151 (t-stat = -1.75), and the third lagged variable is -0.155 (s-stat = -2.04), indicating that media reports will also consider the previous stock price decline as a kind of negative information. In the regression of MRSI, the second lagged variable of return rate is significantly positive (coefficient = 0.146, t-stat = 3.95), indicating that stock prices overreact to MRSI. From the chi-square statistic, MRSI has a more significant causal relationship with return rate than MRII.

[0186] In one example of the present invention, the method further comprises the following steps:

[0187] When multiple negative public opinion indicators are obtained, the request ends are sorted according to the negative public opinion indicators to obtain a request end sequence;

[0188] Divide the request end sequence according to a preset number of groups to obtain multiple request end combinations;

[0189] Create a resource configuration policy to allocate resources to each request end combination according to the preset weights.

[0190] In this embodiment, the following method is used to illustrate how the resource configuration strategy can effectively allocate resources.

[0191] Taking stock investment as an example, referring to Table 7, the mean proportion of audits by the Big Four audit firms (BIG4) is 0.048, which means that 4.8% of the companies are audited by the Big Four audit firms. The mean proportion of SOEs is 0.254, indicating that about 25% of the companies are SOEs. The mean return on assets (ROA) is -0.069, suggesting that the profitability of the penalized companies is relatively weak. In addition, variables such as the shareholding ratio of the top five shareholders (TOP5), the shareholding ratio of institutional investors (POI), the logarithm of total assets (SIZE), the price-to-book ratio (MTB), the book leverage ratio (LEV), and the attention of analysts (Analyst) also provide information about the governance structure, capital structure, and market attention of these companies. The above statistical data reflect the characteristics of the sample companies in different dimensions, which is of great significance for further analyzing the relationship between media risk and stock price fluctuations.

[0192] Through the above analysis, it is determined that stock prices under-react to MRI (market reaction index), which can be used to predict the stock price of the next day. In order to test the predictive ability of MRI, a long-short portfolio is constructed based on MRI and its effectiveness in asset pricing is examined. 250 companies are sorted from high to low according to the MRI of the previous day and divided into five portfolios. Table 11 reports the Jensen alpha and risk loading of the MRI quintile portfolios under the CAPM and Fama-French three-factor model. It is worth noting that since MRI has a negative predictive effect on stock prices, when constructing a long-short portfolio, the high group is subtracted from the low group, so there is a high possibility of achieving positive excess returns:

[0193] Table 11

[0194]

[0195]

[0196] The t statistics are in parentheses. *p<0.10, **p<0.05, ***p<0.01. This table divides all stocks into five groups according to the MRI (media reaction index) of the previous day and examines the abnormal returns of these five stock portfolios. Panel A directly examines the abnormal returns of different stock portfolios and long-short portfolios, while Panels B and C consider the CAPM model and Fama's three-factor model, respectively.

[0197] As shown in Table 11, the empirical results demonstrate the predictive power of MRI. Although only the high groups in Panel B and Panel C show significant alpha values, all three regression analyses generate substantial positive alpha values ​​for the long-short portfolio. The daily excess return of the long-short portfolio reaches more than 0.05% per day, or 12.5% ​​per year (assuming 250 trading days per year). Descriptive statistics show that MRI appears at the 50th percentile, which means that for more than half of the time, the company's MRI is zero. Therefore, no significant decreasing trend of alpha is observed in the third, fourth, and low quintiles. However, significant negative alpha values ​​can be seen in the high and second quintiles, and the negative amplitude of the high quintile is larger. These findings indicate that common factors cannot fully explain the return differences between different MRI portfolios. Investors can gain returns by constructing corresponding resource allocation strategies such as investment strategies through MRI.

[0198] In an embodiment of the present invention, news public opinion data is obtained by responding to an identification request sent by a requesting end; a risk emotion prediction model is used to predict the news public opinion data to determine the corresponding risk type probability and emotional information; weighting is performed according to the risk type probability and emotional information to construct a negative public opinion index; regression analysis is performed on the negative public opinion index to determine the emotional risk index and the information risk index; based on the negative public opinion index, the emotional risk index and the information risk index, the public opinion risk description information corresponding to the requesting end is determined, thereby quickly and accurately predicting and describing the public opinion risk, which helps users to conduct risk warnings and risk defense for the company where the requesting end is located.

[0199] See also Figure 7 , Figure 7 A structural block diagram of a negative public opinion situation awareness risk identification device provided by an embodiment of the present invention is shown.

[0200] The embodiment of the present invention provides a device for identifying negative public opinion situation awareness risks, comprising:

[0201] The data acquisition module 701 is used to respond to the identification request sent by the request end and obtain news public opinion data;

[0202] The risk sentiment determination module 702 is used to predict the news public opinion data using the risk sentiment prediction model to determine the corresponding risk type probability and sentiment information;

[0203] Negative public opinion index determination module 703, used to construct negative public opinion index by weighting according to risk type probability and sentiment information;

[0204] The indicator regression analysis module 704 is used to perform regression analysis on negative public opinion indicators to determine emotional risk indicators and information risk indicators;

[0205] The risk identification module 705 is used to determine the public opinion risk description information corresponding to the requesting end according to the negative public opinion index, the emotional risk index and the information risk index.

[0206] Optionally, the device further comprises:

[0207] Historical data acquisition module, used to obtain historical news and public opinion data;

[0208] The data preprocessing module is used to preprocess the historical news and public opinion data, obtain multiple key words and form a bag-of-words model;

[0209] The vector generation module is used to perform one-hot encoding on the bag-of-words model, generate document vectors, and annotate risk labels and sentiment labels;

[0210] The model determination module is used to determine the public opinion risk description information corresponding to the requesting end based on negative public opinion indicators, emotional risk indicators and information risk indicators.

[0211] Optionally, the model determination module is specifically used for:

[0212] Input document vectors into multiple preset machine learning models to obtain prediction results;

[0213] Compare the prediction results with the risk label and sentiment label respectively, and calculate the loss function value;

[0214] Optimize the parameters of each machine learning model according to the loss function value until the loss function value is less than a preset loss threshold, and obtain multiple intermediate machine learning models;

[0215] Calculate the prediction effect indicators corresponding to each intermediate machine learning model, and select the intermediate machine learning model whose prediction effect indicators meet the selection conditions as the risk sentiment prediction model.

[0216] Optionally, the device further comprises:

[0217] A sorting module is used to sort the request ends according to the negative public opinion indicators when multiple negative public opinion indicators are obtained to obtain a request end sequence;

[0218] A combination division module, used to divide the request end sequence according to a preset number of groups to obtain multiple request end combinations;

[0219] The resource configuration module is used to create a resource configuration strategy to allocate resources to each request end combination according to the preset weight.

[0220] Optionally, the risk sentiment determination module 702 is specifically configured to:

[0221] Convert news and public opinion data into vectors to be input;

[0222] Input the input vector to the risk sentiment prediction model, and determine the risk type probability corresponding to the input vector through the risk prediction branch of the risk sentiment prediction model;

[0223] The emotion prediction branch of the risk emotion prediction model is used to determine the emotion probability corresponding to the input vector as emotion information.

[0224] Optionally, the negative public opinion index determination module 703 is specifically used for:

[0225] The analytic hierarchy process is used to determine the risk weight corresponding to the probability of each risk type;

[0226] Construct negative public opinion indicators according to risk weight, risk type probability and sentiment information;

[0227] Negative public opinion indicators are:

[0228]

[0229] Among them, risk k,t,m is the probability value of risk type m belonging to requester k on day t; n is the total number of news and public opinion data of requester k on day t; is the sentiment information of the a-th news public opinion data of the requesting end k on the t-th day; is the risk type probability of the a-th news public opinion data of requester k on day t belonging to risk type m; θ m is the risk weight of risk type m; b is the total number of risk types; MRI k,t is the total public opinion risk value of the requesting end k on the tth day.

[0230] Optionally, the indicator regression analysis module 704 is specifically used for:

[0231] Using negative public opinion indicators as independent variables and the number of daily comments as dependent variables, regression analysis was performed to construct a regression analysis formula;

[0232] Transform the regression analysis formula to determine the emotional risk index and information risk index;

[0233] The regression analysis formula is:

[0234] MRI k,t =a k +b k ×Tpost k,t +e k,t

[0235] The sentiment risk indicators are:

[0236]

[0237] Information Risk Index MRSI k,t for:

[0238] MPSI k,t =MRI k,t -MRII k,t

[0239] Among them, MRII k,t is the information risk index of requester k on day t, MRSI k,t is the sentiment risk index of requester k on day t, a k and b k are the regression coefficients of the request end k, Tpost k,t is the number of daily comments related to requester k on day t, e k,t is the error term of requester k on day t, and is the estimated value of the regression coefficient.

[0240] Optionally, the risk identification module 705 is specifically used for:

[0241] Calculate the first lagged rate of return according to the negative public opinion index and the preset request end parameters;

[0242] According to the sentiment risk index and the information risk index, combined with the request side parameters, the second lagged rate of return is calculated;

[0243] The first lagged rate of return and the second lagged rate of return are used to determine the public opinion risk description information corresponding to the requesting end.

[0244] An embodiment of the present invention provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of the negative public opinion situation awareness risk identification method as described in any embodiment of the present invention.

[0245] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and modules can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0246] In the several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0247] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0248] In addition, each functional module in each embodiment of the present invention may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of software functional modules.

[0249] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying negative public opinion situation awareness risks, characterized in that: include: Respond to the identification request sent by the requester and obtain news and public opinion data; Using a risk sentiment prediction model to predict the news public opinion data, and determine the corresponding risk type probability and sentiment information; Constructing a negative public opinion index by weighting the risk type probability and the sentiment information; Performing regression analysis on the negative public opinion indicators to determine emotional risk indicators and information risk indicators; According to the negative public opinion index, the emotional risk index and the information risk index, the public opinion risk description information corresponding to the requesting end is determined.

2. The method according to claim 1, characterized in that: The method further comprises: Obtain historical news and public opinion data; Preprocessing the historical news public opinion data to obtain multiple key words and form a bag-of-words model; One-hot encoding is performed on the bag-of-words model to generate a document vector and annotate risk labels and sentiment labels; The document vector is used to train multiple preset machine learning models and respectively calculate prediction effect indicators, and a risk sentiment prediction model is selected according to the prediction effect indicators.

3. The method according to claim 2, characterized in that The adopting of the document vector to train a plurality of preset machine learning models and respectively calculating prediction effect indicators, and selecting a risk sentiment prediction model according to the prediction effect indicators, comprises: Inputting the document vectors into a plurality of preset machine learning models respectively to obtain prediction results; Compare the prediction results with the risk label and the emotion label respectively, and calculate the loss function value; Optimize the parameters of each of the machine learning models according to the loss function value until the loss function value is less than a preset loss threshold, thereby obtaining a plurality of intermediate machine learning models; Calculate the prediction effect indicators corresponding to each of the intermediate machine learning models, and select the intermediate machine learning model whose prediction effect indicators meet the selection conditions as the risk sentiment prediction model.

4. The method according to claim 1, characterized in that: The method further comprises: When multiple negative public opinion indicators are obtained, the requesting ends are sorted according to the negative public opinion indicators to obtain a requesting end sequence; Divide the request end sequence according to a preset number of groups to obtain multiple request end combinations; A resource configuration strategy is created to allocate resources to each of the request end combinations according to preset weights.

5. The method according to claim 1, characterized in that: The risk sentiment prediction model is used to predict the news public opinion data to determine the corresponding risk type probability and sentiment information, including: Convert the news and public opinion data into a vector to be input; Inputting the vector to be input into the risk sentiment prediction model, and determining the risk type probability corresponding to the vector to be input through the risk prediction branch of the risk sentiment prediction model; The emotion probability corresponding to the input vector is determined as emotion information through the emotion prediction branch of the risk emotion prediction model.

6. The method according to claim 1, characterized in that The weighting according to the risk type probability and the emotional information to construct a negative public opinion index includes: The risk weight corresponding to the probability of each risk type is determined by using the analytic hierarchy process; Constructing a negative public opinion index according to the risk weight, the risk type probability and the sentiment information; Negative public opinion indicators are: in, is the probability value of risk type m belonging to requester k on day t; n is the total number of news and public opinion data of requester k on day t; is the sentiment information of the a-th news public opinion data of the requesting end k on the t-th day; is the risk type probability of the a-th news public opinion data of requester k on day t belonging to risk type m; is the risk weight of risk type m; b is the total number of risk types; is the total public opinion risk value of the requesting end k on the tth day.

7. The method according to claim 1, characterized in that The performing regression analysis on the negative public opinion index to determine the emotional risk index and the information risk index includes: Using the negative public opinion index as an independent variable and the number of daily comments as a dependent variable, a regression analysis is performed to construct a regression analysis formula; Transforming the regression analysis formula to determine the emotional risk index and the information risk index; The regression analysis formula is: The sentiment risk indicators are: Information risk indicators for: in, is the information risk index of requester k on day t, is the sentiment risk index of requester k on day t, and are the regression coefficients of the request end k, is the number of daily comments related to requester k on day t, is the error term of requester k on day t, and is the estimated value of the regression coefficient.

8. The method according to claim 1, characterized in that The determining, according to the negative public opinion index, the emotional risk index and the information risk index, the public opinion risk description information corresponding to the requesting end includes: Calculate the first lagged rate of return according to the negative public opinion index and the preset request end parameters; Calculate a second lagged rate of return according to the sentiment risk indicator and the information risk indicator in combination with the requesting end parameter; The first lagged rate of return and the second lagged rate of return are used to determine the public opinion risk description information corresponding to the requesting end.

9. A device for identifying negative public opinion situation awareness risks, characterized in that: include: The data acquisition module is used to respond to the identification request sent by the request end and obtain news and public opinion data; A risk sentiment determination module is used to predict the news public opinion data using a risk sentiment prediction model to determine the corresponding risk type probability and sentiment information; A negative public opinion index determination module, used to construct a negative public opinion index by weighting the risk type probability and the emotional information; An indicator regression analysis module is used to perform regression analysis on the negative public opinion indicators to determine emotional risk indicators and information risk indicators; The risk identification module is used to determine the public opinion risk description information corresponding to the requesting end according to the negative public opinion index, the emotional risk index and the information risk index.

10. An electronic device, characterized in that: It includes a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of the negative public opinion situation awareness risk identification method as described in any one of claims 1-8.

Images