CN116244446A - Method and system for detecting social media cognitive threats - Google Patents

Abstract

The invention belongs to the technical field of network security, and particularly relates to a social media cognitive threat detection method and a social media cognitive threat detection system, which are used for collecting text data of sensitive topics of a network platform and preprocessing the data; aiming at the preprocessed sensitive topic text data, acquiring a cognitive threat topic text through multi-level cognitive threat detection; constructing a cognitive threat propagation knowledge graph through named entity identification and entity relation extraction of the cognitive threat topic text; and carrying out user tracing, event tracing and organization tracing on the text transmission of the cognitive threat topics based on the knowledge graph of the cognitive threat transmission. Aiming at topic texts related to specific topics and sensitive events, the invention utilizes the emotion tendencies behind the topic texts to identify cognitive threats, compared with the traditional manual evidence, the identification period is greatly shortened, the detection information quantity and efficiency are improved, the safety and feasibility are good, the threat judgment accuracy is high, the detection effect is good, and the application scene is very wide.

Description

Method and system for detecting social media cognitive threats

technical field

The invention belongs to the technical field of network security, and in particular relates to a social media cognitive threat detection method and system.

Background technique

Cognition refers to the process of people acquiring knowledge or applying knowledge, or the process of information processing, which includes feeling, perception, memory, thinking, imagination and language. Cognitive threats are based on information that is purposeful, inflammatory, concealed, directional, and untrue to the individual input, and through continuous influence and solidification of the individual cognitive process, the individual forms a distorted, unconventional, and reverse negative cognition, or change the individual's existing normal cognitive system, making it deviate from the core value system of society. The emergence of self-media platforms and social platforms has provided a hotbed for the breeding and spread of cognitive threats. Cyberspace has become the main battlefield of cognitive threat confrontation. Social networks have the characteristics of anonymous identity, "free speech", high real-time performance, and fast communication. Most of the users are young people, who are not sensitive to social issues and are easily permeated by cognition. However, problems such as the high concealment of cognitive threats, the difficulty of traceability, and the difficulty of cross-platform supervision need to be solved urgently. It is imminent to identify, trace and counter cognitive threat information technically. In view of the high concealment of cognitive threats, the difficulty of traceability, and the difficulty of cross-platform supervision, how to technically contain cognitive threats has become an urgent need to purify cyberspace.

Contents of the invention

To this end, the present invention provides a social media cognitive threat detection method and system, which uses the emotional tendency behind it to identify cognitive threats for specific topics and topic texts related to sensitive events. The identification period is shortened, and the amount of information and efficiency of inspection are improved.

According to the design scheme provided by the present invention, a social media cognitive threat detection method is provided, including the following content:

Collect text data of sensitive topics on the network platform and perform preprocessing operations on the data;

For the preprocessed sensitive topic text data, the cognitive threat topic text is obtained through multi-level cognitive threat detection, wherein the multi-level cognitive threat detection includes: dividing the sensitive topic text data into cognitive threat topic text and initial suspected Primary detection of cognitive threat topic texts, classifying initial suspected cognitive threat topic texts into cognitive threat topic texts, suspected cognitive threat topic texts, and intermediate detection of non-cognitive threat topic texts; Threat topic text to obtain the ultimate detection of cognitive threat topic text;

Construct a knowledge map of cognitive threat communication through named entity recognition and entity relationship extraction of cognitive threat topic texts;

Based on the knowledge map of cognitive threat communication, user tracing, event tracing and organizational tracing are carried out on cognitive threat topic text dissemination.

As the social media cognitive threat detection method in the present invention, further, collect sensitive topic text data on the network platform and perform preprocessing operations on the data, including:

First, according to the distributed collection of user authorization information database, sensitive topic text information and related user data on the network platform;

Then, for the collected text information, merge the title with the text, use the redundancy detection algorithm to remove redundant information, de-duplicate the relevant comments, clean and transform the text noise data, and use the word segmentation system to segment the text .

As the social media cognitive threat detection method in the present invention, further, in the primary detection, the sentiment analysis method is used to divide the sensitive topic text data into cognitive threat topic text and initial suspected cognitive threat topic text, wherein the sentiment analysis method divides The process includes:

First, based on the known emotional dictionary and using the word frequency statistics method to construct the basic emotional dictionary, and expand the emotional dictionary by making correlation statistics between the words in the text data and the basic emotional dictionary;

Then, take the text in the sensitive topic text data as the unit and the emotional word as the delimiter, carry out the sentiment weight statistics on the sentences between each delimiter, and calculate the negative sentiment weight according to the proportion of the negative sentiment weight in all the sentiment word weights. Judge the emotional polarity of the text;

Then, according to the emotional polarity of the text, the sensitive topic text data is divided into cognitive threat topic text and initial suspected cognitive threat topic text.

As a social media cognitive threat detection method in the present invention, further, construct a basic emotional dictionary according to a known emotional dictionary and use a word frequency statistical method, including:

First, select a series of emotional words in the known emotional dictionary, sort the emotional words according to the search engine clicks in the series of emotional words, and select some emotional words according to the popularity of the clicks;

Then, based on the word frequency statistics, the emotional vocabulary with the highest relevance to the topic is selected, and a number of selected emotional words and emotional vocabulary are used to form a basic emotional dictionary;

Then, the basic sentiment dictionary is expanded with synonyms and candidate words with sentimental tendencies.

As a social media cognitive threat detection method in the present invention, further, carry out sentiment weight statistics on the sentences between each delimiter, and judge the emotion of the text according to the proportion of negative sentiment weights in all emotional word weights Polarity, including:

First of all, for the sentences between the delimiters, the emotional tendency is counted through the analysis of emotional words, negative words, adverbs, fixed collocations, transition words and exclamation sentences;

Then, the statistical text contains the sum of the negative sentiment tendency values of all clauses and the sum of the absolute value of the overall sentiment weight, and uses the weight of negative sentiment words in the vertical proportion of all sentiment words in the text to judge the sentiment polarity of the text.

As the social media cognitive threat detection method of the present invention, further, in the intermediate detection, the initial suspected cognitive threat topic text is classified into cognitive threat topic text, suspected cognitive threat topic text and non-cognitive threat topic text by using deep learning method For text, the classification process includes:

Construct a deep learning model and use the training data set with labels for pre-training, where the deep learning model includes a BERT model for word vector representation of the input, and cognitive threat detection for the input word vector The BiLSTM model;

Input the initial suspected cognitive threat topic text into the pre-trained deep learning model, use the deep learning model to obtain the cognitive threat probability value, and determine the cognitive threat in the initial suspected cognitive threat topic text through the cognitive threat probability value Topic text, suspected cognitive threat topic text, and non-cognitive threat topic text.

As the social media cognitive threat detection method of the present invention, further, for the preprocessed sensitive topic text data, the cognitive threat topic text is obtained through multi-level cognitive threat detection, which also includes: using the sentiment analysis method in the cognitive threat The comment area in the topic text evaluates the impact of cognitive threats based on the overall emotional tendency.

As the social media cognitive threat detection method of the present invention, further, the knowledge map of cognitive threat propagation is constructed through named entity recognition and entity relationship extraction of the cognitive threat topic text, including:

Construct a named entity extraction model, and optimize the model by using an adversarial training method. The named entity recognition model includes an encoder for mapping input characters to real number space and mining potential semantics, and is used to convert vectors to The BiLSTM neural network layer for extracting contextual semantic information by forward and backward bidirectional features, and the CRF conditional random field layer for taking the bidirectional features extracted by the BiLSTM neural network layer as input and combining the Bioes labeling paradigm to generate the corresponding label of the character;

The cognitive threat topic text is used as the input of the optimized named entity extraction model, and the named entity extraction model is used to identify the entity categories and relationships in the cognitive threat topic text.

As the social media cognitive threat detection method of the present invention, further, constructing the cognitive threat propagation knowledge map through the named entity recognition and entity relationship extraction of the cognitive threat topic text, two named entity extraction models connected by pipeline are built , where the first named entity extraction model uses a single-label multi-classification task method to identify entities in the cognitive threat topic text, and the second named entity extraction model uses a multi-label multi-classification task method to input the first named entity extraction model as input to identify relationships between entities.

Further, the present invention also provides a social media cognitive threat detection system, including: a data collection server, multiple cognitive threat identification servers, a knowledge map server and a web server, wherein,

The data collection server is used to collect text data of sensitive topics on the network platform and perform preprocessing operations on the data;

Multiple cognitive threat identification servers are used to obtain the cognitive threat topic text through multi-level cognitive threat detection for the preprocessed sensitive topic text data, wherein the multiple cognitive threat identification servers specifically include: sensitive topic The text data is divided into cognitive threat topic text and initial suspected cognitive threat topic text, and the primary identification server classifies the initial suspected cognitive threat topic text into cognitive threat topic text, suspected cognitive threat topic text and non-cognitive threat topic text An intermediate authentication server for the text, and an ultimate authentication server for obtaining the cognitive threat topic text from the suspected cognitive threat topic text through manual annotation;

The knowledge map server is used to build a cognitive threat communication knowledge map through named entity recognition and entity relationship extraction of cognitive threat topic texts;

The web server is used to disseminate the knowledge graph based on the cognitive threat and use the web interactive interface to trace the user source, event source and organization source of the cognitive threat topic text dissemination.

Beneficial effects of the present invention:

The present invention can rely on network social platforms such as Weibo, Zhihu, and WeChat official accounts to perform multi-dimensional sentiment analysis on crawled sensitive topic texts and their comments to realize the detection of cognitive threat topic texts, and identify cognitive threat-related names Entities and the relationship between entities are extracted to construct a knowledge map of cognitive threat communication, and the knowledge map is used to realize the visual display of cognitive threat communication user traceability, cognitive threat communication event source tracing, and cognitive threat communication organization source tracing. Cognitive threat propagation prediction, real-time monitoring of key accounts, groups, organizations, and users, analysis of cognitive confrontation strategies, blocking cognitive threat propagation, deepening network cognitive threat supervision, and deterring cognitive threat-related network violations , Effectively purify cyberspace.

Description of drawings:

Fig. 1 is a schematic diagram of social media cognitive threat detection process in an embodiment;

Fig. 2 is a schematic diagram of cognitive threat detection and measurement process in an embodiment;

Figure 3 is a schematic diagram of the confrontation training process in the embodiment;

Fig. 4 is a schematic diagram of the construction hierarchy of the knowledge map visualization in the embodiment.

Detailed ways:

In order to make the purpose, technical solution and advantages of the present invention more clear and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and technical solutions.

The development of the network environment makes the implementation of cognitive domain threats easier and easier, and can be implemented individually or jointly in multiple dimensions and levels, thereby affecting the entire social value form. The embodiment of this case, as shown in Figure 1, provides a social media cognitive threat detection method, including:

S101. Collect text data of sensitive topics on the network platform and perform preprocessing operations on the data;

S102. For the preprocessed sensitive topic text data, obtain the cognitive threat topic text through multi-level cognitive threat detection, wherein the multi-level cognitive threat detection includes: dividing the sensitive topic text data into cognitive threat topic text and The primary detection of the initial suspected cognitive threat topic text, the intermediate detection of classifying the initial suspected cognitive threat topic text into cognitive threat topic text, suspected cognitive threat topic text and non-cognitive threat topic text, and manual labeling from suspected Cognitive threat topic text to obtain the ultimate detection of cognitive threat topic text;

S103. Construct a cognitive threat communication knowledge graph through named entity recognition and entity relationship extraction of the cognitive threat topic text;

S104. Perform user tracing, event tracing, and organization tracing on cognitive threat topic text propagation based on the cognitive threat propagation knowledge graph.

Relying on online social platforms such as Weibo, Zhihu, and WeChat official accounts, multi-dimensional sentiment analysis is performed on crawled sensitive topic texts and their comments to realize the detection of cognitive threat topic texts. By identifying cognitive threat-related named entities and extracting The relationship between entities is used to build a cognitive threat communication knowledge map, and the knowledge map is used to trace the source of cognitive threat communication users. By building a knowledge map of cognitive threats, predicting the propagation path of cognitive threats, and providing analysis of cognitive confrontation strategies. Not only can it achieve accurate identification in the context of short texts on social media platforms, but it can also maintain a high accuracy rate when identifying long media texts, making news media responsible for their words and deeds, and can effectively deter some unscrupulous media.

As a preferred embodiment, further, collect sensitive topic text data on the network platform and perform preprocessing operations on the data, including:

First, according to the distributed collection of user authorization information database, sensitive topic text information and related user data on the network platform;

Then, for the collected text information, merge the title with the text, use the redundancy detection algorithm to remove redundant information, de-duplicate the relevant comments, clean and transform the text noise data, and use the word segmentation system to segment the text .

A large amount of sensitive topic text data of social platforms such as Weibo and official accounts can be obtained through the API interface, which includes ten fields such as article title, text, and comments. In order to facilitate further processing, the title and text of the data are merged, redundant information is removed through redundancy detection algorithms, comments are deduplicated, noise data is cleaned and converted, and word segmentation data preprocessing operations are performed through ICTCLAS. The redundancy detection algorithm for removing redundant information can be designed to include the following steps:

Step1: Segment the text according to the punctuation mark;

Step2: Obtain the first 5 sentences of the article after the sentence clause. If it contains words such as "Follow us" and "Click ** font", delete the sentence and keep the rest;

Step3: Obtain the first 10 sentences of the article after the sentence division. If it contains texts such as "edit", "first review", "click to read", delete the sentence, and keep the rest;

Step4: Merge the retained sentences back into text.

It should be noted that the data collected in this case can be text data of sensitive topics on various social networking platforms. It can be collected through the Weibo platform, and will also be randomly collected through social platforms such as Zhihu and WeChat public accounts. The main process of microblog data collection may include: user authorization, acquisition of newly released microblogs, update of microblog information, and acquisition of user information. User authorization is completed through Oauth2, and the acquisition of newly released Weibo, Weibo information update, and user information acquisition are completed by automatically calling the official API interface of Weibo.

In primary detection, the sentiment analysis method can be used to divide sensitive topic text data into cognitive threat topic text and initial suspected cognitive threat topic text. The process of sentiment analysis method division includes:

First, based on the known emotional dictionary and using the word frequency statistics method to construct the basic emotional dictionary, and expand the emotional dictionary by making correlation statistics between the words in the text data and the basic emotional dictionary;

Then, take the text in the sensitive topic text data as the unit and the emotional word as the delimiter, carry out the sentiment weight statistics on the sentences between each delimiter, and calculate the negative sentiment weight according to the proportion of the negative sentiment weight in all the sentiment word weights. Judge the emotional polarity of the text;

Then, according to the emotional polarity of the text, the sensitive topic text data is divided into cognitive threat topic text and initial suspected cognitive threat topic text.

Multidimensional sentiment analysis is the process of analyzing and processing the emotional polarity, emotional degree, and emotional category of subjective texts with emotional color by using natural language processing and text mining technology. An important research direction in the field of NLP is sentiment analysis. Correct and effective sentiment analysis can quickly get the positive or negative emotions expressed by people from the text, which helps to discover the emotional tendency behind the text, and then separate the hidden information hidden in the massive information. political threats and culturally penetrating cognitive threats. Sentiment analysis tasks can be divided into chapter level, sentence level, word or phrase level according to the granularity of analysis; according to the category of text processing, it can be divided into text-based sentiment analysis and comment-based sentiment analysis. According to the type of task studied, It can be divided into sub-problems such as sentiment classification, sentiment retrieval and sentiment extraction. In the embodiment of this case, as shown in Figure 2, the basic flow of cognitive threat identification and measurement based on dynamic expansion of sentiment dictionary and deep learning

The methods of emotion classification can be roughly divided into classification methods based on sentiment lexicon and classification methods based on deep learning. Both types of methods have their own characteristics and shortcomings. The method based on the sentiment dictionary refers to the use of a sentiment dictionary marked with sentiment polarity to quantify the emotional polarity of the text. This method uses a series of rules and the sentiment dictionary to classify. Analyze the words in the text for matching, and then obtain the emotional value of the sentence through calculation, and finally use the obtained emotional value as the basis for judging the emotional orientation of the sentence. Although this method has a relatively high accuracy rate, the cost of building an emotional dictionary is relatively high. , and the method based on the sentiment dictionary does not consider the connection between words in the text and lacks word meaning information; the method based on deep learning regards sentiment classification as a special text classification, and uses manual annotation and machine learning methods to classify the text Sentiment classification. The method based on deep learning uses marked data and labels. These data and labels are manually marked, and then use the method of deep learning to analyze the sentiment of the text. The commonly used machine learning methods are Naive Bayesian NB (NaiveBayes) , decision tree, support vector machine SVM (SupportVectorMachine), etc. The effect of this method mainly depends on the quantity and quality of manually labeled data, so it is greatly affected by people's subjective consciousness and consumes a lot of labor.

According to the respective characteristics of the two methods, the two methods based on the sentiment dictionary and deep learning are combined and optimized, and a multi-dimensional sentiment analysis method combining the dynamic expansion of the sentiment dictionary and deep learning is proposed to overcome the respective shortcomings of the two methods and achieve comparative results. High accuracy.

Among them, based on the known emotional dictionary and using the word frequency statistical method to construct the basic emotional dictionary, including:

First, select a series of emotional words in the known emotional dictionary, sort the emotional words according to the search engine clicks in the series of emotional words, and select some emotional words according to the popularity of the clicks;

Then, based on the word frequency statistics, the emotional vocabulary with the highest relevance to the topic is selected, and a number of selected emotional words and emotional vocabulary are used to form a basic emotional dictionary;

Then, the basic sentiment dictionary is expanded with synonyms and candidate words with sentimental tendencies.

Further, carry out emotion weight statistics on the sentences between each delimiter, and judge the emotional polarity of the text according to the proportion of negative emotional weight in all emotional word weights, including:

First of all, for the sentences between the delimiters, the emotional tendency is counted through the analysis of emotional words, negative words, adverbs, fixed collocations, transition words and exclamation sentences;

Then, the statistical text contains the sum of the negative sentiment tendency values of all clauses and the sum of the absolute value of the overall sentiment weight, and uses the weight of negative sentiment words in the vertical proportion of all sentiment words in the text to judge the sentiment polarity of the text.

You can select a series of emotional words from Hownet, input them into the search engine one by one, sort the emotional words according to the number of hits (hits value) returned by the search engine, and select several emotional words with the highest hits as the basis In addition, based on the method of word frequency statistics, the basic emotional words with higher relevance to the topic are selected semi-automatically to form the basic emotional dictionary. Because most of the words containing emotional components in the text are adjectives, verbs, and some nouns, after preprocessing, it is only necessary to perform word frequency statistics based on the automatic text with enough entries, and then select the words with the highest word frequency for some words with high word frequency The 20 positive emotional words and the 20 negative emotional words with the highest word frequency, together with the common basic emotional vocabulary, constitute the basic emotional dictionary.

Since the basic sentiment dictionary expresses a strong emotional tendency, the negative sentiment words in the basic sentiment dictionary can be given a sentiment tendency value of -1. The vocabulary in the basic emotional dictionary is small, and it is impossible to contain all the words with emotional tendencies that appear in the text set. Therefore, it is necessary to expand the basic emotional dictionary to build a relatively complete emotional dictionary. You can add synonyms and add candidate words with emotional tendencies for expansion.

Adding synonyms can help identify emotional vocabulary more broadly, and use the existing thesaurus to expand the basic emotional dictionary with synonyms. However, in order to improve the algorithm performance of emotional tendency calculation, it is still necessary to manually screen out commonly used synonyms. After the expansion, the number of words in the sentiment dictionary increases to 256, and the emotional tendency value of synonyms of negative emotional words can be set to -1.

It is very difficult to construct a completely exhaustive sentiment dictionary, but by analyzing the correlation between each word in the text set and the vocabulary in the sentiment dictionary, and incorporating highly correlated words into the dictionary, an emotional dictionary with wider coverage can be effectively constructed.

Pointwise Mutual Information can be used to calculate the correlation of emotional words in the candidate word dictionary, so as to determine whether to add it to the emotional dictionary. The point mutual information method calculates the correlation between words and words based on mutual information theory. The basic idea is to count the co-occurrence probability of two words wordi and wordj in the text. The higher the co-occurrence probability, the higher the correlation between the two words. The calculation formula is as follows:

Where p(wordi^wordj) is the probability of co-occurrence of wordi and wordj in the text, the calculation method is as follows:

Among them, n represents the total number of clauses in the text, and numSentence(wordi, wordj) represents the number of clauses containing both wordi and wordj. P(wordi) and P(wordj) respectively represent the proportion of the number of clauses containing wordi and wordj in the total number of clauses in the text. Calculated as follows:

Among them, numSentence(wordi) indicates the number of clauses containing wordi in the text. In the above formula, PMI(wordi, wordj) means that when one of the variables wordi and wordj appears, the amount of information of the other variable that can be obtained fully shows the statistical correlation between wordi and wordj: when PMI is greater than 0, it means Two words are related, and the larger the PMI value, the stronger the correlation; when the PMI is 0, it means that the two words are statistically independent; when the PMI is less than 0, it means that the two words are mutually exclusive .

The part-of-speech property of the word can be obtained after word segmentation using the ICTCLAS system, and then calculated by word.propertyal∈{a,d,an,ag,al} and word.propertyal∈{vn,vd,vi,vg,vl}, defined by The SO-PMI values of the two candidate words, and the remaining words of the part of speech are directly regarded as neutral words. This method aims to solve the co-occurrence of some words with no emotional tendency and positive or negative emotional words when adding related words The probability is very high, leading to wrong introduction into the sentiment dictionary, causing unnecessary overhead in the performance of sentiment classification, reducing the accuracy problem, and improving the efficiency of the dictionary expansion algorithm. Calculating the SO-PMI value of the two candidate words is as follows: calculating the PMI value of the candidate word and the positive basic dictionary, calculating the PMI value between the candidate word and the negative word, and finally combining the two to obtain the SO-PMI value of the candidate word PMI value, the calculation formula is as follows:

SO-PMI(word)＝

∑ _{posWord∈posWords} PMI(word,posWord)-∑ _{negWord∈negWords} PMI(word,negWord)

The relationship between the value of SO-PMI and emotional tendency can be adjusted as

In summary, the sentiment dictionary expansion method is summarized as follows:

posWords:

If word is a positive word in the basic sentiment dictionary, why is word included in posWords;

If word is a synonym for a positive word in the basic sentiment dictionary, then word is included in posWords;

If word conforms to the formula word.propertyal∈{a,d,an,ag,al} or word.propertyal∈{vn,vd,vi,vg,vl}, and 1.36<SO-PMI(word)<23, word is included posWords.

Similarly, negWords:

If word is a negative word in the basic sentiment dictionary, then word is included in negWords;

If word is a synonym for a negative word in the basic sentiment dictionary, then word is included in negWords;

If word conforms to word.propertyal∈{a,d,an,ag,al} or word.propertyal∈{vn,vd,vi,vg,vl}, and -16<SO-PMI(word)<-1, then Word is included in negWords.

Based on the sentiment dictionary, each text sentence S is taken as a unit, and each sentiment word WS in the sentence is used as a delimiter, and the sentiment weight of the sentence phrase (WSi-1, WSi) between the two delimiters is calculated. Calculate, sentence sentence phrase(WSi-1,WSi) contains the word WSI but does not contain the word WSi-1; the model consists of 5 modules, namely: analysis of emotional words, analysis of negative words, analysis of adverbs, and fixed collocations Analysis of transition words, analysis of exclamatory sentences.

Analysis of emotional words: For each word word in the text to be analyzed, scan the emotional dictionary to determine whether the word exists in the emotional dictionary, if it exists, treat the word as an emotional word and read it from the negative emotional dictionary The sentiment value of the word is returned; if it does not exist, the word is regarded as a neutral word and 0 is returned, and this cycle is repeated until the word judgment of the entire text set is completed. By calculating the emotional tendency value of each word, we obtain accurate emotional words (that is, words whose weight is not equal to 0), and filter out emotional words that do not play an emotional role in a specific sentence (that is, words whose weight is equal to 0 emotional words).

Negative word analysis: in the case where the emotional word Wsi appears in the sentence, calculate the number negNum(Wsi-1, Wsi) of negative words between Wsi and the previous separator Wsi-1 (that is, in a sentence sentence). If negNum is an odd number, the sentiment value of the clause is reversed from that of the emotional word; otherwise, the original sentiment value is maintained.

Adverb analysis: determine whether the vocabulary is in the adverb dictionary, if so, obtain the adverb emotional strength from the adverb dictionary, and multiply the corresponding weight by the current emotional tendency value of the clause as the clause emotional weight.

Turning word analysis: scan backward from the current emotional word Wsi to find the next emotional word Wsi+1. In this process, if a turning word is scanned, the weight(phrase(Wsi-1, Wsi)) is reversed, The emotional tendency of phrase(Wsi-1, Wsi) is biased towards the emotional tendency of phrase(Wsi, Wsi+1) behind the transition word.

Exclamatory sentence analysis: For the analysis of exclamatory sentences, we use the exclamation mark "!" as the identifier of the exclamatory sentence, and record it as exc. The calculation method of its emotional weight is: when the exclamation point is scanned, we search for the emotional word Wsi-1 closest to the exclamation point from the back to the front, and use the emotional tendency value of Wsi-1 as the weight of exc.

Calculate the sum weight(S) of the negative emotional tendency values of all clauses contained in a text S and the sum total(S) of the absolute value of the overall emotional weight, and calculate the weight of negative emotional words in the weight of all emotional words in the text Medium proportion scale(S), scale(S)=weight(S)/total(S), judge the emotional polarity of the text S according to the scale(S), and make a decision on the cognitive threat nature of the text according to the emotional polarity A preliminary judgment is made, the text with scale(S) in the [0.68-1] interval is regarded as a cognitive threat with a high probability; the scale(S) in the [0-0.68) interval is a suspected cognitive threat topic text, So far, the first stage classification of text cognitive threat has been realized.

As a preferred embodiment, further, in the middle-level detection, the initial suspected cognitive threat topic text is classified into cognitive threat topic text, suspected cognitive threat topic text and non-cognitive threat topic text by using a deep learning method. The classification process includes:

Construct a deep learning model and use the training data set with labels for pre-training, where the deep learning model includes a BERT model for word vector representation of the input, and cognitive threat detection for the input word vector The BiLSTM model;

Input the initial suspected cognitive threat topic text into the pre-trained deep learning model, use the deep learning model to obtain the cognitive threat probability value, and determine the cognitive threat in the initial suspected cognitive threat topic text through the cognitive threat probability value Topic text, suspected cognitive threat topic text, and non-cognitive threat topic text.

In the embodiment of this case, according to the respective characteristics of the two methods, the two methods based on emotional dictionary and deep learning are combined and optimized, and a multi-dimensional sentiment analysis method combining dynamic expansion of emotional dictionary and deep learning is proposed to overcome the two methods. Each method has its own disadvantages and achieves a higher accuracy rate.

Cognitive threat recognition based on sentiment analysis performs sentiment analysis on text in two stages. In the first stage, you can refer to the existing HowNet emotional dictionary and BosonNLP emotional dictionary, and use the method based on word frequency statistics to build a basic emotional dictionary, and calculate the statistical correlation between candidate words and vocabulary in the basic emotional dictionary to judge their emotions Tendency to realize the dynamic expansion of the sentiment dictionary. On the basis of sentiment dictionary, negative dictionary, and degree adverb dictionary, each text S is taken as a unit, and each sentiment word WS of the sentence is used as a separator, and the sentence phrase(WSi-1, WSi) calculates the sum of negative emotional weights weight (S) and the sum total (S) of the absolute value of emotional word weights, and defines scale (S) as the proportion of negative emotional weights in all emotional word weights, Judging the emotional polarity of the text S according to the size of scale(S), making a preliminary judgment on the nature of the cognitive threat of the text based on the emotional polarity, and completing the initial recognition of the cognitive threat. The statistical results of the analysis of a large number of collected experimental texts show that texts with a negative emotional weight scale (S) in the [0.68-1] interval are most likely to be cognitive threats; negative emotional weight scale (S) is in [0-0.68) interval is the topic text of suspected cognitive threat; texts whose emotional tendency value is in the [0-0.68) score interval are initially classified as suspected cognitive threat, and the second-stage identification process is performed on it. In the second stage, the BERT+BiLSTM deep learning model can be used as the core to conduct sentiment analysis, further analyze the text's emotional tendency, and complete the re-identification of cognitive threats, using BERT (BidirectionalEncode, Reprsesnation from Transformers, BERT) pre-trained word vectors to replace traditional The word vector trained by the method converts the text processed by word segmentation into a multi-dimensional word vector, and adopts the bidirectional long-short-term memory network (BiLSTM) model that can solve the problem of short-term dependence and long-term dependence to form the core of the emotional tendency analysis of this section. The artificially labeled cognitive threat topic text set and the known cognitive threat topic texts under the same topic are used as the training set to train the BERT+BiLSTM model, and the trained model is used to further analyze the suspected cognitive threat topic texts obtained in the first stage. Sentiment analysis, using the Softmax classifier to divide the text into a set of confirmed cognitive threat topic text sets, a suspected cognitive threat topic text set and a non-cognitive threat topic text set.

The sentiment analysis result based on the dynamic expansion of the emotional dictionary in the first stage is a second-stage identification process for texts that are suspected of cognitive threats, and the BERT+BiLSTM deep learning model is used as the core for further identification of cognitive threats. Firstly, model training is carried out, and the training process is as follows: the training data set can be manually labeled first, whether the label has the nature of cognitive threat, and the BERT model is used to represent the word vector after word segmentation, and finally the converted vector is passed to the BiLSTM neural network network. According to the cognitive threat samples, the BiLSTM model covering the cognitive threat samples is trained. The first-stage processing results are suspected cognitive threat texts through BERT word vectorization, and the converted vectors are respectively passed into the cognitive threat model, and a cognitive threat probability value will be obtained through the model. Through a large number of text experiments, it is shown that, The text with the probability of training results in (0.68-1] can be determined to have the nature of cognitive threat, the probability of (0.32-0.68] is a suspected cognitive threat, which needs manual identification, and the probability of [0-0.32] is non-cognitive threat.

The experimental results show that when the data set contains nearly 5,000 microblog text data, the cognitive threat recognition accuracy rates of purely based on deep learning and purely based on sentiment lexicon are 67.9% and 83.27%, respectively. The accuracy rate of comprehensive cognitive threat recognition based on multidimensional sentiment analysis is 89.9%, which is relatively good.

Furthermore, in the embodiment of this case, for the preprocessed sensitive topic text data, the cognitive threat topic text is obtained through multi-level cognitive threat detection, which also includes: using the sentiment analysis method in the comment area of the cognitive threat topic text The overall emotional tendency to assess the impact of cognitive threats.

Define the influence degree of cognitive threat through the overall emotional tendency of the comment area under the text that has been identified as cognitive threat, merge all the comment texts under a text that has been identified as cognitive threat, after data preprocessing and text word segmentation , the cognitive threat recognition method based on the emotional dictionary mentioned above can be used to judge the overall emotional tendency of the text in the comment area, and the comment orientation triggered by the text of the cognitive threat nature can be used as the basis for judging the threat degree. The threat degree of the text can be evaluated, and the threat degree can be divided into one, two, and three levels from high to low, and the overall negative emotional weight of the comment text is in the range of [0.68-1] The text cognitive threat degree is defined as the first level; the overall negative emotional weight of the comment text is defined as the second level when the overall negative emotional weight of the text is between [0.32-0.68]; the overall negative emotional weight of the comment text is between The text cognitive threat degree is defined as three levels when the weight ratio of the overall emotional words in the text is [0-0.32), and the analysis results can provide an important reference for dealing with cognitive threats.

As a preferred embodiment, further, construct a cognitive threat communication knowledge map through named entity recognition and entity relationship extraction of the cognitive threat topic text, including:

Construct a named entity extraction model, and optimize the model by using an adversarial training method. The named entity recognition model includes an encoder for mapping input characters to real number space and mining potential semantics, and is used to convert vectors to The BiLSTM neural network layer for extracting contextual semantic information by forward and backward bidirectional features, and the CRF conditional random field layer for taking the bidirectional features extracted by the BiLSTM neural network layer as input and combining the Bioes labeling paradigm to generate the corresponding label of the character;

The cognitive threat topic text is used as the input of the optimized named entity extraction model, and the named entity extraction model is used to identify the entity categories and relationships in the cognitive threat topic text.

At present, it is more common to use methods based on statistical machine learning to achieve named entity recognition tasks. The named entity extraction model in this case uses the BERT-BiLSTM-CRF model, which is developed based on the BiLSTM-CRF model. The end-to-end deep learning model can meet the current requirements of Chinese address parsing and address element labeling tasks. The model consists of encoder (Transformer), BiLSTM neural network layer and conditional random field (CRF) layer from bottom to top. The Transformer encoder is based on the character-level Chinese BERT model, which maps the input Chinese address characters into a low-dimensional dense real number space, and mines the potential semantics contained in various address elements in Chinese addresses; the BiLSTM neural network layer converts the encoder into The character vector from the source is used as input to capture the bidirectional features of the Chinese address sequence forward (from left to right) and backward (from right to left), which can fully obtain the semantic information of the context; the CRF conditional random field layer belongs to the probability graph model, The bidirectional feature extracted by the upstream BiLSTM is used as input, and the label corresponding to each character in the address is generated in combination with the Bioes labeling paradigm, so that the Chinese address is further parsed into various address elements according to the label, and the sequence problem is considered in the calculation process, which can be very fast. Improve the recognition effect of named entities to a great extent.

There is no established standard for entities in the field of cognitive threats, and most of the existing network naming and recognition tasks are only for network public opinion recognition. In the example of this case, the crawled data is analyzed, and according to the cognitive threat identification requirements, a total of 6 types of entities in the cognitive threat field can be set, which are users, time, address, platform, organization, and hot events, as shown in Table 1 Shown:

Table 1 Cognitive Threat Entity Types

Entity labeling is the most important issue in named entity recognition tasks, and it is also the basis of model training. Commonly used annotation methods are BIO and BIOES. Although the BIOES labeling method provides more information, more labels need to be predicted. Due to the limited number of data sets constructed in this case, the effect of using the BIOES labeling method may be affected. In the BIO notation system, "B" can be used to mark the beginning of the entity, "I" can be used to mark the inside of the entity, and "O" can be used to mark the non-entity. The labels of each type of entity include "start" and "internal". Therefore, the named entity recognition dataset constructed in this case can be set to 13 labels.

Cognitive threat user knowledge map related entities have complex relationships. When extracting knowledge from cognitive threat topic texts, we should also pay attention to the dependencies between adjacent tags. However, since BiLSTM is good at dealing with long-distance text information, it cannot handle the dependencies between adjacent labels. Therefore, it can be combined with CRF (Conditional Random Field) in the knowledge extraction of cognitive threat user knowledge graphs in BiLSTM output On the basis of the prediction label initially corresponding to each word, the output score is corrected through the relationship of the adjacent labels to obtain an optimal prediction sequence.

The CRF layer takes the output score of the previous layer BILSTM as input, and outputs the most likely sequence of predicted annotations that meet the annotation transfer constraints. For any sequence X=(x1,x2,...,xn); here it is assumed that P is the output score matrix of BiLSTM, and the size of P is n\times k, where n is the number of words, k is the number of labels, Pij represents the score of the j-th label of the i-th word. For the prediction sequence Y=(y1,y2,...,yn), its score function is:

A represents the transfer score matrix, Aij represents the score of label i transferred to label j, and the size of A is k+2. The probability of predicting sequence Y produced is:

Take the logarithm at both ends to get the likelihood function of the predicted sequence:

In the formula, \widetilde{Y} represents the real label sequence, and YX represents all possible label sequences. The output sequence to get the maximum score after decoding:

The output of the CRF layer is the optimal label sequence of the cognitive threat topic text, focusing on the words corresponding to the tags such as reposting user, reposting time, reposting location, reposting platform, text summary information, and text topic. The user knowledge map is also the basis for traceability of forwarding users, forwarding process, forwarding time, forwarding platform and other traceability processes, as well as cognitive threat forwarding user relationship reasoning.

When using BERT and its variants, because it has been pre-trained, the parameters have reached a good level, in order to maintain the training effect, a lower learning rate should be used; on the contrary, because its downstream tasks have been pre-trained, if you set A low learning rate not only makes the training process slow, but also difficult to synchronize with BERT training. Therefore, in the embodiment of this case, a hierarchical learning rate strategy can be adopted: set a smaller learning rate for the upstream BERT pre-training layer, and set a larger learning rate for the lower layer.

In the process of model training, when the loss value declines gradually, if a large learning rate is still used, it will cause the model to oscillate around the optimal point when it converges to the global optimal point. In order to ensure that the loss function is always kept away from the optimal In the range of very close to the value, and gradually approaching the optimal value, it is necessary to adopt the learning rate decay strategy, that is, to reduce the step size of the parameter update. In this case, a learning attenuation strategy can be set: during the training process, when the model effect does not improve, reducing the learning rate can effectively improve the model accuracy.

BERT-BiLSTM-CRF is used as a named entity recognition model, but due to the local instability of the neural network, even a small disturbance may cause a large error to the model. Therefore, the embodiment of this case adopts the confrontation training method to optimize the model. Adversarial training improves the robustness of the model by inputting small perturbations into the model, which can alleviate the defects of local instability of the neural network and improve the robustness of the model. See Figure 3. During the training process, BERT will first generate an initial vector for the input text, and then add some perturbations to the initial vector to generate adversarial samples. These adversarial samples are variants of the original samples, which can easily mislead the model. The initial vector and adversarial samples will be input into BiLSTM for training, and the neural network will learn more robust parameters during the training process to resist adversarial sample attacks.

As a preferred embodiment, further, in constructing the cognitive threat propagation knowledge map through named entity recognition and entity relationship extraction of the cognitive threat topic text, two named entity extraction models connected by pipeline can be built, wherein the first The first named entity extraction model adopts the single-label multi-classification task method to identify entities in the cognitive threat topic text, and the second named entity extraction model uses the multi-label multi-classification task method to use the input of the first named entity extraction model as input to identify entities relationship between.

Knowledge fusion is an important task in building a domain knowledge graph. It makes it a whole through the alignment, association and merging of multiple related entities. The main work is divided into two parts: entity unification and entity disambiguation. Due to the political and offensive characteristics of cognitive threat topic texts, there is a problem of inconsistency between entities identified by named entities, so entity unification and entity disambiguation are required. Among them, entity unification refers to different entity examples with the same meaning, and entity unification is required.

The ambiguity of named entities means that an entity referent can correspond to multiple real-world entities. Due to the richness and complexity of Chinese semantics, the same word may have different meanings in different contexts, so it is necessary to carry out entity Disambiguation. In this case, a link-based entity disambiguation method can be used to link entity references to corresponding entities in the knowledge base. After entity unification, the final effective entity can be obtained.

Graph databases are good at handling large volumes of complex, interconnected, low-structured data that changes rapidly and requires frequent queries—in relational databases, these queries would result in a large number of table joins and thus create performance issues. However, the traditional RDBMS, etc. have performance degradation problems when querying. Therefore, in the embodiment of this case, the persistence engine Neo4j that supports complete transactions can be used, which provides large-scale scalability and can handle billions of nodes on one machine. A relation-property graph that can be extended to run on multiple machines in parallel. At the same time, focus on solving the problem of performance degradation. By modeling data around graphs, Neo4j traverses nodes and edges at the same speed, independent of the amount of data that makes up the graph.

As shown in Figure 4, when realizing the visual display of the knowledge graph based on Neo4j, a set of visual graph elements can be defined in the Neo4j program. The schema of cognitive threat propagation can be mainly expressed by type and property. Define users, events, addresses, platforms, hot events, sentiment labels, and threat intents as entities. In terms of relationship definition, the following relationships can be defined: text-hot event, user-text, forwarding, user-organization, text-emotional label, text-threat intent, etc. A triple can be expressed as: <text, text-hot event, hot event>, <user, user-cognitive threat topic text, text>, <user, forward, user>, <user, user-organization, user >, <text, text-threat-intent, threat-intent>, etc.

By constructing a cognitive threat communication knowledge map, the cognitive threat communication user traceability, event traceability, and organization traceability are realized. Through implicit relationship mining, real-time monitoring of key accounts, groups, organizations, and users can provide a basis for precise positioning and targeted blocking of cognitive threats.

Further, based on the above method, an embodiment of the present invention also provides a social media cognitive threat detection system, including: a data collection server, multiple cognitive threat identification servers, a knowledge graph server, and a web server, wherein,

The data collection server is used to collect text data of sensitive topics on the network platform and perform preprocessing operations on the data;

Multiple cognitive threat identification servers are used to obtain the cognitive threat topic text through multi-level cognitive threat detection for the preprocessed sensitive topic text data, wherein the multiple cognitive threat identification servers specifically include: sensitive topic The text data is divided into cognitive threat topic text and initial suspected cognitive threat topic text, and the primary identification server classifies the initial suspected cognitive threat topic text into cognitive threat topic text, suspected cognitive threat topic text and non-cognitive threat topic text An intermediate authentication server for the text, and an ultimate authentication server for obtaining the cognitive threat topic text from the suspected cognitive threat topic text through manual annotation;

The knowledge map server is used to build a cognitive threat communication knowledge map through named entity recognition and entity relationship extraction of cognitive threat topic texts;

The web server is used to disseminate the knowledge graph based on the cognitive threat and use the web interactive interface to trace the user source, event source and organization source of the cognitive threat topic text dissemination.

The front end can be designed based on JavaScript, and Echart can be used to realize data visualization. Multiple cognitive threat identification servers can be set to be decentralized and distributed, and are jointly responsible for the detection and measurement of cognitive threat information. Only texts that are recognized by multiple servers as cognitive threats will be judged as cognitive threat topic texts. After the text is judged as a cognitive threat topic text, the cognitive threat identification server uploads the text information and the text cognitive threat attribute measurement information to the distributed network. The distributed structure not only improves the accuracy of cognitive threat detection, but also improves the system's ability to resist risks. The damage of one server will not affect the operation of the entire system.

The knowledge graph server corresponds to the cognitive threat knowledge extraction module and the cognitive threat propagation knowledge graph building module. The knowledge map server automatically accesses the distributed network through smart contracts, extracts the text information of the detected cognitive threat topic text, and detects cognitive threat entities such as users, time, address, organization, forwarding platform, and hot events related to the cognitive threat topic text. Carry out named entity recognition and relationship extraction, and build a cognitive threat propagation knowledge map through Neo4j.

The human-computer interaction interface can use the Web to realize the bridge between users and data, and use the Echart visualization tool to convert abstract data and relationships into intuitive charts.

In the embodiment of this case, the system has good security and feasibility, and the complexity of modular operation is low, which is convenient for maintenance. And verified by experimental data, the accuracy rate of threat determination for topic texts within 100 words reaches 93%, which has a high accuracy rate and a good detection effect. In addition, the proposal in this case has a wide range of application scenarios, and can be used in news media supervision, network public opinion supervision, and crackdown on illegal activities.

Relative steps, numerical expressions and numerical values of components and steps set forth in these embodiments do not limit the scope of the present invention unless specifically stated otherwise.

Each embodiment in this specification is described in a progressive manner, each embodiment focuses on the difference from other embodiments, and the same and similar parts of each embodiment can be referred to each other. As for the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and for the related information, please refer to the description of the method part.

The units and method steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, in the above description The composition and steps of each example have been generally described in terms of functions. Whether these functions are performed by hardware or software depends on the specific application and design constraints of the technical solution. Those of ordinary skill in the art may use different methods to implement the described functions for each particular application, but such implementation is not considered to exceed the scope of the present invention.

Those of ordinary skill in the art can understand that all or part of the steps in the above method can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium, such as: a read-only memory, a magnetic disk or an optical disk, and the like. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Correspondingly, each module/unit in the above embodiments can be implemented in the form of hardware, or can be implemented in the form of software function modules. The form is realized. The present invention is not limited to any specific combination of hardware and software.

Finally, it should be noted that: the above-described embodiments are only specific implementations of the present invention, used to illustrate the technical solutions of the present invention, rather than limiting them, and the scope of protection of the present invention is not limited thereto, although referring to the foregoing The embodiment has described the present invention in detail, and those of ordinary skill in the art should understand that any person familiar with the technical field can still modify the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention Changes can be easily thought of, or equivalent replacements are made to some of the technical features; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the scope of the present invention within the scope of protection. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims (10)

1. A social media cognitive threat detection method, is characterized in that, comprises the following content: Collect text data of sensitive topics on the network platform and perform preprocessing operations on the data; For the preprocessed sensitive topic text data, the cognitive threat topic text is obtained through multi-level cognitive threat detection, wherein the multi-level cognitive threat detection includes: dividing the sensitive topic text data into cognitive threat topic text and initial suspected Primary detection of cognitive threat topic texts, classifying initial suspected cognitive threat topic texts into cognitive threat topic texts, suspected cognitive threat topic texts, and intermediate detection of non-cognitive threat topic texts; Threat topic text to obtain the ultimate detection of cognitive threat topic text; Construct a knowledge map of cognitive threat communication through named entity recognition and entity relationship extraction of cognitive threat topic texts; Based on the knowledge map of cognitive threat communication, user tracing, event tracing and organizational tracing are carried out on cognitive threat topic text dissemination. 2. the social media cognitive threat detection method according to claim 1, is characterized in that, collects network platform sensitive topic text data and carries out preprocessing operation to data, comprises: First, according to the distributed collection of user authorization information database, sensitive topic text information and related user data on the network platform; Then, for the collected text information, merge the title with the text, use the redundancy detection algorithm to remove redundant information, de-duplicate the relevant comments, clean and transform the text noise data, and use the word segmentation system to segment the text . 3. The social media cognitive threat detection method according to claim 1, wherein, in primary detection, sensitive topic text data is divided into cognitive threat topic text and initial suspected cognitive threat topic text by using sentiment analysis method, Among them, the process of sentiment analysis method division includes: First, based on the known emotional dictionary and using the word frequency statistics method to construct the basic emotional dictionary, and expand the emotional dictionary by making correlation statistics between the words in the text data and the basic emotional dictionary; Then, take the text in the sensitive topic text data as the unit and the emotional word as the delimiter, carry out the sentiment weight statistics on the sentences between each delimiter, and calculate the negative sentiment weight according to the proportion of the negative sentiment weight in all the sentiment word weights. Judge the emotional polarity of the text; Then, according to the emotional polarity of the text, the sensitive topic text data is divided into cognitive threat topic text and initial suspected cognitive threat topic text. 4. the social media cognitive threat detection method according to claim 3, is characterized in that, according to known sentiment lexicon and utilizes word frequency statistical method to construct basic sentiment lexicon, comprises: First, select a series of emotional words in the known emotional dictionary, sort the emotional words according to the search engine clicks in the series of emotional words, and select some emotional words according to the popularity of the clicks; Then, based on the word frequency statistics, the emotional vocabulary with the highest relevance to the topic is selected, and a number of selected emotional words and emotional vocabulary are used to form a basic emotional dictionary; Then, the basic sentiment dictionary is expanded with synonyms and candidate words with sentimental tendencies. 5. the social media cognitive threat detection method according to claim 3, is characterized in that, carry out emotion weight statistics to the segmented sentence between each delimiter, according to negative emotion weight in all emotion word weights The proportion is used to judge the emotional polarity of the text, including: First of all, for the sentences between the delimiters, the emotional tendency is counted through the analysis of emotional words, negative words, adverbs, fixed collocations, transition words and exclamation sentences; Then, the statistical text contains the sum of the negative sentiment tendency values of all clauses and the sum of the absolute value of the overall sentiment weight, and uses the weight of negative sentiment words in the vertical proportion of all sentiment words in the text to judge the sentiment polarity of the text. 6. The social media cognitive threat detection method according to claim 1, characterized in that, in the intermediate detection, the initial suspected cognitive threat topic text is classified into cognitive threat topic text, suspected cognitive threat topic text by using deep learning method Text and non-cognitive threat topic text, the classification process includes: Construct a deep learning model and use the training data set with labels for pre-training, where the deep learning model includes a BERT model for word vector representation of the input, and cognitive threat detection for the input word vector The BiLSTM model; Input the initial suspected cognitive threat topic text into the pre-trained deep learning model, use the deep learning model to obtain the cognitive threat probability value, and determine the cognitive threat in the initial suspected cognitive threat topic text through the cognitive threat probability value Topic text, suspected cognitive threat topic text, and non-cognitive threat topic text. 7. The social media cognitive threat detection method according to claim 1, wherein, for the preprocessed sensitive topic text data, the cognitive threat topic text is obtained through multi-level cognitive threat detection, further comprising: utilizing The sentiment analysis method evaluates the impact of cognitive threats based on the overall emotional tendency in the comment area of cognitive threat topic texts. 8. The social media cognitive threat detection method according to claim 1, wherein the cognitive threat propagation knowledge map is constructed through named entity recognition and entity relationship extraction of the cognitive threat topic text, including: Construct a named entity extraction model and optimize the model by using an adversarial training method. Among them, the named entity recognition model includes an encoder for mapping input characters to real number space and mining potential semantics, and is used to convert vectors to The BiLSTM neural network layer for extracting contextual semantic information by forward and backward bidirectional features, and the CRF conditional random field layer for taking the bidirectional features extracted by the BiLSTM neural network layer as input and combining the Bioes labeling paradigm to generate the corresponding label of the character; The cognitive threat topic text is used as the input of the optimized named entity extraction model, and the named entity extraction model is used to identify the entity categories and relationships in the cognitive threat topic text. 9. The social media cognitive threat detection method according to claim 8, characterized in that, in constructing the cognitive threat propagation knowledge map through named entity recognition and entity relationship extraction of the cognitive threat topic text, a pipeline connection is set up There are two named entity extraction models, in which, the first named entity extraction model uses a single-label multi-classification task method to identify entities in cognitive threat topic texts, and the second named entity extraction model adopts a multi-label multi-classification task method. The first named entity extraction model input is used as input to identify relationships between entities. 10. A social media cognitive threat detection system, characterized in that it comprises: a data acquisition server, a plurality of cognitive threat identification servers, a knowledge graph server and a web server, wherein, The data collection server is used to collect text data of sensitive topics on the network platform and perform preprocessing operations on the data; Multiple cognitive threat identification servers are used to obtain the cognitive threat topic text through multi-level cognitive threat detection for the preprocessed sensitive topic text data, wherein the multiple cognitive threat identification servers specifically include: sensitive topic The text data is divided into cognitive threat topic text and initial suspected cognitive threat topic text, and the primary identification server classifies the initial suspected cognitive threat topic text into cognitive threat topic text, suspected cognitive threat topic text and non-cognitive threat topic text An intermediate authentication server for the text, and an ultimate authentication server for obtaining the cognitive threat topic text from the suspected cognitive threat topic text through manual annotation; The knowledge map server is used to build a cognitive threat communication knowledge map through named entity recognition and entity relationship extraction of cognitive threat topic texts; The web server is used to disseminate the knowledge graph based on the cognitive threat and use the web interactive interface to trace the user source, event source and organization source of the cognitive threat topic text dissemination.

Images