A tracking and tracing method based on network information

Abstract

The present invention relates to the field of basic network and information security, and specifically discloses a tracking and tracing method based on network information. The present invention is used for the security monitoring of network information, including: determining key sections related to specific public opinion events; extracting discussion contents within each topic, extracting keywords in the discussion contents of each post, and establishing an event corpus; constructing a sub-corpus of public opinion eigen-objects and a sub-corpus of public opinion derivative objects; constructing an LSTM model and training the LSTM model; using the LSTM model to classify each piece of discussion content into discussions on public opinion eigen-objects and discussions on public opinion derivative objects; determining the total duration of the discussion on the public opinion event, dividing time periods based on the total duration of the discussion, determining the public opinion inflection point based on the discussion, comparing the accounts and remarks near the inflection point, and obtaining associated accounts during the public opinion inflection point period.

Description

Technical Field

[0001] The present invention relates to the field of basic network and information security, and in particular to a tracking and tracing method based on network information. Background Art

[0002] With the development of the Internet, online speech has increasingly affected people's lives, and even cyber violence has occurred. Moreover, some activities are accompanied by the participation of online trolls.

[0003] At present, the research on online events is mainly focused on the rationality of the speech itself, and the main means used are blocking certain specific keywords and restricting sensitive words. However, the update iteration of online vocabulary is extremely rapid, and the update of blocked sensitive words is often not as fast as the development of online vocabulary.

[0004] The applicant noted in the study that the fermentation of network events often has two characteristics: one is the change of the object of network content discussion, and the other is the change of the direction of network discussion. In order to better prevent network events from causing unnecessary harm to social life, it is of great significance to track the development of network events in a timely manner and quickly judge the change of the subject of network discussion. However, there is no relevant report on the current network tracking technology. Summary of the invention

[0005] In view of the above problems existing in the prior art, the present invention proposes a method for tracing the source of guiding speech in network events. The present invention can quickly determine the occurrence of the turning point of the event, screen out the account or account cluster that initiated the key speech that has a certain guiding effect on the public opinion, and issue an early warning for the occurrence of public opinion events, and provide a data basis for finding network water army or organized public opinion propaganda.

[0006] During the research, the applicant found that guiding speech often does not exist in isolation, and that it is often not simply insulting or negative language that plays a decisive role in guiding public opinion, but speech that often has a certain directional change process. Therefore, the applicant studied the directional change process of public opinion from this perspective and found that the directional changes of public opinion of some accounts have certain similarities. Such directional guidance of public opinion is very unfavorable, and it is of great significance to find such directional guidance accounts.

[0007] Specifically, the present invention provides a tracking and tracing method based on network information, the method comprising:

[0008] (3.1) Determine the key sectors related to a specific public opinion event, capture the discussion content of relevant topics within any key sector, and obtain the corresponding relationship pairs between the account information of the public opinion expression subject, the public opinion expression content, and the public opinion expression time within the corresponding sector;

[0009] (3.2) Extract the discussion content within each topic, extract the keywords in the discussion content of each post, and establish an event corpus;

[0010] (3.3) Extract the words expressing the intrinsic object of public opinion and the words expressing the derivative object of public opinion in the corpus, and construct an intrinsic object corpus of public opinion and a derivative object corpus of public opinion;

[0011] (3.4) Construct an LSTM model, select the discussion content about the object of public opinion for marking, mark them as the intrinsic object of public opinion and several derivative objects of public opinion respectively, and use the marked text to train the LSTM model;

[0012] (3.5) Use the LSTM model to classify each discussion content into discussions about the intrinsic object of public opinion and discussions about the derivative object of public opinion;

[0013] (3.6) Determine the total discussion time of the public opinion event, divide the time period based on the total discussion time, divide all discussion content according to multiple time nodes into several public opinion stages. For each public opinion stage, perform weighted summation on the directions of each public opinion discussion object within this stage to obtain the public opinion direction vector of this stage. In the public opinion direction vector, for each object of public opinion, take it as a dimension and assign a dimension modulus value. The dimension modulus value is proportional to the frequency of discussion of this object. For the directionality of the expression statements of the discussion of this object, based on the thesaurus of directional expressions, assign values to it. Take the ratio of the difference between positive discussions and negative discussions to the total discussion as the discussion direction angle of this object. Determine the target object of public opinion according to the magnitude order of the dimension modulus values of the objects of public opinion;

[0014] For the target object of public opinion, take the time when the public opinion direction passes through zero as the inflection point, and intercept the accounts and speech contents participating in the discussion of this sector within the predetermined time range before and after the inflection point;

[0015] (3.7) For the speech contents within the inflection point area range, extract all the speech contents of the same account about the relevant object within the inflection point range in chronological order. For each speech content, determine the directional angle value of the speech content, and construct a directional angle change curve of this account with time as the horizontal axis;

[0016] (3.8) Crawl the discussion content on the current public opinion platform regarding relevant topics in other key sections of this public opinion event, and repeat the above steps (3.1)-(3.8) to construct the directional angle change curves of all accounts related to inflection points within the corresponding section.

[0017] (3.9) Perform similarity matching on the directional angle change curves of all obtained accounts, and extract the account information with a similarity of the directional angle change curve higher than the set threshold.

[0018] In a preferred implementation manner, for all content whose discussion content has not been clearly determined to be a derivative object, it is all considered as the discussion of the main and secondary objects in this discussion item.

[0019] In a preferred implementation manner, the step (3.2) includes: using the Word2vec model to convert the obtained corpus, and representing the words in the corpus in the form of one-hot encoding vectors.

[0020] In another preferred implementation manner, the discussion directions of the object include positive, negative, and neutral, and different scores are assigned to different corpora.

[0021] In another preferred implementation manner, in step (3.6), the discussion direction angle value of the object is calculated based on the following formula:

[0022] n represents the number of subjects participating in the discussion of this object, i ∈ [1, n], P i is the assignment for this object based on positive discussion content, N i is the assignment for this object based on negative discussion content, B i is the assignment for this object based on negative discussion content, and σ represents the discount coefficient for neutral discussion content.

[0023] In another preferred implementation manner, the method further includes: (1) Based on keywords related to a specific public opinion event, obtain public opinion platforms highly relevant to this specific public opinion event;

[0024] (2) Calculate the weighted coefficients of the occurrence frequencies of the core keywords of this specific public opinion event in each public opinion platform, and sort the weighted coefficients of the core keywords in each public opinion platform;

[0025] (3) Select the target analysis platform according to the sorting of each public opinion platform.

[0026] In another preferred implementation manner, for each account, in step (3.7), the directional angle value uses φ j = P j + σB j - Nj Perform calculations, where P j is the assignment value based on positive discussion content in a single speech of the account, and N j is the assignment value based on negative discussion content in a single speech of the account, and B j is also the assignment value based on negative discussion content in a single speech of the account. σ represents the discount coefficient for neutral discussion content. Use this formula to calculate the directional angle values of multiple speeches of the account, and then calculate the directional angle change curve of the account.

[0027] The method also includes counting the proportion of accounts with a similarity of the directional angle change curve higher than a set threshold in the total discussion accounts. When this proportion is higher than 3 - 5%, an early warning is issued.

[0028] In the present invention, although three directions are given for the directionality of speeches, those skilled in the art can further expand the directionality and divide it into more directional classifications.

[0029] It should be noted that the assignment values based on positive, negative, and neutral discussion content in steps (3.6) and (3.7) are obtained by summing the assignment values of keywords in the speech. Taking P j as an example, it is equal to where K is the number of keywords included in the speech, and p k is the assignment value of the k-th keyword, and so on. The negative and neutral directionality is calculated in this way.

[0030] By using the method of the present invention, it is possible to effectively determine the object of public opinion discussion and the directional change of the derivative object discussion, and timely determine the guiding speeches and corresponding accounts based on the direction of public opinion discussion. Furthermore, it is possible to more accurately determine whether there is a guiding group controlling and guiding the direction of the discussion, reduce the amount of data required for analysis, and improve the analysis accuracy. Brief Description of the Drawings

[0031] Figure 1 is a schematic flowchart of the method of the present invention.

[0032] Figure 2 is a screenshot of some speeches before the inflection point of public opinion in Event A;

[0033] Figure 3 is a screenshot of some speeches after the inflection point of public opinion in Event A.

[0034] Figure 4 is an exemplary architecture diagram of the LSTM model;

[0035] Figure 5 is a curve graph of the public opinion direction of the intrinsic object of the example event in the present invention. Detailed Embodiments

[0036] The present invention will be described in detail below in conjunction with the accompanying drawings and their embodiments, but the protection scope of the present invention is not limited to the scope described in the embodiments.

[0037] The method of the present invention includes the following steps:

[0038] (1) Based on keywords related to a specific public opinion event, obtain the public opinion platforms highly relevant to the specific public opinion event;

[0039] (2) Calculate the weighted coefficients of the occurrence frequencies of the core keywords of the specific public opinion event in each public opinion platform, and sort the weighted coefficients of the core keywords in each public opinion platform;

[0040] (3) Select the target analysis platform according to the sorting of each public opinion platform.

[0041] The steps for analyzing the target platform include:

[0042] (3.1) Determine the key sections related to the specific public opinion event, capture the discussion content of the relevant topics in any one of the key sections, and establish a mapping relationship pair between the account information of the public opinion expression subject, the public opinion expression content, and the public opinion expression time within the corresponding section;

[0043] (3.2) Extract the discussion content within each topic, extract the keywords in the discussion content of each post, establish an event corpus, and use the Word2vec model to transform the obtained corpus, and represent the words in the corpus in the form of one-hot encoding;

[0044] (3.3) Extract the words expressing the intrinsic object of the public opinion and the words expressing the derivative object of the public opinion in the corpus, and construct an intrinsic object corpus of the public opinion and a derivative object corpus of the public opinion;

[0045] (3.4) Construct an LSTM model, select some discussion content for marking, and mark them as the intrinsic object of the public opinion and several derivative objects of the public opinion (the number of derivative objects can be determined according to the specific situation of the derivative objects to be studied. Train the LSTM model.

[0046] The structure of an exemplary LSTM model of the present invention is as Figure 2 shown. In the figure, t represents the current moment, W f represents the forgetting gate, W i represents the input gate, W o represents the output gate, tanh and σ represent non-linear activation functions, c represents the long-term memory state at time t in the network structure, and h represents the short-term memory state and output at the moment.

[0047] Those skilled in the art should understand that for the classification of the object and the derivative object, other classification models can also be used to achieve it, which will not be limited here.

[0048] Taking Event A as an example, the intrinsic object event of public opinion is Event A itself, and the derivative objects of public opinion can be various derivative discussion targets, etc. For the sake of simplicity of description, here an example with two derivative objects of public opinion will be described.

[0049] (3.5) Use the LSTM model to classify each piece of discussion content, and classify it into discussions on the intrinsic object of public opinion, discussions on the primary derivative object (in this example, discussions on the first derivative object of public opinion), and discussions on the secondary derivative object (discussions on the second derivative object of public opinion).

[0050] (3.6) Determine the total discussion time of the public opinion event, divide the time period based on the total discussion time, and divide all the discussion content according to multiple time nodes (for example, every 5 - 10 minutes) into several public opinion stages. For each public opinion stage, respectively perform weighted summation on the public opinion discussion object directions within this stage to obtain the public opinion direction vector of this stage. In the public opinion direction vector, for each public opinion subject, take it as a dimension and assign a dimension modulus value. The dimension modulus value reflects the frequency of discussion of this subject. For the directionality of the expression statements of the discussion on this object, based on the thesaurus of directional expressions, assign values to it. Based on (the difference between positive discussions and negative discussions divided by the total discussions as the discussion direction angle of this object), thus construct the public opinion direction curve for each public opinion object, and determine the target public opinion object according to the order of the dimension modulus values of the public opinion objects (from large to small).

[0051] In a preferred implementation manner, the discussion direction angle value of the object is calculated based on the following formula: n represents the number of subjects participating in the discussion of this object, i ∈ [1, n], P i is the assignment based on positive discussion content, N i is the assignment based on negative discussion content. It should be noted that for the attenuation coefficient of neutral discussions, it can be set to 0. Each assignment of positive, negative, and neutral is the sum of assignments based on keywords. The assignment of each positive discussion is determined based on the number of positive keywords and the assignment weight of each positive keyword. Assign assignment weights to each positive keyword in the corpus, and for words not appearing in the corpus, assign values to them based on the assignment of the keyword closest to them.

[0052] In the data on the change of public opinion direction or the direction curve, the area passing through the zero point of the vertical axis is regarded as the inflection point area, and the accounts participating in the discussion of this section and the content of the remarks within a predetermined time range before and after the inflection point are intercepted. If the number of participating accounts or the number of discussion posts within this time range is too large and exceeds a certain number, the inflection point range is gradually reduced until the number of discussion posts near the public opinion inflection point can be controlled below the predetermined number (such as 5,000 posts).

[0053] Taking Event A as an example, all the remarks during the event discussion are divided into 15 time series (public opinion stages). Taking the discussion of the eigen-object as an example, the normalized discussion direction angle curve for the eigen-object (Event A itself) is as Figure 5 shown.

[0054] As can be seen from the figure, the early discussions are mainly positive discussions. However, as time goes on, starting from time series 4, the discussion turns, and by the time between time series 6 - 8, the discussion has turned from positive to negative. Therefore, intercepting the remarks within the inflection point range for research is more likely to find guiding remarks.

[0055] (3.7) For the content of the remarks within the inflection point area, extract all the remarks of the same account regarding the relevant object within the inflection point range in chronological order or sentence order. For each content of the remarks published, determine the directional angle value of the remarks published content, and construct the directional angle change curve of this account with time as the horizontal axis,

[0056] The directional angle value uses φ j = P j + σB j - N j for calculation, and it can be normalized within the interval [0, 1] or [0, 2π]. Among them, P j is the assignment value based on the positive discussion content in a single remark of this account, N j is the assignment value based on the negative discussion content in a single remark of this account, B j is the assignment value based on the negative discussion content in a single remark of this account, and σ represents the discount coefficient for neutral discussion content. The assignment values for positive, negative, and neutral discussion content are obtained by summing the assignment values of the keywords in the remarks. Taking P j as an example, it is equal to K is the number of keywords contained in the remark, p k is the assignment value of the k-th keyword, and so on. The negative and neutral directions are calculated in this way. Using such a calculation method can calculate the directionality of the remarks more clearly. For example, assign 1 to the language with obvious positive energy, 0.3 to the neutral language, and -1 to the negative energy language.

[0057] More preferably, for remarks with a single comment exceeding a certain length, the remarks are split, and the directional angle value and the angle curve are calculated based on the split remarks. At this time, the abscissa of the curve is the number of split remarks.

[0058] The present invention uses the change of the directional angle value to extract guiding remarks, which has obvious advantages compared with simply extracting guiding remarks based on the directionality itself. If judged solely by the directionality of the remarks, a large number of inappropriate remarks will be extracted. Such remarks are often issued by a specific group, and their remarks will not cause people to follow, and the harm of such remarks is relatively small. Guiding remarks, on the other hand, are remarks with a certain clear logic and context, and such remarks are the ones with social harm.

[0059] (3.8) Crawl the discussion content of relevant topics in other key sections of the current public opinion platform regarding this public opinion event, and repeat the above steps (3.1)-(3.8) to construct the directional angle change curves of all relevant accounts within the corresponding section.

[0060] (3.9) Perform similarity matching on the directional angle change curves of all obtained accounts, and extract the account information with a similarity of the directional angle change curve higher than the set threshold.

[0061] In this way, accounts that have obvious semantic differences but similar directionality can be found. In some cases, the remarks of two accounts seem to have very low similarity in terms of expression, but have similar directional variability. Such guiding remarks need to be closely monitored and are worthy of early warning. Existing methods are basically studied according to semantic similarity and do not pay attention to the guiding role of the directionality of remarks.

[0062] Preferably, select the total difference between the individual with the largest angle value and the individual with the smallest angle value. For each speech subject (the angle values of the account's remarks in each publication are normalized with the radian of a circle as the basic unit), project the obtained result into the angular coordinate system. Different quadrants in the angular coordinate system correspond to different directional intensities. Eliminate the subjects in the first and fourth quadrants to filter out extreme remarks.

[0063] As Figure 2 shown, it is the information content extracted from some online discussion content before the public opinion inflection point occurred when event A happened. From Figure 2 it can be seen that in the initial stage of the event, people's focus was mainly on event A itself, and the object of discussion was more about event A itself, and the direction of discussion was also a positive direction and a relatively neutral direction.

[0064] Figure 3The figure shows the information content extracted from some online discussions after the inflection point of public opinion occurred when Event A happened.

[0065] When studying the remarks near the inflection point by switching the research subject to the derivative object, it is found that when the inflection point of the speech direction occurs, the discussion of the object also changes. Figure 3 After the midpoint inflection, the object of public opinion discussion gradually changes to the derivative object, and the direction of discussion also changes. From the comparison of the remarks in the attached figure, it can be seen that the direction of public opinion discussion has changed from the most basic Event A itself to the discussion of other derivative objects. Such gradual remarks are likely to cause changes in the direction of discussion and are more likely to trigger corresponding consequences. The method of the present invention can effectively monitor such changes in public opinion objects and the relevance of corresponding discussion accounts. Therefore, the method of the present invention can determine the change of people's emotions for each object one by one. By studying the people with similar thinking evolution processes, the process of people's thinking evolution can be better determined, the occurrence of guiding remarks can be better found, and early warnings can be provided.

[0066] Although the principle of the present invention has been described in detail above in conjunction with the preferred embodiments of the present invention, those skilled in the art should understand that the above embodiments are only explanations of the illustrative implementation modes of the present invention and do not limit the scope of the present invention. The details in the embodiments do not constitute a limitation on the scope of the present invention. Without departing from the spirit and scope of the present invention, any obvious changes such as equivalent transformation and simple substitution based on the technical solution of the present invention fall within the protection scope of the present invention.

Claims

1. A tracking and tracing method based on network information, characterized in that: The method comprises: (3.1) Identify key sections related to specific public opinion events, capture discussion content of relevant topics in any key section, and obtain the corresponding relationship between the account information of the subject of public opinion expression and the content and time of public opinion expression in the corresponding section; (3.2) Extract the discussion content within each topic, extract the keywords in the discussion content of each post, and build an event corpus; (3.3) Extract the words that express the intrinsic object of public opinion and the words that express the derived object of public opinion in the corpus, and construct the corpus sub-database of the intrinsic object of public opinion and the corpus sub-database of the derived object of public opinion; (3.4) Construct an LSTM model, select discussion content about public opinion objects and mark them as public opinion intrinsic objects and several public opinion derived objects, and use the marked text to train the LSTM model; (3.5) Use the LSTM model to classify each discussion content into discussions on the intrinsic objects of public opinion and discussions on the derived objects of public opinion; (3.6) Determine the total length of discussion time of the public opinion event, divide the time periods based on the total length of discussion time, divide all discussion contents according to multiple time nodes, and divide them into several public opinion stages. For each public opinion stage, perform weighted summation on the directions of the public opinion discussion objects in the stage to obtain the public opinion direction vector of the stage. In the public opinion direction vector, for each public opinion object, take it as a dimension and assign a dimension modulus value. The dimension modulus value is proportional to the frequency of the object being discussed. For the directionality of the statement of the discussion of the object, assign it based on the lexicon of directional expressions. Take the ratio of the difference between positive discussion and negative discussion to the total discussion as the discussion direction angle of the object. Determine the target public opinion object according to the order of the dimension modulus value of the public opinion object. For the target public opinion object, the time when the public opinion direction passes through zero is taken as the turning point, and the accounts and speech content participating in the discussion of this section within a predetermined time range before and after the turning point are intercepted; (3.7) For the speech content within the inflection point area, extract all speech content about the target public opinion object in the inflection point range of the same account in chronological order. For each speech content, determine the directional angle value of the speech content, and construct the directional angle change curve of the account with time or number of sentences as the horizontal axis; (3.8) Capture the discussion content of the current public opinion platform on the public opinion event and related topics in other key sections, repeat the above steps (3.6)-(3.8), and construct the directional angle change curve of all inflection point related accounts in the corresponding section; (3.9) Perform similarity matching on the directional angle change curves of all the accounts obtained, and extract account information whose directional angle change curve similarity is higher than a set threshold.

2. The tracking and tracing method based on network information according to claim 1 is characterized in that: The step (3.2) includes: using the Word2vec model to convert the obtained corpus, and vectorizing the words in the corpus in a one-hot encoding form.

3. The tracking and tracing method based on network information according to claim 1 is characterized in that: The discussion directions of the object include positive, negative, and neutral, and different scores are assigned to different corpora.

4. The tracking and tracing method based on network information according to claim 3 is characterized in that: In step (3.6), the angle value of the discussion direction of the object is calculated based on the following formula: n represents the number of subjects participating in the discussion of the object, i∈[1,n], P i is the assignment based on the positive discussion content, N i is the assignment based on negative discussion content, B i is the value assigned based on neutral discussion content, and σ represents the discount coefficient for neutral discussion content.

5. The tracking and tracing method based on network information according to claim 1 is characterized in that: The method further includes: (1) based on keywords related to a specific public opinion event, obtaining a public opinion platform that is highly related to the specific public opinion event; (2) Calculate the weighted coefficients of the frequency of occurrence of the core keywords of the specific public opinion event on each public opinion platform, and rank the weighted coefficients of the core keywords on each public opinion platform; (3) Select the target analysis platform according to the ranking of various public opinion platforms.

6. The tracking and tracing method based on network information according to claim 3 is characterized in that: For each account, the directional angle value in step (3.7) is calculated using φ j =P j +σB j -N j Calculate, where P j N is the value assigned to the account based on the positive discussion content in a single speech. j B is the value assigned to the negative discussion content in a single speech of the account, j is the value assigned based on the neutral discussion content in a single speech of the account, σ represents the discount coefficient for the neutral discussion content, and the formula is used to calculate the directional angle values ​​of multiple speeches of the account, and then calculate the directional angle change curve of the account.

7. The tracking and tracing method based on network information according to claim 1 is characterized in that: The method also includes calculating the proportion of accounts whose similarity of directional angle change curves is higher than a set threshold to the total discussion accounts, and issuing an early warning when the proportion is higher than 3-5%.

Images