A system for detecting all-media content security based on broadcasting and television technology

Abstract

The present invention relates to the technical field of content security detection, and specifically discloses a system for realizing full-media content security detection based on broadcasting and television technology, comprising: a model building module: setting a number of sentiment polarity labels, marking the sentiment polarity labels of historical texts, and forming sample data; using a convolutional neural network model, training according to the sample data, and obtaining a sentiment label analysis model; an initial module: presetting initial keywords, obtaining original texts, using the sentiment label analysis model to perform sentiment polarity analysis on these original texts, and obtaining the sentiment polarity labels of the original texts; a keyword analysis module: setting an analysis cycle, obtaining context-like texts, counting the proportion of occurrences and the proportion of other times, obtaining keyword quantity curves and other quantity curves, and judging whether the other quantity curves meet complementary constraints; a detection module: using the initial keywords and alternative keywords to perform content security detection on subsequent media texts.

Description

Technical Field

[0001] The present invention relates to the technical field of content security detection, and in particular to a system for realizing full-media content security detection based on broadcasting and television technology. Background Art

[0002] All-media content security detection mainly involves the use of advanced information technology to conduct all-round, multi-level security monitoring and management of radio, television and new media content; it mainly includes image review, video review, text review, real-time monitoring, security protection and data analysis.

[0003] In the prior art, when performing content security detection on the text content in the media file, the method mainly relies on presetting illegal keywords. This method searches for these illegal keywords in the text content and shields the retrieved illegal keywords to achieve the purpose of maintaining content security. However, with the improvement of the audience's or users' thinking and creativity, some audiences or users are gradually able to cleverly avoid the retrieval of these preset illegal keywords. They will use homophones, homophones, pinyin abbreviations, decomposition, similar characters, etc. to replace the original illegal keywords, thereby circumventing system detection. In addition, the speed of change of Internet buzzwords is very fast, and new Internet terms and expressions are constantly emerging. Simply relying on the preset illegal keyword list cannot keep up with this change in time, resulting in the latest illegal keywords often being difficult to retrieve in time. Summary of the invention

[0004] The purpose of the present invention is to provide a system for realizing security detection of all media contents based on broadcasting and television technology, so as to solve the above technical problems.

[0005] The purpose of the present invention can be achieved through the following technical solutions:

[0006] A system for detecting the security of all media contents based on broadcasting and television technology, comprising:

[0007] Model building module: set a number of sentiment polarity labels, obtain historical media files, extract the text content in the historical media files, record them as historical texts, mark the sentiment polarity labels of each historical text, and obtain sample data; establish a convolutional neural network model, train the convolutional neural network model according to the sample data, and obtain the final sentiment label analysis model;

[0008] Initial module: preset a number of initial keywords, obtain the historical text where the initial keywords are located, and record it as the original text; obtain the sentiment polarity label of the original text through the sentiment label analysis model, and record it as the original sentiment label;

[0009] Keyword analysis module: set an analysis period; obtain all media files within the analysis period in real time, and extract the text content in the media files, which is recorded as media text; select the media text whose sentiment polarity label is the original sentiment label, which is recorded as context-like text; obtain all words in the context-like text, and obtain the occurrence ratio of the initial keyword in all words; record the words in all words except the initial keyword as the remaining words, obtain the occurrence ratio of each remaining word, and record it as the remaining frequency ratio;

[0010] According to the proportion of the number of occurrences in each analysis period, the keyword quantity curve is fitted, and according to the proportion of the remaining times of each remaining word in each analysis period, the remaining quantity curves of each remaining word are obtained; if the keyword quantity curve G(t) , where t is the number of the analysis period, t i represents the i-th analysis cycle, then it is determined in turn whether the keyword quantity curve and the other quantity curves satisfy the complementary constraint: , where n is the number of analysis cycles, t n represents the nth analysis period, Q(t) is the remaining quantity curve, G´(t i ) means that G(t) is at t=t i The derivative value at Q´(t i ) means Q(t) at t=t i The derivative value at , Cv is a preset error threshold, k is a preset slope error threshold; if the keyword quantity curve and the remaining quantity curves Q(t) satisfy the complementary constraint, the remaining words corresponding to the remaining quantity curves Q(t) are recorded as replacement keywords for the initial keywords;

[0011] Detection module: performs content security detection on subsequent media texts based on the initial keywords and alternative keywords.

[0012] As a further solution of the present invention: the setting of the sentiment polarity label includes positive, negative and neutral.

[0013] As a further solution of the present invention: the establishment of the convolutional neural network model is based on natural language processing technology.

[0014] As a further solution of the present invention: the setting process of the analysis cycle includes:

[0015] A time interval threshold is set, and the moment of setting the initial keyword is set as the initial moment. Starting from the initial moment, a time node is obtained every time interval threshold, and each two adjacent time nodes constitute an analysis cycle, and the composition of the analysis cycle includes the initial moment.

[0016] As a further solution of the present invention: the process of obtaining the occurrence frequency ratio includes:

[0017] Obtain the total number M of all words in the class context text, and obtain the occurrence times m of the initial keyword among all words, then the occurrence frequency ratio p = m / M is obtained.

[0018] As a further solution of the present invention: the fitting process of the keyword quantity curve includes:

[0019] Number each analysis period, and obtain the occurrence frequency ratio under each numbered analysis period; use the number as the abscissa and the occurrence frequency ratio as the ordinate to establish a coordinate system; according to each numbered analysis period and its corresponding occurrence frequency ratio, obtain several coordinate points in the coordinate system; connect each coordinate point with a smooth curve, and record the curve as the keyword quantity curve.

[0020] As a further solution of the present invention: when obtaining all words in the class context text, filter out the stop words in the class context text, and the stop words include "de" and "wo".

[0021] As a further solution of the present invention: the process of performing content security detection on subsequent media texts according to the initial keyword and the alternative keyword includes:

[0022] Retrieve the words in the media text. If the initial keyword or the alternative keyword is retrieved, it is prompted that the content of the media text is abnormal.

[0023] Advantages of the present invention:

[0024] The present invention sets sentiment polarity labels and uses a convolutional neural network model to train historical texts, so that the system can learn and identify texts with different sentiment polarities. This deep learning-based method is more accurate and flexible than traditional rule-based or statistical methods because it can capture complex structures and patterns in texts. By setting the analysis cycle, media files in each cycle can be acquired and analyzed in real time. This real-time performance enables the system to quickly respond to new content and conduct security detection in a timely manner. At the same time, by continuously updating and optimizing the keyword library (including initial keywords and alternative keywords), it can adapt to the dynamic changes of media content and maintain high accuracy. And by analyzing all the words in the context-like text and their proportion of occurrence, possible alternative keywords are discovered. This method ensures comprehensive coverage and detailed analysis of media content and reduces the risk of missing important information. According to the current network language situation, a number of initial keywords are set, and then the proportion of the number of occurrences of the initial keywords in each analysis cycle is collected. If the proportion of the number of occurrences of the initial keywords begins to decrease, it is analyzed whether there are other words in each analysis cycle whose proportion of the remaining times increases. If the complementary constraints are met, it indicates that the remaining words gradually replace the initial keywords and become new keywords. When performing security detection on the text content of media files, the setting of keywords follows the latest changes in network vocabulary. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The present invention will be further described below in conjunction with the accompanying drawings.

[0026] Figure 1 It is a flow chart of a system for realizing a security detection of all-media content based on broadcasting and television technology according to the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] See also Figure 1 As shown, the present invention is a system for realizing full media content security detection based on broadcasting and television technology, comprising:

[0029] Model building module: set a number of sentiment polarity labels, obtain historical media files, extract the text content in the historical media files, record them as historical texts, mark the sentiment polarity labels of each historical text, and obtain sample data; establish a convolutional neural network model, train the convolutional neural network model according to the sample data, and obtain the final sentiment label analysis model;

[0030] It is understandable that a large number of historical media files are collected, such as news reports, social media posts, comments, etc.; text content is extracted from these files as the basis of training data; the extracted text content is manually or automatically annotated and assigned corresponding sentiment polarity labels. This step is the key to building a supervised learning model, because the model needs to learn how to identify the sentiment tendency of the text through annotated data; the annotated historical texts and their corresponding sentiment polarity labels are combined to form a sample data set, which will be used for subsequent model training; the convolutional neural network model is trained using the obtained sample data set; during the training process, the model will learn how to adjust its internal parameters to minimize the difference between the predicted sentiment polarity label and the actual label. Through multiple iterations and optimization, the model gradually improves its ability to recognize the sentiment tendency of the text;

[0031] As a preferred embodiment of the present invention, the setting of the sentiment polarity label includes positive, negative and neutral;

[0032] It should be noted that sentiment polarity refers to the positive, negative or neutral sentiment tendency expressed in a text; the sentiment polarity tag is used to annotate the sentiment polarity of a text; sentiment polarity analysis is a research field in natural language processing (NLP), which focuses on extracting and identifying the author's emotional attitude from text data;

[0033] As a preferred embodiment of the present invention, the establishment of the convolutional neural network model is based on natural language processing technology;

[0034] It is worth noting that before building a convolutional neural network model, the sample data needs to be fully preprocessed, including text cleaning, word segmentation, and stop word removal;

[0035] It should be noted that the architecture of the convolutional neural network is designed according to the requirements, including convolutional layers, pooling layers and fully connected layers; the convolutional layer is used to extract text features, the pooling layer is used for dimensionality reduction, and the fully connected layer maps the features to the output space to achieve classification or regression tasks; during the model training process, the input sample is passed layer by layer through the convolutional layer, activation layer (such as ReLU), pooling layer, etc., and finally the output result is obtained. This process is called forward propagation; during the forward propagation process, each neuron receives the input from the neurons in the previous layer, and calculates according to its weight and bias, and then passes the result to the neurons in the next layer; after the forward propagation, it is necessary to calculate the difference between the model prediction result and the true label, that is, Loss; Common loss functions include cross entropy loss and mean square error; According to the definition of loss function, the loss of each sample can be calculated, and then the average loss of the entire training set can be calculated; After calculating the loss, the backpropagation algorithm needs to be used to update the parameters of the model to reduce the loss value; The backpropagation is to transfer the loss from the output layer to the input layer layer by layer through the chain rule, and update the value of each parameter according to the gradient descent algorithm; After completing a forward propagation and backpropagation, a new set of parameter values ​​will be obtained; In the process of multiple iterations, the model will continuously optimize its parameters to gradually reduce the loss value and improve the prediction accuracy;

[0036] Initial module: preset a number of initial keywords, obtain the historical text where the initial keywords are located, and record it as the original text; obtain the sentiment polarity label of the original text through the sentiment label analysis model, and record it as the original sentiment label;

[0037] It can be understood that the acquired original text is input into the previously trained sentiment label analysis model, which processes the text content and outputs one or more sentiment polarity labels; the sentiment polarity labels of the original text obtained from the sentiment label analysis model are recorded as a reference or benchmark for subsequent analysis; these original sentiment labels reflect the emotional color of the text in the context of specific keywords;

[0038] Keyword analysis module: set an analysis period; obtain all media files within the analysis period in real time, and extract the text content in the media files, which is recorded as media text; select the media text whose sentiment polarity label is the original sentiment label, which is recorded as context-like text; obtain all words in the context-like text, and obtain the occurrence ratio of the initial keyword in all words; record the words in all words except the initial keyword as the remaining words, obtain the occurrence ratio of each remaining word, and record it as the remaining frequency ratio;

[0039] According to the proportion of the number of occurrences in each analysis period, the keyword quantity curve is fitted, and according to the proportion of the remaining times of each remaining word in each analysis period, the remaining quantity curves of each remaining word are obtained; if the keyword quantity curve G(t) , where t is the number of the analysis period, t i represents the i-th analysis cycle, then it is determined in turn whether the keyword quantity curve and the other quantity curves satisfy the complementary constraint: , where n is the number of analysis cycles, t n represents the nth analysis period, Q(t) is the remaining quantity curve, G´(t i ) means that G(t) is at t=t i The derivative value at Q´(t i ) means Q(t) at t=t i The derivative value at , Cv is a preset error threshold, k is a preset slope error threshold; if the keyword quantity curve and the remaining quantity curves Q(t) satisfy the complementary constraint, the remaining words corresponding to the remaining quantity curves Q(t) are recorded as replacement keywords for the initial keywords;

[0040] It is understandable that a fixed time period needs to be set as the analysis cycle, which can be adjusted according to actual needs, such as daily, weekly or monthly; the purpose of setting the analysis cycle is to regularly collect and analyze media files so as to timely discover content change trends and potential risks; in each analysis cycle, the system automatically obtains all media files in the cycle; these media files may include news reports, social media posts, blog posts and other forms of content; by obtaining these files in real time, the system can ensure that the analyzed data is up to date, thereby improving the accuracy and timeliness of the analysis; according to the sentiment polarity labels previously obtained through the sentiment label analysis model, media texts matching the original sentiment labels are screened out; these texts are called context-like texts, which reflect the text content in the same sentiment background as the initial keywords; further analysis is performed on the context-like texts to extract all the words therein, and the proportion of the number of occurrences of each word to the total number of words is calculated, which helps to understand which words are more common in the same sentiment background as the initial keywords; the words in the context-like texts are divided into two categories: initial keywords and remaining words; the initial keywords refer to those preset keywords, and the remaining words are all words except the initial keywords, and their proportion of occurrences is calculated respectively;

[0041] It is worth noting that in order to determine whether there are other words that can replace the initial keyword, the system needs to check whether there is a complementary relationship between the keyword quantity curve and the remaining quantity curves; specifically, the following two conditions need to be met:

[0042] Integration threshold condition: Cv is a preset error threshold used to measure the difference between the two curves. If the difference is less than or equal to Cv, the two curves are considered to be complementary in terms of the value of the integral area. The closer the absolute value of the difference between the integral values ​​of the two curves is to 0, the more complementary they are.

[0043] Slope error threshold condition: , where k is a preset slope error threshold, which is used to limit the sum of the change rates of the two curves; if the sum of the change rates of the two curves belongs to [0, k], it is considered that the change trends of the two curves are just opposite; if both of the above two conditions are met, it is considered that there is a complementary relationship between the keyword quantity curve and other quantity curves, that is, there are potential alternative keywords;

[0044] As a preferred embodiment of the present invention, the setting process of the analysis cycle includes:

[0045] Setting a time interval threshold, taking the time when the initial keyword is set as the initial time, starting from the initial time, obtaining a time node every time interval threshold, and each two adjacent time nodes constitute an analysis cycle, and the composition of the analysis cycle includes the initial time;

[0046] It is understandable that a time interval threshold needs to be determined; this threshold defines the length of time between two adjacent time nodes; for example, if the time interval threshold is set to 24 hours, then a new time node will be generated every 24 hours; from the initial moment, according to the previously set time interval threshold, the time is gradually increased to generate a new time node; for example, if the initial moment is 8 o'clock on Monday morning and the time interval threshold is 24 hours, then the next time node will be 8 o'clock on Tuesday morning, and so on; each analysis cycle consists of two adjacent time nodes; for example, the first analysis cycle may be from 8 o'clock on Monday morning to 8 o'clock on Tuesday morning, the second analysis cycle is from 8 o'clock on Tuesday morning to 8 o'clock on Wednesday morning, and so on;

[0047] As a preferred embodiment of the present invention, the process of obtaining the occurrence frequency ratio includes:

[0048] Obtain the total number M of all words in the class context text, and obtain the number of occurrences m of the initial keyword in all words, and then obtain the number of occurrences p=m / M;

[0049] It can be understood that all the words in the text can be obtained by extracting them from the context-like text through natural language processing (NLP) techniques, such as word segmentation, removal of stop words, etc. Then, the total number of these words M is counted;

[0050] In a preferred embodiment of the present invention, the fitting process of the keyword quantity curve includes:

[0051] Number each analysis period, and obtain the proportion of the occurrence times in each analysis period with the corresponding number; use the number as the abscissa and the proportion of the occurrence times as the ordinate to establish a coordinate system; according to each analysis period with the corresponding number and the proportion of the occurrence times, obtain a number of coordinate points in the coordinate system; connect each coordinate point with a smooth curve, and record the curve as the keyword quantity curve;

[0052] It can be understood that a mathematical method (such as the least squares method, spline interpolation, etc.) is used to connect the obtained coordinate points with a smooth curve; this curve is the keyword quantity curve, which reflects the change trend of the initial keywords in different analysis periods;

[0053] In a preferred embodiment of the present invention, when obtaining all the words in the class context text, filter out the stop words in the class context text, and the stop words include "de" and "wo";

[0054] It should be noted that the stop words refer to the words with a relatively high frequency of occurrence in the text but a relatively small contribution to the analysis task; these words usually include common words without actual meaning such as "de" and "wo", and removing them can improve the accuracy and efficiency of the analysis; the stop word filtering technology identifies and removes irrelevant words by using a predefined stop word list; common stop word libraries include the Harbin Institute of Technology stop word list, the Baidu stop word list, etc., and these tools can effectively reduce the influence of noise data; the filtering of the stop words first loads the text data and performs tokenization processing to split the text into words; then obtains the stop word list and filters out the stop words in the text, and finally counts the frequency of the remaining words;

[0055] Detection module: Perform content security detection on subsequent media texts according to the initial keywords and alternative keywords;

[0056] In a preferred embodiment of the present invention, the process of performing content security detection on subsequent media texts according to the initial keywords and alternative keywords includes:

[0057] Retrieve the words in the media text. If the initial keyword or alternative keyword is retrieved, it is prompted that the content of the media text is abnormal;

[0058] It is understandable that the system will perform natural language processing (NLP) on the input media text, including steps such as word segmentation and removal of stop words, so as to extract all the words in the text; then, the system will search these words to check whether they match the initial keywords or alternative keywords; if the system finds the initial keywords or their alternative keywords during the retrieval process, this means that the media text may contain sensitive or inappropriate content, and a warning or prompt will be generated to inform the user that there is an abnormality in the content of the media text.

[0059] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A system for realizing all-media content security detection based on broadcasting and television technology, characterized in that: Including: Model establishment module: Set a number of sentiment polarity labels, obtain historical media files, extract the text content in the historical media files, denoted as historical text, label the sentiment polarity labels of each historical text to obtain sample data; establish a convolutional neural network model, and train the convolutional neural network model according to the sample data to obtain a final sentiment label analysis model. Initial module: Preset a number of initial keywords, obtain the historical text where the initial keywords are located, denoted as the original text; through the sentiment label analysis model, obtain the sentiment polarity label of the original text, denoted as the original sentiment label. Keyword analysis module: Set an analysis period; obtain all media files within the analysis period in real time, and extract the text content in the media files, denoted as media text. Select the media text with the sentiment polarity label being the original sentiment label, denoted as context-like text; obtain all the words in the context-like text, and obtain the proportion of the occurrence times of the initial keywords in all the words. Denote the words other than the initial keywords in all the words as other words, and obtain the proportion of the occurrence times of each other word, denoted as the other occurrence proportion. According to the proportion of the number of occurrences in each analysis period, the keyword quantity curve is fitted, and according to the proportion of the remaining times of each remaining word in each analysis period, the remaining quantity curves of each remaining word are obtained; if the keyword quantity curve G(t) , where t is the number of the analysis period, t i represents the i-th analysis cycle, then it is determined in turn whether the keyword quantity curve and the other quantity curves satisfy the complementary constraint: , where n is the number of analysis cycles, t n represents the nth analysis period, Q(t) is the remaining quantity curve, G´(t i ) means that G(t) is at t=t i The derivative value at Q´(t i ) means Q(t) at t=t i The derivative value at , Cv is a preset error threshold, k is a preset slope error threshold; if the keyword quantity curve and the remaining quantity curves Q(t) satisfy the complementary constraint, the remaining words corresponding to the remaining quantity curves Q(t) are recorded as replacement keywords for the initial keywords; Detection module: Perform content security detection on subsequent media texts according to the initial keywords and alternative keywords.

2. According to claim 1, a system for realizing all-media content security detection based on broadcasting and television technology is characterized in that: The setting of the sentiment polarity labels includes positive, negative, and neutral.

3. According to claim 1, a system for realizing all-media content security detection based on broadcasting and television technology is characterized in that: The establishment of the convolutional neural network model is based on natural language processing technology.

4. According to claim 1, a system for realizing all-media content security detection based on broadcasting and television technology is characterized in that: The process of setting the analysis period includes: Set a time interval threshold, take the moment when the initial keywords are set as the initial moment, starting from the initial moment, obtain a time node every other time interval threshold, and each adjacent two time nodes form an analysis period, and the composition of the analysis period includes the initial moment.

5. According to claim 1, a system for realizing all-media content security detection based on broadcasting and television technology is characterized in that: The process of obtaining the proportion of occurrence times includes: Obtain the total number M of all the words in the context-like text, and obtain the occurrence times m of the initial keywords in all the words, then obtain the proportion of occurrence times p = m / M.

6. The system for realizing all-media content security detection based on broadcasting and television technology according to claim 1 is characterized in that: The process of fitting the keyword quantity curve includes: Number each analysis period, and obtain the proportion of occurrence times under each numbered analysis period; use the number as the abscissa and the proportion of occurrence times as the ordinate to establish a coordinate system; obtain a number of coordinate points in the coordinate system according to each numbered analysis period and its corresponding proportion of occurrence times; connect each coordinate point with a smooth curve, and denote the curve as the keyword quantity curve.

7. The system for realizing all-media content security detection based on broadcasting and television technology according to claim 1 is characterized in that: When obtaining all the words in the context-like text, filter out the stop words in the context-like text, and the stop words include "de" and "wo".

8. The system for realizing all-media content security detection based on broadcasting and television technology according to claim 1 is characterized in that: The process of performing content security detection on subsequent media texts according to the initial keywords and alternative keywords includes: Retrieve the words in the media text, if the initial keyword or alternative keyword is retrieved, it is prompted that there is an abnormality in the content of the media text.

Images