Content detection method for information board content security management system

Abstract

The application relates to the technical field of natural language processing, in particular to a content detection method for an intelligence board content security management and control system, which comprises the following steps: by setting different traffic field entity type weights, by adding corresponding weights when calculating the similarity of different keywords and entity type words, the gap between the similarity of different keywords and entity type words is increased, and then the semantic similarity between each keyword and the most matched entity type is determined; the segmented intelligence board content to be detected is dynamically evaluated, dynamic threshold setting is carried out according to the finally confirmed entity type, the semantic similarity of each keyword is compared with the corresponding dynamic threshold, the illegal words are determined, and the intelligence board content security monitoring is completed; compared with the case where the weight is not added, the gap between the similarity of illegal words and safe words can be increased, which is beneficial to the division of illegal words and safe words, and avoids the omission of high-risk words of the entity type with high aggregation.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, specifically to a content detection method for an information board content security management system. Background Technology

[0002] Information boards, as a key component of intelligent transportation systems, play an irreplaceable role in modern traffic management. By dynamically displaying real-time traffic information, they provide accurate guidance and warnings to road users, effectively improving traffic efficiency and safety. However, the accuracy and completeness of the content displayed on information boards directly affect their functionality. Content violations or formatting issues can not only mislead drivers' decisions but also potentially lead to serious traffic safety hazards. Therefore, establishing an effective information board content security control system to ensure the accurate presentation of information is a crucial step in guaranteeing the effectiveness of traffic management.

[0003] Currently, research on information boards mainly focuses on improving hardware reliability and optimizing communication protocols, while a systematic solution for automated detection of displayed content quality has not yet been developed. Traditional detection technologies still lack experience in the security management of information board content, making it difficult to accurately determine whether the output content is compliant and potentially leading to missed detections of prohibited words. Summary of the Invention

[0004] To address the aforementioned technical issues, this application provides a content detection method for an information board content security management system, thereby resolving the existing problems.

[0005] The content detection method for the information board content security management system in this application adopts the following technical solution:

[0006] One embodiment of this application provides a content detection method for an information board content security management system, the method comprising the following steps:

[0007] Collect information from the traffic information board content in the historical logs of the traffic management system, perform entity type recognition, and obtain a vocabulary set for each entity type; collect images of the traffic information boards to be detected.

[0008] Extract the text content from the traffic information board image and obtain the keywords of the text content;

[0009] The word vectors of each keyword and each word of each entity type are obtained through a word vector generation model; the entity type weights of each entity type are constructed based on the similarity between the word vectors of words in each entity type.

[0010] Based on the similarity between each keyword and the word vectors of words in each entity type, and combined with the entity type weights, the average similarity of each keyword in each entity type is constructed; based on the average similarity, the best matching entity type of each keyword and the semantic similarity between the keyword and the best matching entity type are determined.

[0011] Dynamic thresholds for various entity types are determined based on the distribution characteristics of the similarity between word vectors of words in each entity type; information board content security monitoring is performed based on the semantic similarity of each keyword and the dynamic threshold of the corresponding best matching entity type.

[0012] In one embodiment, the acquisition of keywords from the information board text content specifically includes:

[0013] The text content on the traffic information board is cleaned of non-text characters using regular expressions. The cleaned text content is then segmented and stop words are removed. All the words obtained after the process are used as keywords for the text content of the information board.

[0014] In one embodiment, the process of obtaining the entity type weights of each entity type is as follows:

[0015] Calculate the cosine similarity between the word vectors of any two words in each entity type, and take the mean of all the cosine similarities of each entity type as the average similarity between all words in each entity type.

[0016] The entity type weight of each entity type is determined based on the average similarity among all words in each entity type.

[0017] In one embodiment, the expression for the entity type weight of each entity type is:

[0018] In the formula, λ t μ represents the entity type weight of the t-th entity type. t μ' represents the average similarity among all words of the t-th entity type; μ' represents the set of average similarities among all entity types; min() is the minimum value function; ε is a pre-set positive number.

[0019] In one embodiment, the expression for the average similarity of each keyword in each entity type is:

[0020]

[0021] In the formula, b a,t V represents the average similarity of the a-th keyword in the information board text across the t-th entity type; t w represents the number of words for the t-th entity type. av is the word vector corresponding to the a-th keyword in the information board text; t,i λ is the word vector corresponding to the i-th word in the t-th entity type; t Let w be the entity type weight for the t-th entity type; cos() represents the calculation of the cosine similarity function, cos(w a ,v t,i ) represents the word vector w a With word vector v t,i Cosine similarity between them.

[0022] In one embodiment, the most matching entity type of each keyword and the semantic similarity between them are specifically defined as follows:

[0023] The entity type corresponding to the maximum average similarity of each keyword across all entity types is taken as the most matching entity type for each keyword; the average similarity of each keyword among the most matching entity types is taken as the semantic similarity between each keyword and the most matching entity type.

[0024] In one embodiment, the process of obtaining the dynamic thresholds for the various entity types is as follows:

[0025] Calculate the dispersion of cosine similarity among word vectors of all words within each entity type; determine the dynamic threshold of each entity type based on the average similarity among all words of each entity type, the standard deviation, and the entity type weight.

[0026] In one embodiment, the dispersion of the cosine similarity is the standard deviation of the cosine similarity between the word vectors of all words within each entity type.

[0027] In one embodiment, the expression for the dynamic threshold of each entity type is:

[0028] β t =(μ t -k×σ t )×λ t In the formula, β t μ represents the dynamic threshold for the t-th entity type; t σ represents the average similarity among all words of the t-th entity type; k represents the preset sensitivity coefficient; σ t λ represents the standard deviation of the cosine similarity among the word vectors of all words within the t-th entity type. t This represents the entity type weight for the t-th entity type.

[0029] In one embodiment, the security monitoring of the information board content specifically includes:

[0030] If the semantic similarity between each keyword and its most matching entity type is less than the dynamic threshold of that most matching entity type, then the keyword is determined to be a violation word; otherwise, it is determined to be a compliant word.

[0031] This application has at least the following beneficial effects:

[0032] This application proposes a content detection method for a traffic information board content security management system. By setting weights for different entity types in the traffic domain, it addresses the issue of missed detections when illegal and safe words are close in distance in the word vector space during traffic information board content security management. By adding corresponding weights when calculating the similarity between different keywords and words in entity types, the method increases the gap between the similarity between different keywords and words in entity types. Compared to not adding this weight, it increases the gap between the similarity between illegal and safe words, which is beneficial for the classification of illegal and safe words and avoids missing high-risk words in highly clustered entity types. Furthermore, it determines the semantic similarity between each keyword and the best matching entity type. The method dynamically evaluates the information board content to be detected after word segmentation, sets dynamic thresholds based on the finally confirmed entity types, compares the semantic similarity of each keyword with the corresponding dynamic threshold, identifies illegal words, completes the information board content security monitoring, and further avoids the problem of missed detection of illegal words. Attached Figure Description

[0033] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 A flowchart of the content detection method for the information board content security management system provided in this application;

[0035] Figure 2 A schematic diagram illustrating the process of obtaining keywords from the text content of the information board. Detailed Implementation

[0036] To further illustrate the technical means and effects adopted by this application to achieve the intended purpose of the invention, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the content detection method for an information board content security management system proposed in this application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0037] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0038] The following, in conjunction with the accompanying drawings, details the specific scheme of the content detection method for the information board content security management system provided in this application.

[0039] This application provides a content detection method for an information board content security management system according to one embodiment.

[0040] Specifically, the following content detection methods for the information board content security management system are provided. Please refer to [link / reference]. Figure 1 The method includes the following steps:

[0041] Step S1: Collect information from the information board content in the historical logs of the traffic management system, perform entity type identification, and obtain a vocabulary set for each entity type.

[0042] First, a knowledge database is constructed. Specifically, information related to the information board content is collected from the historical logs of the traffic management system. Named entity recognition (NAME) technology is then used to identify the entity types of the collected information, resulting in a vocabulary set for each entity type. The database composed of the vocabulary of all identified entity types serves as the knowledge database described in this application. The entity types in this application include: traffic events, road sections, traffic control measures, time attributes, and information board equipment. The NAME recognition technology is a well-known technology, and its specific process will not be elaborated upon here.

[0043] The traffic incidents mentioned above are nouns used to describe dynamic events that affect traffic, such as accidents, construction, congestion, and severe weather.

[0044] The content corresponding to the road segment is the specific interval name in the road network, such as "G2 Expressway Shanghai Section K12", etc.

[0045] The traffic control measures refer to specific management methods, such as speed limits, lane closures, and detour routes.

[0046] The content corresponding to the time attribute consists of terms related to the timeliness of traffic events, such as start time and estimated duration.

[0047] The information board device contains information board ID, location, screen type, etc.

[0048] Next, a high-definition camera is used to capture images of the traffic information boards to be detected. Since background information in the captured images may interfere with the recognition of the information board content, the captured images are processed to extract regions of interest (ROIs) to obtain the information board region images, ensuring that the text extracted subsequently comes from the traffic information boards. ROI extraction is a well-known technique, and the specific process will not be described in detail.

[0049] For the extracted information board area image, it is converted to grayscale to obtain a grayscale image. Then, the grayscale image is enhanced and denoised using the Limiting Contrast Adaptive Histogram Equalization (CLAHE) algorithm and the Non-Local Means Denoising (NLM) algorithm to improve the quality of the text area of ​​the traffic information board. Among them, the CLAHE algorithm and the NLM algorithm are well-known technologies, and the specific process will not be described in detail.

[0050] Step S2: Extract the text content from the traffic information board image and obtain the keywords of the text content.

[0051] (1) To obtain the text region in the information board, the grayscale values ​​of all pixels in the grayscale image are used as input to the Otsu method to obtain a segmentation threshold. Pixels with grayscale values ​​greater than the segmentation threshold are designated as foreground text region pixels, and pixels with grayscale values ​​less than the segmentation threshold are designated as background pixels. The grayscale values ​​of the foreground text region pixels are set to 255 (white foreground), and the grayscale values ​​of the background pixels are set to 0 (black background), thus obtaining a binary image. The Otsu method is a well-known technique, and its specific process will not be elaborated further. The binary image more clearly contrasts the foreground text region.

[0052] Furthermore, morphological operations are performed on the text foreground in the binary image. First, an opening operation is performed, which eliminates small noise points by erosion followed by dilation. Then, a closing operation is performed, which fills the internal holes of the text or repairs text breaks by dilation followed by erosion, thereby enhancing the binary image. Next, the Canny edge detection algorithm is used to detect edges in the enhanced binary image and extract clear text edge contours. The opening, closing, and Canny edge detection algorithms are all well-known techniques, and their specific processes will not be elaborated further.

[0053] Furthermore, Tesseract OCR is used to convert the text in the image with marked text outlines into editable text. Since the text may contain characters unrelated to the content, such as special symbols like

、

[0054] (2) In the text detection work of the information board security content control system, the standardized word segmentation and stop word filtering of the original text is the key link to ensure the accuracy of content security analysis.

[0055] For the cleaned text content, the jieba word segmentation tool was used to segment the text, and the segmentation results were processed to remove stop words. During word segmentation, a professional dictionary for the traffic field was loaded to ensure that industry terms such as "variable speed limit sign" and "traffic incident detection" could be fully recognized. The stop word library was specifically optimized for the characteristics of traffic information boards, including common function words while integrating high-frequency but meaningless words from the traffic field. Through a multi-level filtering mechanism of basic stop word filtering and domain-irrelevant word filtering, general meaningless words and high-frequency but non-controlling words in the traffic field, such as "prompt," "current," and "announcement," were filtered out, eliminating as many distractors as possible and retaining words with actual detection value. All text processing adopted the UTF-8 encoding standard to avoid garbled characters. The jieba word segmentation and stop word removal are well-known technologies, and their specific processes will not be elaborated further.

[0056] Taking a warning message on an information board, “A traffic accident has occurred 500 meters ahead, please slow down,” as an example, the keywords retained after processing are [“ahead,” “500 meters,” “occurred,” “traffic accident,” “slow down”]. These words will be input into subsequent content safety control calculations.

[0057] (3) All words obtained after the above word segmentation and stop word removal are used as keywords in the information board text content.

[0058] Step S3: Obtain word vectors for each keyword and each word of each entity type through a word vector generation model; construct entity type weights for each entity type based on the similarity between word vectors of words in each entity type.

[0059] In the content security management of traffic information boards, words that meet the content security standards of the information boards usually have a high degree of fit with one of the entity types in the knowledge database, while words that violate the rules have a low degree of fit with each entity type in the knowledge database. For example, the word "congestion," which is common in traffic management scenarios, has a relatively high degree of fit with the entity type "traffic incident" in the knowledge database; while the word "injury or fatality," which is completely unrelated to traffic management scenarios, has a low degree of fit with each entity type in the knowledge database of this application.

[0060] A pre-trained Word2Vec model is obtained from an online platform. All words from the knowledge database and all keywords from the information board text are used as input to this Word2Vec model to generate word vectors for each of these words. The Word2Vec model is a well-known technology, and its specific process will not be elaborated upon.

[0061] It should be noted that this application only provides one word vector generation method for obtaining word vectors. There are many existing word vector generation methods, and implementers may also use other word vector generation models to obtain word vectors. This application does not impose any specific restrictions.

[0062] Because the knowledge database constructed from historical intelligence board data exhibits both clustering and dispersion in the word similarity across entity types, for example, in the "time attribute" category, words representing time periods show high clustering, while in the "traffic event" category, words such as accident, construction, congestion, and severe weather are relatively dispersed. In entities with concentrated word vectors, some violating words may have similar word vectors to certain words within that entity, leading to missed detections of violating words.

[0063] To avoid missing violations during content security control, an entity type weight is constructed for each entity type in the knowledge database based on the similarity between word vectors of words within each entity type. The expression is as follows:

[0064]

[0065] In the formula, λ t μ represents the entity type weight of the t-th entity type in the knowledge database. t represents the average similarity among all words of the t-th entity type in the knowledge database; μ' represents the set of average similarities among all entity types in the knowledge database; min() is the minimum value function; ε is a preset positive number to prevent the denominator from being 0. Preferably, in this embodiment, the value of ε is set to 0.01; as other embodiments of this application, the implementer can set the value of ε according to the actual situation.

[0066] V t v represents the number of words of the t-th entity type in the knowledge database; t,i and v t,j Let represent the word vectors corresponding to the i-th and j-th words in the t-th entity type, respectively; cos() represents the calculation of the cosine similarity function, cos(v t,i ,v t,j ) represents the word vector v t,i With word vector v t,jCosine similarity between them.

[0067] The degree of word clustering in each entity type is evaluated using cosine similarity, μ t The larger the value, the more similar the word vectors of the words in the t-th entity type, the more concentrated the vocabulary is in that entity type, and the more concentrated the vocabulary is in the corresponding entity type weight. Entity types with more concentrated vocabulary should be assigned larger weights.

[0068] Step S4: Based on the similarity between each keyword and the word vectors of words in each entity type, and combined with the entity type weights, construct the average similarity of each keyword in each entity type; based on the average similarity, determine the best matching entity type of each keyword and the semantic similarity between the keyword and the best matching entity type.

[0069] Based on the cosine similarity between the word vectors of each keyword in the information board text and the word vectors of each entity type in the knowledge database, the average similarity of each keyword in the information board text for each entity type is constructed, and the expression is:

[0070]

[0071] S a =max(B a )

[0072] In the formula, b a,t V represents the average similarity of the a-th keyword in the information board text across the t-th entity type; t w represents the number of words of the t-th entity type in the knowledge database. a v is the word vector corresponding to the a-th keyword in the information board text; t,i λ is the word vector corresponding to the i-th word in the t-th entity type; t Let w be the entity type weight for the t-th entity type; cos() represents the calculation of the cosine similarity function, cos(w a ,v t,i ) represents the word vector w a With word vector v t,i Cosine similarity between them;

[0073] S a B represents the semantic similarity between the a-th keyword in the information board text and the best-matching entity type. a This represents the set of average similarities of the a-th keyword in the information board text across all entity types; max() is the maximum value function, max(B a ) represents the maximum average similarity of the a-th keyword across all entity types.

[0074] The entity type corresponding to the maximum average similarity of the a-th keyword among all entity types is taken as the best matching entity type of the a-th keyword.

[0075] The matching degree between each keyword and each entity type was determined by calculating the similarity between the word vectors of each keyword and all words in each entity type. Since illegal words may be missed during content security control, the matching degree was enhanced by constructing a weighted index for each entity type to determine the final similarity. Due to λ... t The value of λ is greater than 1, and the higher the clustering of entity types, the greater the weight. Therefore, when calculating the similarity between different keywords and words in highly clustered entity types, a weight λ is added. t This increases the gap between the similarity of different keywords and words in the highly clustered entity type. Compared with not adding this weight, it can increase the gap between the similarity between illegal words and safe words, which is beneficial for the division between illegal words and safe words.

[0076] Maximum value calculation can determine the entity type closest to each keyword in the information board text; when the distance between violation words and safety words in the word vector space is relatively close during the traffic board content security control process, increasing the entity type weight can avoid missing high-risk words due to highly clustered entity types; thus S a The smaller the value, the less semantically similar the keyword a is to its most matching entity type, and the more likely the keyword a is to be a violation.

[0077] Step S5: Determine the dynamic thresholds for various entity types based on the distribution characteristics of the similarity between word vectors of words in each entity type; perform security monitoring of the information board content based on the semantic similarity of each keyword and the dynamic threshold of the corresponding most matching entity type.

[0078] The similarity distribution among word vectors for each entity type approximates a normal distribution. Therefore, the value corresponding to μ-kσ is used, combined with the entity type weight for each entity type, to construct a dynamic threshold for each entity type. This threshold determines whether the word to be tested is a dangerous word that deviates from its corresponding entity type, thus assessing the semantic similarity of keywords. The expression for the dynamic threshold for each entity type is:

[0079] β t =(μ t -k×σ t )×λ t

[0080] In the formula, β t μ represents the dynamic threshold for the t-th entity type in the knowledge database. tσ represents the average similarity among all words of the t-th entity type in the knowledge database; k represents a preset sensitivity coefficient. Preferably, in this embodiment, k is set to 2 to balance the threshold and avoid missed detections; t λ represents the standard deviation of the cosine similarity among the word vectors of all words within the t-th entity type in the knowledge database. t This represents the entity type weight of the t-th entity type in the knowledge database.

[0081] The calculation of standard deviation is a well-known technique; in this application, σ... t The expression is:

[0082] Furthermore, if the semantic similarity between each keyword and its most matching entity type is less than the dynamic threshold of that most matching entity type, then the keyword is determined to be a non-compliant word; otherwise, it is determined to be a compliant word.

[0083] After identifying the prohibited words, locate the corresponding information board, report the information from the information board to the relevant departments, and promptly resolve any issues that arise.

[0084] A diagram illustrating the process of obtaining keywords from the information board text content is shown below. Figure 2 As shown.

[0085] In summary, this application embodiment sets weights for different entity types in the transportation domain. When the distance between illegal words and safe words in the word vector space is relatively close during the safety management of traffic board content, there is a risk of missed detection. By adding corresponding weights when calculating the similarity between different keywords and words in entity types, the gap between the similarity between different keywords and words in entity types is increased. Compared with not adding this weight, the gap between the similarity between illegal words and safe words can be increased, which is beneficial for the division of illegal words and safe words and avoids the omission of high-risk words in highly clustered entity types. Furthermore, the semantic similarity between each keyword and the best matching entity type is determined. The information board content to be detected after word segmentation is dynamically evaluated, and a dynamic threshold is set according to the finally confirmed entity type. The semantic similarity of each keyword is compared with the corresponding dynamic threshold to determine illegal words, complete the safety monitoring of information board content, and further avoid the problem of missed detection of illegal words.

[0086] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this application. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0087] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0088] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them; modifications to the technical solutions described in the foregoing embodiments, or equivalent substitutions of some of the technical features, do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A content detection method for an information board content security management system, characterized in that, The method includes the following steps: Collect information from the traffic information board content in the historical logs of the traffic management system, perform entity type recognition, and obtain a vocabulary set for each entity type; collect images of the traffic information boards to be detected. Extract the text content from the traffic information board image and obtain the keywords of the text content; The word vectors of each keyword and each word of each entity type are obtained through a word vector generation model; the entity type weights of each entity type are constructed based on the similarity between the word vectors of words in each entity type. Based on the similarity between each keyword and the word vectors of words in each entity type, and combined with the entity type weights, the average similarity of each keyword in each entity type is constructed; based on the average similarity, the best matching entity type of each keyword and the semantic similarity between the keyword and the best matching entity type are determined. Dynamic thresholds for various entity types are determined based on the distribution characteristics of the similarity between word vectors of words in each entity type; information board content security monitoring is performed based on the semantic similarity of each keyword and the dynamic threshold of the corresponding most matching entity type. The process for obtaining the dynamic thresholds for various entity types is as follows: Calculate the standard deviation of the cosine similarity among the word vectors of all words within each entity type; based on the average similarity among all words of each entity type, the standard deviation, and the entity type weights, determine the dynamic threshold for each entity type, expressed as: In the formula, Represents the dynamic threshold for the t-th entity type; This represents the average similarity among all words of the t-th entity type. This indicates the preset sensitivity coefficient; The standard deviation of the cosine similarity among the word vectors of all words within the t-th entity type; This represents the entity type weight for the t-th entity type.

2. The content detection method for an information board content security management system as described in claim 1, characterized in that, The keywords for obtaining the text content of the information board are specifically: The text content on the traffic information board is cleaned of non-text characters using regular expressions. The cleaned text content is then segmented and stop words are removed. All the words obtained after processing are used as keywords for the text content of the information board.

3. The content detection method for an information board content security management system as described in claim 1, characterized in that, The process for obtaining the entity type weights for each entity type is as follows: Calculate the cosine similarity between the word vectors of any two words in each entity type, and take the mean of all the cosine similarities of each entity type as the average similarity between all words in each entity type. The entity type weight of each entity type is determined based on the average similarity among all words in each entity type.

4. The content detection method for an information board content security management system as described in claim 3, characterized in that, The expression for the entity type weight of each entity type is as follows: In the formula, This represents the entity type weight for the t-th entity type; This represents the average similarity among all words of the t-th entity type. This represents the set of average similarities across all entity types. The function is for finding the minimum value; Humans are pre-set to positive numbers.

5. The content detection method for an information board content security management system as described in claim 1, characterized in that, The expression for the average similarity of each keyword in each entity type is: In the formula, Let be the average similarity of the a-th keyword in the information board text across the t-th entity type; Let be the number of words for the t-th entity type; Let be the word vector corresponding to the a-th keyword in the information board text; Let be the word vector corresponding to the i-th word in the t-th entity type; Let t be the entity type weight of the t-th entity type; This represents the calculation of the cosine similarity function. Word vectors With word vectors Cosine similarity between them.

6. The content detection method for an information board content security management system as described in claim 1, characterized in that, The most matching entity type for each keyword and the semantic similarity between them are as follows: The entity type corresponding to the maximum average similarity of each keyword across all entity types is taken as the most matching entity type for each keyword; the average similarity of each keyword among the most matching entity types is taken as the semantic similarity between each keyword and the most matching entity type.

7. The content detection method for an information board content security management system as described in claim 1, characterized in that, The aforementioned security monitoring of the information board content specifically includes: If the semantic similarity between each keyword and its most matching entity type is less than the dynamic threshold of that most matching entity type, then the keyword is determined to be a violation word; otherwise, it is determined to be a compliant word.

Images