Short message content multi-level auditing and intelligent sensitive word filtering method and system

Abstract

The present application relates to the technical field of short message content security, and particularly relates to a short message content multi-level auditing and intelligent sensitive word filtering method and system. A semantic reconstruction text is obtained by performing multi-modal deformation identification on a short message to be audited, an auditing decision tree is constructed based on semantic density distribution and sender historical credit weight, a risk propagation coefficient of a semantic unit is calculated to match a decision node and an auditing path is generated, a local risk label and a credibility measure value are outputted by executing auditing and global auditing results are summarized, a semantic vector space representation is constructed when there is a risk, a plurality of adversarial perturbation vectors are generated and gradient detection is performed along the direction thereof to identify a critical semantic boundary, a compliant replacement word is determined by nearest neighbor search under the constraint of keeping the direction of the theme vector unchanged and a compliant short message content is obtained by replacement. The method can efficiently identify deformation sensitive words, improve auditing accuracy and sensitive word filtering efficiency, and ensure the compliance and safety of short message content.

Description

Technical Field

[0001] This invention relates to the field of SMS content security technology, and in particular to a method and system for multi-level review and intelligent sensitive word filtering of SMS content. Background Technology

[0002] SMS content review is a crucial step in ensuring mobile communication security and preventing spam. Current conventional review methods rely on sensitive word databases, scanning SMS text using precise keyword matching and regular expression rules; content matching these rules is blocked or flagged. To improve coverage, some systems introduce machine learning-based text classification models, training classifiers with a large number of labeled SMS messages to predict risk probabilities, followed by manual review of suspected content. These methods are relatively simple to deploy, can handle SMS messages explicitly containing sensitive words, and are the standard technical approach used by most SMS service platforms. However, with the development of communication networks and the increasing diversity of SMS content formats, review systems need to contend with more and more circumvention methods, gradually revealing the limitations of traditional methods.

[0003] Conventional keyword and regular expression-based matching methods struggle to handle deliberately crafted variant texts by senders. These variants include inserting special symbols between sensitive words, using homophones or near-synonyms, and splitting Chinese characters into radicals. These variations render rule matching ineffective, allowing risky SMS messages to bypass review. Fixed rules and classification models lack understanding of semantic context, relying primarily on local lexical features. This can easily misclassify normal SMS messages containing ambiguity or a sarcastic tone as risky, while overlooking messages that express risk through semantic cues. Conventional review strategies apply the same criteria to all senders, failing to consider differences in sender historical reputation. This results in many low-risk users being reviewed repeatedly, while high-risk users receive insufficient monitoring, leading to inefficient allocation of review resources. While some systems introduce manual review mechanisms, the efficiency of manual review is limited by the volume of text, making it difficult to handle large-scale SMS flows continuously, further reducing the timeliness and accuracy of reviews. Summary of the Invention

[0004] This invention provides a method and system for multi-level review of SMS content and intelligent sensitive word filtering, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a method for multi-level review and intelligent sensitive word filtering of SMS content, comprising:

[0006] Multimodal deformation recognition is performed on the SMS messages to be reviewed to obtain semantically reconstructed text;

[0007] Based on the semantic density distribution features of the semantically reconstructed text and the historical reputation weight of the SMS sender, an audit decision tree is constructed.

[0008] Calculate the risk propagation coefficient of each semantic unit in the semantically reconstructed text, match a corresponding decision node for each semantic unit in the review decision tree, dynamically determine the review level corresponding to the current semantic unit along the branch path of the review decision tree, and generate the review path of the SMS to be reviewed;

[0009] Perform the corresponding audit operations according to the audit path, output the local risk label and credibility metric value of the semantic unit, and summarize the results after traversing all decision nodes of the audit decision tree to obtain the global audit result;

[0010] When the global review result is risky, a semantic vector space representation of the semantically reconstructed text is constructed, multiple adversarial perturbation vectors are generated, gradient detection is performed along the direction of the adversarial perturbation vectors, and the critical semantic boundary that causes the decision result of the decision node to change abruptly is identified;

[0011] Under the constraint of keeping the direction of the semantically reconstructed text topic vector unchanged, the semantic units located at the critical semantic boundary are used to determine compliant replacement words and replace them through nearest neighbor search in the semantic vector space to obtain compliant SMS content.

[0012] Multimodal deformation recognition is performed on the SMS messages to be reviewed to obtain semantically reconstructed text, including:

[0013] The stroke connection relationship of each character is mapped to a topological graph structure. By calculating the isomorphic similarity between the topological graph structure of the target character in the SMS to be reviewed and the topological graph structure of each standard character in the standard character library, the glyph deformation characters formed by stroke deformation, component splitting or symbol substitution are identified, and the glyph deformation characters are restored to the corresponding glyph restoration standard characters.

[0014] Extract the initial consonant features, final vowel features, and tone features of each character to construct a phonological feature vector. By calculating the distance metric between the phonological feature vector of the target character in the SMS to be reviewed and the phonological feature vector of each standard character in the standard character library, identify phonologically deformed characters formed by homophonic substitution or mixed pinyin, and restore the phonologically deformed characters to the corresponding phonologically restored standard characters.

[0015] Based on the combination of the glyph restoration standard characters and the phonological restoration standard characters to form the restored text, semantic field migration analysis is performed on the restored text to extract the semantic field attribution identifier of each lexical unit in the restored text, identify lexical units that have undergone abnormal semantic field migration, and restore the lexical units that have undergone semantic field migration to the standard lexical units in the original semantic field through the semantic field backtracking mechanism, thus obtaining the semantically reconstructed text.

[0016] Based on the semantic density distribution features of the semantically reconstructed text and the historical reputation weight of the SMS sender, a review decision tree is constructed, including:

[0017] The semantically reconstructed text is segmented into semantic units. The ratio of the number of key semantic components within a semantic unit to the total length of the semantic unit is calculated to obtain the semantic density value of each semantic unit. The semantic density values ​​of all semantic units are arranged in order of position to form a semantic density distribution sequence.

[0018] The semantic density distribution sequence is subjected to distribution pattern recognition, and the peak position, fluctuation amplitude and gradient change rate of the semantic density distribution sequence are extracted and combined to form semantic density distribution features;

[0019] Obtain the historical sending records of the SMS sender, calculate the ratio of the number of illegal SMS messages in the historical sending records to the total number of SMS messages sent, determine the corresponding reputation decay coefficient based on the numerical range of the ratio, and determine the historical reputation weight of the SMS sender based on the reputation decay coefficient.

[0020] The semantic density distribution feature is used as the content risk decision dimension, and the historical reputation weight is used as the sender reputation decision dimension. Multiple decision segmentation surfaces are set in the decision space formed by the content risk decision dimension and the sender reputation decision dimension. Each decision segmentation surface divides the decision space into different decision regions. Each decision region corresponds to a different review level configuration. The mapping relationship between the decision segmentation surface and the decision region is constructed into a review decision tree.

[0021] Calculate the risk propagation coefficient of each semantic unit in the semantically reconstructed text, match a corresponding decision node for each semantic unit in the review decision tree, dynamically determine the review level corresponding to the current semantic unit along the branch path of the review decision tree, and generate the review path of the SMS to be reviewed, including:

[0022] Dependency relations are parsed for each semantic unit in the semantically reconstructed text. Syntactic dependency edges and semantic dependency edges between each semantic unit and other semantic units are identified, and the number of dependency edge connections is counted. The dependency correlation degree of the semantic unit is calculated based on the number of dependency edge connections.

[0023] Within a preset sliding window, the frequency of co-occurrence of each semantic unit and risk semantic units is counted, and the risk co-occurrence intensity of the semantic units is calculated based on the frequency of co-occurrence.

[0024] The risk propagation coefficient of each semantic unit is obtained by weighted fusion of the dependency correlation degree and the risk co-occurrence strength.

[0025] The semantic density distribution features and the risk propagation coefficient of each semantic unit are used as decision input features, and a corresponding decision node is matched for each semantic unit in the review decision tree;

[0026] Obtain the review level configuration corresponding to the decision region where the decision node is located, and parse it into a review level sequence and review granularity parameters;

[0027] According to the positional order of each semantic unit, the review level sequence and review granularity parameters corresponding to each semantic unit are sequentially concatenated to form the review path of the SMS to be reviewed.

[0028] Perform the corresponding audit operations according to the audit path, output the local risk label and credibility metric value of the semantic unit, and summarize the results after traversing all decision nodes of the audit decision tree to obtain the global audit result, including:

[0029] According to the positional order of each semantic unit in the review path, the decision node and review level sequence corresponding to each semantic unit are obtained sequentially.

[0030] At each decision node, the type of review operation to be performed for the current semantic unit is determined according to the review level sequence. Risk features are extracted from the text content of the semantic unit. The extracted risk features are matched and calculated with a preset risk feature template. Based on the matching result, a local risk label for the semantic unit is generated.

[0031] The confidence level of the local risk label is evaluated, and the number and matching strength of the risk features matched during the generation of the local risk label are counted. The confidence metric of the local risk label is calculated based on the number of risk features and the matching strength.

[0032] The local risk label and the trust metric value are associated with the location information of the semantic unit and stored to form the audit record of the semantic unit;

[0033] Traverse all decision nodes of the review decision tree and collect the review records of all semantic units;

[0034] Aggregate and analyze the local risk tags in all collected review records. Based on the distribution pattern of the local risk tags and the statistical characteristics of the credibility metric, determine the overall risk level and risk type of the SMS to be reviewed, and generate the global review result.

[0035] Construct a semantic vector space representation of the semantically reconstructed text, generate multiple adversarial perturbation vectors, perform gradient probing along the directions of the adversarial perturbation vectors, and identify critical semantic boundaries that cause abrupt changes in the decision results of the decision nodes, including:

[0036] Each semantic unit in the semantically reconstructed text is vectorized and encoded, and the vector representation of each semantic unit is mapped to a unified semantic vector space. The semantic vector space representation of the semantically reconstructed text is constructed based on the semantic relationships between each semantic unit.

[0037] Multiple perturbation directions are selected in the semantic vector space, and a perturbation vector sequence with increasing perturbation amplitude is generated in each perturbation direction. The perturbation vector sequence is used as an adversarial perturbation vector.

[0038] The adversarial perturbation vector is superimposed on the semantic vector space representation to obtain the perturbed semantic vector space representation.

[0039] The perturbated semantic vector space representation is input into the review decision tree to obtain the judgment results of each decision node in the review decision tree;

[0040] The perturbation amplitude is gradually increased along the direction of the adversarial perturbation vector, and the change of the decision result of the decision node is monitored in real time. When the decision result changes from a non-risk state to a risk state or from a risk state to a non-risk state, the perturbation amplitude and perturbation direction at the time of the state change are recorded.

[0041] Based on the recorded perturbation amplitude and perturbation direction, a critical semantic boundary that causes a sudden change in the judgment result is determined in the semantic vector space.

[0042] Under the constraint of keeping the direction of the semantically reconstructed text topic vector unchanged, the semantic units located at the critical semantic boundary are used to determine and replace compliant replacement words through nearest neighbor search in the semantic vector space to obtain compliant SMS content, including:

[0043] Extract semantic units located at the critical semantic boundary from the semantically reconstructed text, and mark the semantic units as semantic units to be replaced;

[0044] The vector representations of all semantic units in the semantically reconstructed text are weighted and aggregated to calculate the topic vector of the semantically reconstructed text, and the direction of the topic vector in the semantic vector space is determined as the topic constraint direction.

[0045] The vector representation of the semantic unit to be replaced is decomposed into components along the topic constraint direction and components perpendicular to the topic constraint direction.

[0046] In the semantic vector space, a nearest neighbor search is performed with the vector representation of the semantic unit to be replaced as the center. Candidate replacement word vectors with a distance less than a preset distance threshold from the vector representation of the semantic unit to be replaced are selected. The deviation value between the projection component of each candidate replacement word vector along the topic constraint direction and the component of the semantic unit to be replaced along the topic constraint direction is calculated.

[0047] Candidate replacement words corresponding to candidate replacement word vectors whose deviation values ​​are less than a preset deviation threshold are selected as compliant replacement words;

[0048] Replace the semantic unit to be replaced with the compliant replacement word, and update the semantically reconstructed text;

[0049] The updated semantically reconstructed text is then restored to its original form to obtain compliant SMS content.

[0050] A second aspect of this invention provides a multi-level SMS content review and intelligent sensitive word filtering system, comprising:

[0051] The multimodal recognition unit is used to perform multimodal deformation recognition on the SMS messages to be reviewed, and obtain semantically reconstructed text.

[0052] The semantic reconstruction unit is used to construct an audit decision tree based on the semantic density distribution features of the semantically reconstructed text and the historical reputation weight of the SMS sender.

[0053] A decision tree construction unit is used to calculate the risk propagation coefficient of each semantic unit in the semantically reconstructed text. In the review decision tree, a corresponding decision node is matched for each semantic unit. The review level corresponding to the current semantic unit is dynamically determined along the branch path of the review decision tree, and the review path of the SMS to be reviewed is generated.

[0054] The risk propagation unit is used to perform corresponding audit operations according to the audit path, output the local risk label and credibility measurement value of the semantic unit, and summarize the global audit results after traversing all decision nodes of the audit decision tree.

[0055] The review execution unit is used to construct the semantic vector space representation of the semantically reconstructed text when the global review result is risky, generate multiple adversarial perturbation vectors, perform gradient detection along the direction of the adversarial perturbation vectors, and identify the critical semantic boundary that causes the decision result of the decision node to change abruptly.

[0056] A critical semantic unit is used to determine compliant replacement words and replace semantic units located at the critical semantic boundary by means of nearest neighbor search in the semantic vector space, while keeping the direction of the semantically reconstructed text topic vector unchanged, so as to obtain compliant SMS content.

[0057] A third aspect of the present invention provides an electronic device, comprising:

[0058] processor;

[0059] Memory used to store processor-executable instructions;

[0060] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0061] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0062] The beneficial effects of this application are as follows:

[0063] Multimodal deformation recognition performs deep semantic reconstruction of SMS content, effectively cracking deformation avoidance methods such as homophones, word splitting, and symbol combination, restoring the true semantic information. A review decision tree is constructed based on semantic density distribution features and the sender's historical reputation weight, integrating the content's inherent risk attributes with behavioral profiles. This gives the decision path personalized adaptability, significantly improving the capture rate of high-risk content while reducing interference with normal users.

[0064] The risk propagation coefficient and decision node matching mechanism dynamically determine the review level along the branch path, transforming the review process from static rule matching to dynamic path reasoning. Local risk labels and credibility metrics are aggregated level by level, accurately depicting the distribution and propagation chain of risks in SMS messages, avoiding missed or false detections caused by isolated judgments. Even if risk semantics are scattered and hidden in long texts, they can be effectively identified through multi-node collaborative analysis.

[0065] When there is a risk to the overall result, the critical semantic boundary that causes a sudden change in the review judgment is accurately located by adversarial perturbation and gradient detection in the semantic vector space. The critical semantic unit is replaced with compliant words, and the topic vector direction is kept unchanged during the replacement process. This ensures that the rewritten text message does not deviate from the original meaning and is semantically natural and fluent, which greatly reduces the false negative rate of the review. While ensuring security, it improves the usability and user experience of compliant content and achieves a balance between security filtering and information expression. Attached Figure Description

[0066] Figure 1 A flowchart illustrating the multi-level review and intelligent sensitive word filtering method for SMS content;

[0067] Figure 2 A schematic diagram illustrating the process of obtaining semantically reconstructed text from multimodal deformation recognition of SMS messages awaiting review;

[0068] Figure 3This is a schematic diagram of semantic density distribution and decision segmentation surface;

[0069] Figure 4 This is a heatmap of the critical semantic boundary and a schematic diagram of the unit distribution. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0072] Figure 1 This is a flowchart illustrating the multi-level review and intelligent sensitive word filtering method for SMS content according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0073] Multimodal deformation recognition is performed on the SMS messages to be reviewed to obtain semantically reconstructed text;

[0074] Based on the semantic density distribution features of the semantically reconstructed text and the historical reputation weight of the SMS sender, an audit decision tree is constructed.

[0075] Calculate the risk propagation coefficient of each semantic unit in the semantically reconstructed text, match a corresponding decision node for each semantic unit in the review decision tree, dynamically determine the review level corresponding to the current semantic unit along the branch path of the review decision tree, and generate the review path of the SMS to be reviewed;

[0076] Perform the corresponding audit operations according to the audit path, output the local risk label and credibility metric value of the semantic unit, and summarize the results after traversing all decision nodes of the audit decision tree to obtain the global audit result;

[0077] When the global review result is risky, a semantic vector space representation of the semantically reconstructed text is constructed, multiple adversarial perturbation vectors are generated, gradient detection is performed along the direction of the adversarial perturbation vectors, and the critical semantic boundary that causes the decision result of the decision node to change abruptly is identified;

[0078] Under the constraint of keeping the direction of the semantically reconstructed text topic vector unchanged, the semantic units located at the critical semantic boundary are used to determine compliant replacement words and replace them through nearest neighbor search in the semantic vector space to obtain compliant SMS content.

[0079] In one optional implementation, the SMS message to be reviewed undergoes multimodal deformation recognition to obtain semantically reconstructed text, including:

[0080] The stroke connection relationship of each character is mapped to a topological graph structure. By calculating the isomorphic similarity between the topological graph structure of the target character in the SMS to be reviewed and the topological graph structure of each standard character in the standard character library, the glyph deformation characters formed by stroke deformation, component splitting or symbol substitution are identified, and the glyph deformation characters are restored to the corresponding glyph restoration standard characters.

[0081] Extract the initial consonant features, final vowel features, and tone features of each character to construct a phonological feature vector. By calculating the distance metric between the phonological feature vector of the target character in the SMS to be reviewed and the phonological feature vector of each standard character in the standard character library, identify phonologically deformed characters formed by homophonic substitution or mixed pinyin, and restore the phonologically deformed characters to the corresponding phonologically restored standard characters.

[0082] Based on the combination of the glyph restoration standard characters and the phonological restoration standard characters to form the restored text, semantic field migration analysis is performed on the restored text to extract the semantic field attribution identifier of each lexical unit in the restored text, identify lexical units that have undergone abnormal semantic field migration, and restore the lexical units that have undergone semantic field migration to the standard lexical units in the original semantic field through the semantic field backtracking mechanism, thus obtaining the semantically reconstructed text.

[0083] like Figure 2 As shown, the method includes:

[0084] For the glyph-level deformation problem in processing short messages to be reviewed, for each character, the connection relationship of its strokes is abstracted into a topological graph structure, where the nodes of the graph correspond to stroke endpoints or intersection points, and the edges of the graph correspond to the connection paths between stroke segments. For the target character in the short message to be reviewed, its corresponding topological graph structure is extracted, and compared one by one with the topological graph structure of each standard character pre-stored in the standard character library, to calculate the isomorphism similarity between the two graphs. The calculation of isomorphism similarity comprehensively considers the matching degree of node degree sequences, the consistency of edge connection patterns, and the coverage ratio of subgraph structures. When the isomorphism similarity between the topological graph of a certain target character and the topological graph of a certain standard character exceeds a preset threshold, the target character is determined as a glyph-deformed character camouflaged by means of stroke deformation, component splitting or symbol substitution, and is restored to the corresponding standard character for glyph restoration. For example, the scenario where a "Jiong"-shaped symbol is split and camouflaged into other glyphs, or the scenario where local deformation of Chinese character strokes is performed to avoid keyword detection, can be effectively identified through topological graph isomorphism similarity comparison. After the glyph restoration is completed, all glyph-deformed characters are replaced with the corresponding standard characters for glyph restoration, to form a preliminary restored text at the glyph level.

[0085] On the basis of completing the identification of glyph deformation, deformation at the phonological level is further identified. For each character in the short message to be reviewed, its initial consonant feature, final vowel feature and tone feature are extracted, and the three types of features are spliced to construct a phonological feature vector. The initial consonant feature can be encoded as a combined representation of the consonant articulation position and articulation manner, the final vowel feature can be encoded as a multi-dimensional representation of the vowel tongue position height, front / back position, and rounded lip state, and the tone feature corresponds to the category codes of the four tones of Mandarin and the neutral tone. For the target character in the short message to be reviewed, the distance metric value between its phonological feature vector and the phonological feature vector of each standard character in the standard character library is calculated. The distance metric can adopt a weighted Euclidean distance, wherein the weight of the tone feature can be adjusted according to the actual application scenario. When the distance metric value between a certain target character and a certain standard character is lower than a preset threshold, the target character is determined as a phonologically deformed character formed by means of homophonic substitution or mixed use of pinyin, and is restored to the corresponding standard character for phonological restoration. Typical phonological deformation scenarios include replacing a certain character in a sensitive word with a homophonic Chinese character, mixing pinyin letters into Chinese words, or using dialect homophonic characters to avoid filtering. After the phonological restoration is completed, all phonologically deformed characters are replaced with the corresponding standard characters for phonological restoration.

[0086] The glyphic and phonological standard characters are recombined according to their position in the original text message to form the restored text. This text has undergone dual correction at the character level, involving both glyphic and phonological deformations. However, hidden semantic deformations may still exist, i.e., circumventing censorship by introducing words into unconventional semantic fields. When performing semantic field transfer analysis on the restored text, each lexical unit is mapped to a pre-constructed multi-level semantic field system, and the semantic field attribution identifier of each lexical unit is extracted. The semantic field attribution identifier describes the semantic category domain to which the word belongs in the current context.

[0087] Lexical units exhibiting anomalous semantic field migration are identified by comparing their current semantic field affiliation with the affiliation of the mainstream semantic field in the standard corpus, thus calculating the semantic field migration distance. This distance is measured based on the path length between two semantic field nodes in the semantic field hierarchy tree; a longer path indicates a greater degree of semantic field migration. When the semantic field migration distance of a lexical unit exceeds a preset threshold, it is determined that the lexical unit has undergone anomalous semantic field migration. For identified anomalous semantic field migration lexical units, a semantic field backtracking mechanism is initiated. This mechanism retrieves the standard lexical unit in the standard semantic field that best matches the current context semantics, restoring the anomalously migrated lexical unit to its original standard semantic field. The core of the semantic field backtracking mechanism lies in utilizing the semantic field affiliation information of other lexical units within the context window to jointly constrain the selection of the target semantic field for backtracking, avoiding ambiguity during single-lexical unit backtracking.

[0088] After three levels of processing—character shape restoration, phonological restoration, and semantic field backtracking—the semantically reconstructed text is obtained. This text has undergone deformation removal processing in three dimensions: character morphology, pronunciation features, and semantic attribution, effectively resisting SMS senders' attempts to circumvent sensitive word filtering through multi-level deformation techniques. The quality of the semantically reconstructed text, as input to subsequent multi-level review processes, directly determines the accuracy of the review decision tree construction and risk propagation coefficient calculation. In practical applications, the three sub-processes—character shape deformation recognition, phonological deformation recognition, and semantic field transfer analysis—can be executed in parallel to improve processing efficiency. The final results are fused at the lexical unit granularity. When deformation is detected in multiple dimensions for the same lexical unit, the semantic field backtracking result is prioritized as the final restoration basis to ensure semantic consistency.

[0089] In constructing the standard character library, the glyph topology library needs to cover common stroke deformation patterns, including stroke additions and subtractions, stroke deformations, and component position adjustments; the phonological feature vector library needs to cover all Mandarin phonetic combinations and common dialect homophonic mapping relationships; the semantic field system needs to be trained on a large-scale corpus to ensure that the boundaries of each semantic field are clear and the coverage is complete. Regular updates and maintenance of these three types of libraries are crucial for ensuring the accuracy of multimodal deformation recognition, and special attention should be paid to the timely inclusion of emerging internet slang and new deformation methods.

[0090] In one optional implementation, a review decision tree is constructed based on the semantic density distribution features of the semantically reconstructed text and the historical reputation weight of the SMS sender, including:

[0091] The semantically reconstructed text is segmented into semantic units. The ratio of the number of key semantic components within a semantic unit to the total length of the semantic unit is calculated to obtain the semantic density value of each semantic unit. The semantic density values ​​of all semantic units are arranged in order of position to form a semantic density distribution sequence.

[0092] The semantic density distribution sequence is subjected to distribution pattern recognition, and the peak position, fluctuation amplitude and gradient change rate of the semantic density distribution sequence are extracted and combined to form semantic density distribution features;

[0093] Obtain the historical sending records of the SMS sender, calculate the ratio of the number of illegal SMS messages in the historical sending records to the total number of SMS messages sent, determine the corresponding reputation decay coefficient based on the numerical range of the ratio, and determine the historical reputation weight of the SMS sender based on the reputation decay coefficient.

[0094] The semantic density distribution feature is used as the content risk decision dimension, and the historical reputation weight is used as the sender reputation decision dimension. Multiple decision segmentation surfaces are set in the decision space formed by the content risk decision dimension and the sender reputation decision dimension. Each decision segmentation surface divides the decision space into different decision regions. Each decision region corresponds to a different review level configuration. The mapping relationship between the decision segmentation surface and the decision region is constructed into a review decision tree.

[0095] The semantically reconstructed text is segmented into semantic units, based on the fundamental principle of semantic integrity. The text is divided into several independent semantic units. These units are defined by semantic boundaries rather than simple character boundaries, typically phrases, clauses, or functional semantic fragments. For each semantic unit, the number of key semantic components within it is counted. Key semantic components include entity words, action words, sentiment words, and modifiers with clear referential relationships. The semantic density value of the semantic unit is obtained by dividing the number of key semantic components by the total length of the semantic unit (in characters or lexical units). The calculation formula is: ,in Indicates the first The number of key semantic components within a semantic unit Indicates the first The total length of each semantic unit. A higher semantic density value indicates that the semantic unit carries more dense semantic information within a finite length, and is usually highly related to the way risk content is expressed. Arranging all semantic units in order of their position in the original text yields the semantic density distribution sequence. ,in This represents the total number of semantic units.

[0096] Distribution pattern recognition is performed on semantic density distribution sequences, focusing on extracting three types of features: peak position, fluctuation amplitude, and gradient rate of change. The peak position is determined by detecting the index of local maxima in the sequence, denoted as . , indicating the first The position number of the semantic unit where each peak is located; the distribution pattern of the peak positions reflects the concentration and dispersion of high semantic density content in the SMS message. If the peaks are concentrated at the end or beginning of the SMS message, it suggests a specific censorship avoidance strategy. The fluctuation amplitude is obtained by calculating the difference between the maximum and minimum density values ​​in the sequence, denoted as . ,in The maximum value in the sequence. The minimum value in the sequence is denoted as . Large fluctuations indicate a significantly uneven distribution of semantic density in the SMS content, a distribution characteristic often associated with deliberately diluting sensitive content. The gradient rate of change is obtained by calculating the mean absolute value of the differences between the density values ​​of adjacent semantic units, denoted as . This reflects the drastic changes in semantic density within the text. The set of peak positions, fluctuation amplitude, and gradient change rate are combined to form a semantic density distribution feature vector, which serves as the input for the content risk decision dimension in the subsequent decision space construction.

[0097] After obtaining the sender's historical sending records, the ratio of the number of non-compliant SMS messages sent by that sender within a certain time window to the total number of SMS messages sent is calculated and denoted as . ,in Indicates the number of violation text messages. This indicates the total number of text messages sent. (Based on the proportion) The corresponding reputation decay coefficient is determined based on the pre-defined grading rules within the given numerical range. Specifically, when Falling into the lower range (e.g.) )hour, Taking a smaller value indicates that the sender has a good historical reputation and a low degree of reputation decay; when Falling into the middle range (e.g.) )hour, Taking the median value reflects a certain history of violations; when Falling into a higher range (e.g.) )hour, A larger value indicates a higher tendency for the sender to violate regulations and a greater degree of reputational degradation. Historical reputation weight. Based on reputation decay coefficient The calculation shows that it satisfies... This results in a negative correlation between historical reputation weight and the proportion of violations. Historical reputation weight The range of values ​​is constrained within a reasonable range to avoid extreme values ​​interfering with the subsequent construction of the decision space.

[0098] Using semantic density distribution features as a dimension for content risk decision-making, and historical reputation weighting As the sender reputation decision dimension, the two dimensions together constitute a two-dimensional decision space. In this decision space, the coordinates of the content risk decision dimension are represented by the comprehensive score of the semantic density distribution feature vector, while the coordinates of the sender reputation decision dimension are represented by the historical reputation weight. This is a direct representation. Multiple decision-making facets are set within this decision space. The position and direction of these facets are predetermined based on historical review data and risk distribution patterns. Each facet divides the decision space into two or more sub-regions, each corresponding to a different review level configuration. For example, when the overall content risk score is high and the historical reputation weight is low, the corresponding decision region is mapped to the strictest review level configuration, requiring full manual review of the SMS message. When the overall content risk score is at a medium level and the historical reputation weight is high, the corresponding decision region is mapped to a medium review level configuration, employing automated review combined with sampling review. When the overall content risk score is low and the historical reputation weight is high, the corresponding decision region is mapped to a lightweight review level configuration, requiring only basic keyword matching to pass.

[0099] The mapping relationship between decision segmentation surfaces and decision regions is organized into a tree structure, forming a review decision tree. The root node of the tree corresponds to the entrance of the overall decision space, each internal node corresponds to a decision segmentation surface, and the leaf nodes correspond to specific review level configurations. When traversing along the branch paths of the tree, based on the coordinate position of the SMS message to be reviewed in the decision space, its relative positional relationship with each decision segmentation surface is determined sequentially, ultimately falling into the decision region corresponding to a certain leaf node, thus determining the review level that the SMS message should be subject to. This tree-like organization method makes the process of determining the review level have a clear logical path, facilitates the traceability and interpretation of review decisions, and also supports the dynamic adjustment of the position of decision segmentation surfaces after the accumulation of new historical data, in order to adapt to the evolution of illegal SMS methods.

[0100] By incorporating the two heterogeneous information types of semantic density distribution features and historical reputation weights into the same decision framework, the organic integration of content-level risk features and sender-level reputation features is achieved. This enables the review decision tree to respond simultaneously to the risk signals of the SMS content itself and the sender's historical behavior patterns. Thus, under the condition of limited review resources, the most stringent review resources are prioritized for SMS messages with the highest overall risk, thereby improving the overall review efficiency and accuracy.

[0101] Figure 3 This diagram illustrates the semantic density distribution and decision segmentation surface. It shows the density distribution sequence of semantic units in the semantically reconstructed text. The horizontal axis represents the position of semantic units (in text order), and the vertical axis represents the semantic density value (number of key components / unit length). The three horizontal dashed lines represent the segmentation thresholds in the decision tree based on density and reputation weights, corresponding to lightweight, medium, and strict review regions, respectively. Peak positions and fluctuation amplitudes are used to construct semantic density distribution features. Regions with concentrated peaks trigger higher-level review, fully demonstrating the guiding role of density distribution features in the decision tree branch paths and providing a quantitative basis for determining dynamic review levels.

[0102] In one optional implementation, the risk propagation coefficient of each semantic unit in the semantically reconstructed text is calculated, a corresponding decision node is matched for each semantic unit in the review decision tree, the review level corresponding to the current semantic unit is dynamically determined along the branch path of the review decision tree, and the review path of the SMS to be reviewed is generated, including:

[0103] Dependency relations are parsed for each semantic unit in the semantically reconstructed text. Syntactic dependency edges and semantic dependency edges between each semantic unit and other semantic units are identified, and the number of dependency edge connections is counted. The dependency correlation degree of the semantic unit is calculated based on the number of dependency edge connections.

[0104] Within a preset sliding window, the frequency of co-occurrence of each semantic unit and risk semantic units is counted, and the risk co-occurrence intensity of the semantic units is calculated based on the frequency of co-occurrence.

[0105] The dependency correlation degree and the risk co-occurrence intensity are weighted and fused to obtain the risk propagation coefficient of each semantic unit;

[0106] The semantic density distribution features and the risk propagation coefficient of each semantic unit are used as decision input features, and a corresponding decision node is matched for each semantic unit in the review decision tree;

[0107] Obtain the review level configuration corresponding to the decision region where the decision node is located, and parse it into a review level sequence and review granularity parameters;

[0108] According to the positional order of each semantic unit, the review level sequence and review granularity parameters corresponding to each semantic unit are sequentially concatenated to form the review path of the SMS to be reviewed.

[0109] The risk propagation coefficient of each semantic unit in the semantically reconstructed text is calculated to complete dependency parsing. The core objective of dependency parsing is to reveal the structural connections between semantic units, specifically including two types of dependency edges: syntactic dependency edges and semantic dependency edges. Syntactic dependency edges reflect the dominance and subordination relationships of semantic units at the syntactic level, such as subject-predicate, verb-object, and attributive-head relationships; semantic dependency edges reflect the semantic connections between semantic units at the semantic level, such as agent, patient, instrument, and purpose relationships. For each semantic unit, the number of syntactic dependency edges and the number of semantic dependency edges between it and other semantic units are counted, and the sum of the two constitutes the total number of dependency edge connections for that semantic unit. Let the first semantic unit be... The number of syntactic dependency edges for a semantic unit is The number of semantic dependency edges is Then the dependency relation degree of this semantic unit Defined as: ,in and These are the weight coefficients for grammatical dependency edges and semantic dependency edges, respectively. These can be pre-set based on corpus statistical experience; generally, semantic dependency edges contribute more significantly to risk propagation. The value is generally greater than The higher the degree of dependency, the closer the semantic unit is to other units in the text structure, and the wider the potential risk impact range.

[0110] To calculate the co-occurrence strength of risks, a sliding window mechanism is introduced to constrain the context range. Let the width of the sliding window be... For the first Each semantic unit, with it as the center, is arranged before and after it. Within a window, the frequency of co-occurrence of a semantic unit with each member of a predefined risk semantic unit set is counted. The risk semantic unit set is provided by a pre-built risk semantic lexicon, covering multiple categories of risk terms such as fraud, inducement, and illegality. Let the frequency of co-occurrence within the window be... The set of semantic units and risk semantic units. The frequency of the common occurrence of these risk terms is: Then the risk co-occurrence strength of this semantic unit The calculation is as follows: ,in This represents the total number of risk terms in the risk semantic unit set. For the first The risk weight of each risk term reflects the difference in the degree of harm caused by different categories of risk terms. (Sliding window width) The setting needs to strike a balance between context coverage and computational efficiency, and can typically be set to 5 to 10 semantic units.

[0111] Dependency Co-occurrence intensity with risk Perform weighted fusion to obtain the first Risk propagation coefficient of a semantic unit : ,in and To integrate the weighting coefficients, satisfy the following conditions: And all of them are greater than zero. and The value of can be obtained through supervised training or grid search optimization on labeled corpora. Risk propagation coefficient The value comprehensively reflects the semantic unit's radiation capacity in the dependency structure and its co-occurrence tendency in the risk context. The larger the value, the more critical the role of the semantic unit in the risk propagation chain.

[0112] After calculating the risk propagation coefficient, the semantic density distribution features and risk propagation coefficient of each semantic unit are used as decision input features. A corresponding decision node is then matched for each semantic unit in the review decision tree. Internal nodes of the review decision tree correspond to feature threshold judgment logic, while leaf nodes correspond to specific review decision regions. For the first... Each semantic unit is assigned its semantic density value and risk propagation coefficient. The feature vector is constructed by traversing the decision tree layer by layer from the root node, based on the feature threshold conditions of each internal node, until a leaf node is reached. The decision region where the leaf node is located is the location of the decision node matched by the semantic unit. Different decision regions correspond to different risk level classifications, such as low-risk, medium-risk, and high-risk regions, and each region is pre-configured with corresponding review level information.

[0113] After obtaining the review level configuration corresponding to the decision region where the decision node is located, it is parsed into two types of parameters: review level sequence and review granularity parameter. The review level sequence describes the order of review levels that the semantic unit needs to go through, such as different combinations like "lexical review → syntactic review → semantic review" or "direct pass"; the review granularity parameter specifies the fineness of the inspection operation in each review level, including different granularity options such as character level, phrase level, and sentence level. The review level sequence and review granularity parameter together determine the review intensity and depth applied to the semantic unit. Decision regions with higher risk levels usually have a review level sequence containing more review levels and more refined review granularity parameters.

[0114] Based on the positional order of each semantic unit in the semantically reconstructed text, the corresponding review level sequence and review granularity parameters of each semantic unit are sequentially concatenated to form a complete review path for the SMS message to be reviewed. During the concatenation process, the connection between review levels of adjacent semantic units needs to be addressed: if adjacent semantic units have the same review level, they can be merged into the same review segment to reduce redundant operations; if adjacent semantic units have different review levels, a level switching marker is inserted into the review path so that the boundaries of review operations can be correctly identified in subsequent execution stages. The final generated review path is stored in the form of a structured sequence, with each element containing three fields: the position index of the semantic unit, the corresponding review level sequence, and the review granularity parameter, forming a complete and sequentially executable review operation guide.

[0115] The dynamic generation mechanism of the review path enables differentiated allocation of review resources for SMS messages with different risk characteristics. For SMS messages with a low overall risk propagation coefficient and a stable semantic density distribution, the review path tends to allocate shallower review operations to save computational resources. For SMS messages containing semantic units with high risk propagation coefficients or abnormally concentrated semantic density distributions, the review path will allocate deeper review operations to the corresponding semantic units to ensure the comprehensiveness and accuracy of risk identification. Compared with a fixed strategy of uniformly applying the highest level of review to all SMS messages, this decision tree-based dynamic path generation method significantly reduces the overall computational overhead while ensuring review quality, making it suitable for deployment in large-scale SMS traffic scenarios.

[0116] In one optional implementation, the corresponding audit operation is performed according to the audit path, the local risk label and credibility metric value of the semantic unit are output, and the global audit result is obtained by traversing all decision nodes of the audit decision tree, including:

[0117] According to the positional order of each semantic unit in the review path, the decision node and review level sequence corresponding to each semantic unit are obtained sequentially.

[0118] At each decision node, the type of review operation to be performed for the current semantic unit is determined according to the review level sequence. Risk features are extracted from the text content of the semantic unit. The extracted risk features are matched and calculated with a preset risk feature template. Based on the matching result, a local risk label for the semantic unit is generated.

[0119] The confidence level of the local risk label is evaluated, and the number and matching strength of the risk features matched during the generation of the local risk label are counted. The confidence metric of the local risk label is calculated based on the number of risk features and the matching strength.

[0120] The local risk label and the trust metric value are associated with the location information of the semantic unit and stored to form the audit record of the semantic unit;

[0121] Traverse all decision nodes of the review decision tree and collect the review records of all semantic units;

[0122] Aggregate and analyze the local risk tags in all collected review records. Based on the distribution pattern of the local risk tags and the statistical characteristics of the credibility metric, determine the overall risk level and risk type of the SMS to be reviewed, and generate the global review result.

[0123] Following the positional order of semantic units in the review path, the decision node and review level sequence corresponding to each semantic unit are obtained sequentially. The review path records the traversal order of each semantic unit in the review decision tree, as well as the review level assigned to each semantic unit. Review levels are typically divided into multiple tiers, such as light review, medium review, and strict review, with different levels corresponding to different review operation types and computational resource inputs. Obtaining the positional order ensures that the review operations proceed sequentially according to the natural arrangement of semantic units in the original SMS text, avoiding the loss of contextual semantic information due to out-of-order processing.

[0124] At each decision node, the type of review operation to be performed on the current semantic unit is determined based on the review level sequence. Review operation types include precise keyword matching, semantic similarity comparison, contextual analysis, and multimodal fusion judgment. When extracting risk features from the text content of the semantic unit, the extracted risk features cover sensitive words at the lexical level, illegal expression patterns at the phrase level, circumventional language structures at the syntactic level, and implicit expression tendencies at the semantic level. After extraction, these risk features are matched against a pre-built risk feature template. The risk feature template is a pre-built and continuously updated structured knowledge base that includes feature representations of known illegal content, covering feature patterns of various risk types such as fraud, harassment, and illegal advertising. The matching calculation uses a multi-dimensional similarity measurement method, comprehensively considering the cosine similarity and edit distance of the feature vectors, thus maintaining a strong ability to identify circumvention methods such as variant words, homophones, and word splitting. Based on the matching results, a local risk label is generated for the current semantic unit, which contains information in two dimensions: risk type identification and risk level rating.

[0125] Confidence assessment is performed on local risk labels, and the number and strength of risk features matched during the generation of these local risk labels are statistically analyzed. Let the first... The number of risk features matched by each semantic unit is The mean matching strength of each matching feature is Then the credibility metric of the local risk label of the semantic unit Through comprehensive and The calculation yields the following result, specifically expressed as: ,in This is a normalization function used to map the number of features to... The design incorporates a range to ensure consistent dimensions for the credibility metric. When the number of matched risk features is small but the match strength of each individual feature is extremely high, the credibility metric can still reach a high level, reflecting strong single evidence with high confidence. Conversely, when the number of matched risk features is large but the match strengths of each feature are dispersed, the credibility metric reflects the overall confidence level supported by multiple pieces of evidence. This two-dimensional credibility metric design enables subsequent global aggregation analysis to distinguish between high-confidence risks and low-confidence suspected risks, improving the accuracy of the overall review results.

[0126] Local risk labels and credibility metrics are associated with the location information of semantic units to form the review record for that semantic unit. The location information includes the starting and ending character offsets of the semantic unit in the original SMS text, as well as the decision node number corresponding to that semantic unit in the review decision tree. The review record is stored in a structured manner, with fields including: semantic unit identifier, location range, decision node number, review level, local risk label, credibility metric, and review operation type. The structured storage of review records provides a complete data foundation for subsequent global aggregation analysis after traversal, and also supports traceability queries of the review process, facilitating rapid location of specific risk semantic units during manual review.

[0127] The process traverses all decision nodes in the review decision tree, collecting review records for all semantic units. The traversal follows a breadth-first or depth-first strategy to ensure no semantic unit assigned to a decision node is overlooked. During traversal, semantic units with a light review level have a local risk label of "no risk" in their review records; these records are also collected and included in the global aggregation analysis to ensure the complete coverage of the overall SMS content by the global review results. After traversal, a complete set of review records covering all semantic units is generated.

[0128] Aggregate analysis is performed on local risk tags from all collected review records to analyze their distribution patterns and statistical characteristics of credibility metrics. Distribution pattern analysis focuses on the positional distribution of risk tags within the SMS text, such as whether they are concentrated at the beginning or scattered throughout the text. Concentrated distribution often corresponds to clearly defined violations, while scattered distribution suggests deliberate segmentation to evade detection. Statistical characteristics of credibility metrics include the mean, variance, and percentage of semantic units exceeding a preset high confidence threshold for all semantic units. Let the set of credibility metrics for all semantic units be... (here) (Using the meaning of the total number of semantic units), the high confidence threshold is... The percentage of semantic units exceeding this threshold By statistical analysis Quantity and The ratio is obtained. Based on the distribution pattern of local risk labels and the statistical characteristics of credibility metrics, the overall risk level and risk type of the SMS messages to be reviewed are comprehensively determined. The overall risk level is divided into four levels: low risk, medium risk, high risk, and very high risk. The risk type corresponds to the predefined category system in the risk feature template, including major types such as telecom fraud, illegal marketing, pornographic content, and gambling-related traffic. When multiple local risk labels correspond to different risk types, the final risk type is determined by a weighted voting method based on credibility metrics, ensuring that the global review results accurately reflect the main illegal attributes of the SMS content. The final generated global review results include the overall risk level, risk type, a list of the locations of major risk semantic units, and a global confidence score, providing a complete basis for subsequent compliance processing or blocking decisions.

[0129] In one optional implementation, a semantic vector space representation of the semantically reconstructed text is constructed, multiple adversarial perturbation vectors are generated, gradient probing is performed along the direction of the adversarial perturbation vectors, and critical semantic boundaries that cause abrupt changes in the decision results of the decision nodes are identified, including:

[0130] Each semantic unit in the semantically reconstructed text is vectorized and encoded, and the vector representation of each semantic unit is mapped to a unified semantic vector space. The semantic vector space representation of the semantically reconstructed text is constructed based on the semantic relationships between each semantic unit.

[0131] Multiple perturbation directions are selected in the semantic vector space, and a perturbation vector sequence with increasing perturbation amplitude is generated in each perturbation direction. The perturbation vector sequence is used as an adversarial perturbation vector.

[0132] The adversarial perturbation vector is superimposed on the semantic vector space representation to obtain the perturbed semantic vector space representation.

[0133] The perturbated semantic vector space representation is input into the review decision tree to obtain the judgment results of each decision node in the review decision tree;

[0134] The perturbation amplitude is gradually increased along the direction of the adversarial perturbation vector, and the change of the decision result of the decision node is monitored in real time. When the decision result changes from a non-risk state to a risk state or from a risk state to a non-risk state, the perturbation amplitude and perturbation direction at the time of the state change are recorded.

[0135] Based on the recorded perturbation amplitude and perturbation direction, a critical semantic boundary that causes a sudden change in the judgment result is determined in the semantic vector space.

[0136] Each semantic unit in the semantically reconstructed text is vectorized and encoded. A pre-trained context-aware language model is used to generate a corresponding dense vector representation for each semantic unit. Specifically, the text content of each semantic unit undergoes preprocessing steps such as word segmentation and sub-word splitting before being input into the encoder to obtain a fixed-dimensional embedding vector. Since the embedding vectors of different semantic units come from different encoding scales, they need to be uniformly mapped to the same semantic vector space. Spatial alignment is achieved through a linear projection matrix, making the vector representations of each semantic unit comparable under the same metric system. Based on this, a holistic semantic vector space representation of the semantically reconstructed text is constructed according to the semantic relationships between semantic units—including co-occurrence relationships, contextual dependencies, and semantic similarity. This spatial representation not only preserves the semantic information of each semantic unit itself but also encodes the interaction structure between semantic units through relationships, forming a vectorized description that reflects the overall semantic distribution of the text.

[0137] In the semantic vector space, perturbation directions are selected by comprehensively considering the geometric structure of the semantic space and the distribution characteristics of the review decision boundary. The selection of perturbation directions is not random, but prioritizes sampling along the principal component directions with large variance in the semantic vector space, as well as directions related to the center directions of known risk semantic categories, to improve the targeting and efficiency of gradient detection. For each selected perturbation direction... (in (where the index number is the perturbation direction), and a perturbation vector sequence with increasing perturbation amplitude is generated in that direction. Let the step size of the perturbation amplitude be... , No. The perturbation vector of the step is ,in The perturbation steps are numbered, starting from 1 and increasing incrementally until the preset maximum perturbation amplitude is reached. By controlling The size of the perturbation vector can be flexibly adjusted between coarse-grained rapid positioning and fine-grained precise positioning, thereby achieving a balance between computational efficiency and positioning accuracy. The generated perturbation vector sequence is used as an adversarial perturbation vector for subsequent gradient detection.

[0138] The adversarial perturbation vectors are superimposed onto the semantic vector space representation to obtain the perturbed semantic vector space representation. Let the original semantic vector space representation be... , superimposed on the first The disturbance direction is the first After the perturbation vector of the step, the perturbation semantic vector space is represented as This perturbation operation simulates minor modifications to semantic content at the vector space level, enabling the review decision tree to re-execute judgments under different semantic states, thereby revealing the location information of the decision boundary. The perturbation superposition operation maintains the overall structure of the semantic vector space, introducing only controllable offsets in specific directions, thus avoiding significant damage to the text semantics.

[0139] Represent the perturbed semantic vector space The input is a review decision tree. Based on the judgment rules of each decision node, the review decision tree evaluates the perturbed semantic representation node by node, outputting the judgment result of each decision node under the current perturbed state. The judgment result is recorded in the form of binary labels, corresponding to the "risk" and "non-risk" states respectively. Since the amplitude of the adversarial perturbation vector gradually increases, the judgment result of the review decision tree remains stable under different perturbation amplitudes, but a state transition can also occur at a certain perturbation amplitude. The changes in the judgment result of each decision node are monitored in real time. When the judgment result of a decision node changes from a "non-risk" state to a "risk" state, or from a "risk" state to a "non-risk" state, the corresponding perturbation amplitude is recorded. (in (The perturbation step number and perturbation direction during the state transition) This state transition point is the critical position where the decision node's judgment result changes abruptly in the direction of the disturbance, and it has important semantic boundary localization significance.

[0140] Based on the recorded perturbation amplitude and the direction of disturbance In the semantic vector space, the critical semantic boundary that causes a sudden change in the judgment result is determined. The spatial coordinates of the critical semantic boundary can be represented as follows: That is, the original semantic vector space represents the position reached after shifting along the perturbation direction by the critical perturbation amplitude. The critical points are calculated separately for multiple perturbation directions. The data is then aggregated, and a complete critical semantic boundary hypersurface is constructed in the semantic vector space using fitting or interpolation methods. This hypersurface divides the semantic vector space into "risk regions" and "non-risk regions." The semantic units located near the critical semantic boundary are the most sensitive to the review decision results and are the key targets for subsequent compliance replacement operations.

[0141] To further improve the localization accuracy of critical semantic boundaries, after initial localization using coarse-grained perturbation, a binary search strategy can be employed to reduce the perturbation step size within the known state transition interval, thereby refining the estimation of the critical perturbation amplitude. Let the state transition interval obtained from coarse-grained localization be... Corresponding disturbance amplitude range Within this interval, with smaller step sizes Perform binary iterations until the estimation error of the disturbance amplitude is lower than the preset accuracy threshold. This refinement process significantly reduces boundary positioning errors caused by excessively large step sizes, ensuring the geometric accuracy of the critical semantic boundary in the semantic vector space and providing a reliable boundary reference for subsequent semantic unit recognition and compliance replacement.

[0142] Figure 4 This image presents a heatmap and schematic diagram of the critical semantic boundary and unit distribution. It is a two-dimensional projection heatmap of the semantic vector space, with color intensity representing the sensitivity of decision nodes to perturbations (red for high sensitivity, blue for low sensitivity). The black dashed line represents the fitted critical semantic boundary hypersurface. The scattered points represent the semantic unit vectors in the semantically reconstructed text, with star-shaped markers indicating units to be replaced near the critical boundary. This image reveals the distribution density of semantic units around the critical boundary and shows that most sensitive units are concentrated on both sides of the boundary. Compliant replacement words can be located by maintaining the direction of the topic vector through nearest neighbor search. This image visually verifies the effectiveness and spatial interpretability of identifying critical boundaries using adversarial perturbation combined with gradient detection.

[0143] In one optional implementation, the semantic units located at the critical semantic boundary are replaced by compliant replacement words determined through nearest neighbor search in the semantic vector space, while maintaining the direction of the semantically reconstructed text topic vector, to obtain compliant SMS content, including:

[0144] Extract semantic units located at the critical semantic boundary from the semantically reconstructed text, and mark the semantic units as semantic units to be replaced;

[0145] The vector representations of all semantic units in the semantically reconstructed text are weighted and aggregated to calculate the topic vector of the semantically reconstructed text, and the direction of the topic vector in the semantic vector space is determined as the topic constraint direction.

[0146] The vector representation of the semantic unit to be replaced is decomposed into components along the topic constraint direction and components perpendicular to the topic constraint direction.

[0147] In the semantic vector space, a nearest neighbor search is performed with the vector representation of the semantic unit to be replaced as the center. Candidate replacement word vectors with a distance less than a preset distance threshold from the vector representation of the semantic unit to be replaced are selected. The deviation value between the projection component of each candidate replacement word vector along the topic constraint direction and the component of the semantic unit to be replaced along the topic constraint direction is calculated.

[0148] Candidate replacement words corresponding to candidate replacement word vectors whose deviation values ​​are less than a preset deviation threshold are selected as compliant replacement words;

[0149] Replace the semantic unit to be replaced with the compliant replacement word, and update the semantically reconstructed text;

[0150] The updated semantically reconstructed text is then restored to its original form to obtain compliant SMS content.

[0151] After confirming the risk in the overall audit results, it is necessary to perform compliance replacement operations on the semantic units located at the critical semantic boundaries in the semantically reconstructed text. This involves iterating through the critical semantic boundary coordinates corresponding to all perturbation directions. The semantic units falling near each critical semantic boundary are extracted and marked as semantic units to be replaced. "Near a critical semantic boundary" means that the distance between the vector representation of the semantic unit and the coordinates of the corresponding critical boundary is less than a preset boundary proximity threshold. Semantic units meeting this condition are identified as key semantic components causing a sudden change in the decision node's judgment result and require compliance processing. During the marking process, if the same semantic unit is simultaneously covered by critical boundaries from multiple perturbation directions, the priority of that semantic unit to be replaced is increased accordingly, and it is processed first in subsequent replacement operations.

[0152] After extracting the semantic units to be replaced, it is necessary to calculate the overall topic vector of the semantically reconstructed text to maintain the semantic theme of the text during the replacement process. The vector representations of all semantic units in the semantically reconstructed text are weighted and aggregated. The weight of each semantic unit is determined by its semantic density value and positional importance in the text. Semantic units located in the core of the text and with higher semantic density contribute more to the topic vector. Let the first semantic unit be... The vector representation of a semantic unit is as follows The corresponding aggregation weight is Then the topic vector We obtain the following result after weighted summation and normalization: Normalized This is the unit vector of the topic constraint direction in the semantic vector space, and all subsequent replacement operations use this direction as the constraint reference.

[0153] After determining the topic constraint direction, an orthogonal decomposition is performed on the vector representation of each semantic unit to be replaced. Let the vector representation of the semantic unit to be replaced be... Decompose it into directions along the subject constraint. The amount and components perpendicular to the subject constraint direction The decomposition formula is: , .in This reflects the degree of alignment between the semantic unit and the text topic. This represents a semantic component that deviates from the topic direction. During the replacement process, the projected component of the compliant replacement word in the topic constraint direction is required to be consistent with... Maintain a high degree of consistency to ensure that the overall semantic theme of the replaced text does not change substantially.

[0154] Vector representation of the semantic unit to be replaced Centered on the semantic vector space, a nearest neighbor search is performed to filter out the nearest neighbors with the same as the target vector. The Euclidean distance between them is less than a preset distance threshold. All candidate replacement word vectors constitute the candidate replacement word set. The threshold value needs to strike a balance between semantic similarity and candidate word diversity. If the threshold is too small, the number of candidate words will be insufficient; if the threshold is too large, candidate words with significant semantic bias will be introduced. In practical applications, The value can be adaptively set based on the overall distribution density of the semantic vector space, for example, taking a multiple of the average nearest neighbor distance of word vectors in that space. For each candidate replacement word vector in the candidate set... Calculate its projection components along the subject constraint direction. Then, the projection component and the topic direction component of the semantic unit to be replaced are calculated. Deviation between : Deviation value The smaller the value, the higher the alignment between the candidate replacement word and the original semantic unit in the direction of topic constraints, and the smaller the impact on the text topic after replacement.

[0155] Filter all deviation values Less than the preset deviation threshold Candidate replacement words are selected as compliant candidate words that meet the topic constraints. If multiple compliant candidate words still exist after filtering by the deviation threshold, they are further sorted according to their distance from the semantic unit to be replaced in the overall vector space, and the one with the smallest distance is selected as the final compliant replacement word to ensure that the local semantics of the replaced semantic unit is as close as possible to the original content. If the filtering result is empty, that is, there are no candidate words that simultaneously meet the distance threshold and deviation threshold conditions, the deviation threshold is appropriately relaxed and a new search is performed, or the semantic unit is marked as not being able to be automatically replaced and submitted for manual review.

[0156] After identifying compliant replacement words, the corresponding semantic units to be replaced in the semantically reconstructed text are replaced with compliant replacement words, completing the local update of the semantically reconstructed text. If there are multiple semantic units to be replaced, the replacement operations are performed sequentially according to priority. After each replacement, the topic vector of the updated text is recalculated to dynamically adjust the topic constraint direction of subsequent replacement operations, avoiding topic shift caused by multiple replacements accumulating. After each round of replacement, the global semantic consistency of the updated semantically reconstructed text is checked to confirm that the offset of the topic vector direction does not exceed the allowable range. If it does, the current replacement operation is rolled back and the candidate word selection strategy is adjusted.

[0157] After all semantic units to be replaced have been replaced and passed consistency verification, a text restoration operation is performed on the updated semantically reconstructed text. Text restoration is the inverse process of multimodal deformation recognition. It maps the normalized semantic units in the semantically reconstructed text back to natural language expressions, restoring normal word boundaries, punctuation marks, and grammatical structures, while retaining the expression of non-sensitive content in the original SMS message. This ensures that the compliant SMS content conforms to normal communication habits in terms of readability and naturalness. During the restoration process, for locations in the original SMS message that contain homophone substitutions, character deformations, or other deformation methods, the standard written form of the compliant replacement words is output, and the original form is not restored. The final output compliant SMS content eliminates the risky semantic components in the original SMS message while maintaining the overall semantic coherence and thematic consistency of the text, and can be directly used for subsequent SMS sending or archiving.

[0158] A second aspect of this invention provides a multi-level SMS content review and intelligent sensitive word filtering system, comprising:

[0159] The multimodal recognition unit is used to perform multimodal deformation recognition on the SMS messages to be reviewed, and obtain semantically reconstructed text.

[0160] The semantic reconstruction unit is used to construct an audit decision tree based on the semantic density distribution features of the semantically reconstructed text and the historical reputation weight of the SMS sender.

[0161] A decision tree construction unit is used to calculate the risk propagation coefficient of each semantic unit in the semantically reconstructed text. In the review decision tree, a corresponding decision node is matched for each semantic unit. The review level corresponding to the current semantic unit is dynamically determined along the branch path of the review decision tree, and the review path of the SMS to be reviewed is generated.

[0162] The risk propagation unit is used to perform corresponding audit operations according to the audit path, output the local risk label and credibility measurement value of the semantic unit, and summarize the global audit results after traversing all decision nodes of the audit decision tree.

[0163] The review execution unit is used to construct the semantic vector space representation of the semantically reconstructed text when the global review result is risky, generate multiple adversarial perturbation vectors, perform gradient detection along the direction of the adversarial perturbation vectors, and identify the critical semantic boundary that causes the decision result of the decision node to change abruptly.

[0164] A critical semantic unit is used to determine compliant replacement words and replace semantic units located at the critical semantic boundary by means of nearest neighbor search in the semantic vector space, while keeping the direction of the semantically reconstructed text topic vector unchanged, so as to obtain compliant SMS content.

[0165] A third aspect of the present invention provides an electronic device, comprising:

[0166] processor;

[0167] Memory used to store processor-executable instructions;

[0168] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0169] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0170] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for multi-level review and intelligent sensitive word filtering of SMS content, characterized in that, include: Multimodal deformation recognition is performed on the SMS messages to be reviewed to obtain semantically reconstructed text; Based on the semantic density distribution features of the semantically reconstructed text and the historical reputation weight of the SMS sender, an audit decision tree is constructed. Calculate the risk propagation coefficient of each semantic unit in the semantically reconstructed text, match a corresponding decision node for each semantic unit in the review decision tree, dynamically determine the review level corresponding to the current semantic unit along the branch path of the review decision tree, and generate the review path of the SMS to be reviewed; Perform the corresponding audit operations according to the audit path, output the local risk label and credibility metric value of the semantic unit, and summarize the results after traversing all decision nodes of the audit decision tree to obtain the global audit result; When the global review result is risky, a semantic vector space representation of the semantically reconstructed text is constructed, multiple adversarial perturbation vectors are generated, gradient detection is performed along the direction of the adversarial perturbation vectors, and the critical semantic boundary that causes the decision result of the decision node to change abruptly is identified; Under the constraint of keeping the direction of the semantically reconstructed text topic vector unchanged, the semantic units located at the critical semantic boundary are used to determine compliant replacement words and replace them through nearest neighbor search in the semantic vector space to obtain compliant SMS content.

2. The method according to claim 1, characterized in that, Multimodal deformation recognition is performed on the SMS messages to be reviewed to obtain semantically reconstructed text, including: The stroke connection relationship of each character is mapped to a topological graph structure. By calculating the isomorphic similarity between the topological graph structure of the target character in the SMS to be reviewed and the topological graph structure of each standard character in the standard character library, the glyph deformation characters formed by stroke deformation, component splitting or symbol substitution are identified, and the glyph deformation characters are restored to the corresponding glyph restoration standard characters. Extract the initial consonant features, final vowel features, and tone features of each character to construct a phonological feature vector. By calculating the distance metric between the phonological feature vector of the target character in the SMS to be reviewed and the phonological feature vector of each standard character in the standard character library, identify phonologically deformed characters formed by homophonic substitution or mixed pinyin, and restore the phonologically deformed characters to the corresponding phonologically restored standard characters. Based on the combination of the glyph restoration standard characters and the phonological restoration standard characters to form the restored text, semantic field migration analysis is performed on the restored text to extract the semantic field attribution identifier of each lexical unit in the restored text, identify lexical units that have undergone abnormal semantic field migration, and restore the lexical units that have undergone semantic field migration to the standard lexical units in the original semantic field through the semantic field backtracking mechanism, thus obtaining the semantically reconstructed text.

3. The method according to claim 1, characterized in that, Based on the semantic density distribution features of the semantically reconstructed text and the historical reputation weight of the SMS sender, a review decision tree is constructed, including: The semantically reconstructed text is segmented into semantic units. The ratio of the number of key semantic components within a semantic unit to the total length of the semantic unit is calculated to obtain the semantic density value of each semantic unit. The semantic density values ​​of all semantic units are arranged in order of position to form a semantic density distribution sequence. The semantic density distribution sequence is subjected to distribution pattern recognition, and the peak position, fluctuation amplitude and gradient change rate of the semantic density distribution sequence are extracted and combined to form semantic density distribution features; Obtain the historical sending records of the SMS sender, calculate the ratio of the number of illegal SMS messages in the historical sending records to the total number of SMS messages sent, determine the corresponding reputation decay coefficient based on the numerical range of the ratio, and determine the historical reputation weight of the SMS sender based on the reputation decay coefficient. The semantic density distribution feature is used as the content risk decision dimension, and the historical reputation weight is used as the sender reputation decision dimension. Multiple decision segmentation surfaces are set in the decision space formed by the content risk decision dimension and the sender reputation decision dimension. Each decision segmentation surface divides the decision space into different decision regions. Each decision region corresponds to a different review level configuration. The mapping relationship between the decision segmentation surface and the decision region is constructed into a review decision tree.

4. The method according to claim 1, characterized in that, Calculate the risk propagation coefficient of each semantic unit in the semantically reconstructed text, match a corresponding decision node for each semantic unit in the review decision tree, dynamically determine the review level corresponding to the current semantic unit along the branch path of the review decision tree, and generate the review path of the SMS to be reviewed, including: Dependency relations are parsed for each semantic unit in the semantically reconstructed text. Syntactic dependency edges and semantic dependency edges between each semantic unit and other semantic units are identified, and the number of dependency edge connections is counted. The dependency correlation degree of the semantic unit is calculated based on the number of dependency edge connections. Within a preset sliding window, the frequency of co-occurrence of each semantic unit and risk semantic units is counted, and the risk co-occurrence intensity of the semantic units is calculated based on the frequency of co-occurrence. The risk propagation coefficient of each semantic unit is obtained by weighted fusion of the dependency correlation degree and the risk co-occurrence strength. The semantic density distribution features and the risk propagation coefficient of each semantic unit are used as decision input features, and a corresponding decision node is matched for each semantic unit in the review decision tree; Obtain the review level configuration corresponding to the decision region where the decision node is located, and parse it into a review level sequence and review granularity parameters; According to the positional order of each semantic unit, the review level sequence and review granularity parameters corresponding to each semantic unit are sequentially concatenated to form the review path of the SMS to be reviewed.

5. The method according to claim 1, characterized in that, Perform the corresponding audit operations according to the audit path, output the local risk label and credibility metric value of the semantic unit, and summarize the results after traversing all decision nodes of the audit decision tree to obtain the global audit result, including: According to the positional order of each semantic unit in the review path, the decision node and review level sequence corresponding to each semantic unit are obtained sequentially. At each decision node, the type of review operation to be performed for the current semantic unit is determined according to the review level sequence. Risk features are extracted from the text content of the semantic unit. The extracted risk features are matched and calculated with a preset risk feature template. Based on the matching result, a local risk label for the semantic unit is generated. The confidence level of the local risk label is evaluated, and the number and matching strength of the risk features matched during the generation of the local risk label are counted. The confidence metric of the local risk label is calculated based on the number of risk features and the matching strength. The local risk label and the trust metric value are associated with the location information of the semantic unit and stored to form the audit record of the semantic unit; Traverse all decision nodes of the review decision tree and collect the review records of all semantic units; Aggregate and analyze the local risk tags in all collected review records. Based on the distribution pattern of the local risk tags and the statistical characteristics of the credibility metric, determine the overall risk level and risk type of the SMS to be reviewed, and generate the global review result.

6. The method according to claim 1, characterized in that, Construct a semantic vector space representation of the semantically reconstructed text, generate multiple adversarial perturbation vectors, perform gradient probing along the directions of the adversarial perturbation vectors, and identify critical semantic boundaries that cause abrupt changes in the decision results of the decision nodes, including: Each semantic unit in the semantically reconstructed text is vectorized and encoded, and the vector representation of each semantic unit is mapped to a unified semantic vector space. The semantic vector space representation of the semantically reconstructed text is constructed based on the semantic relationships between each semantic unit. Multiple perturbation directions are selected in the semantic vector space, and a perturbation vector sequence with increasing perturbation amplitude is generated in each perturbation direction. The perturbation vector sequence is used as an adversarial perturbation vector. The adversarial perturbation vector is superimposed on the semantic vector space representation to obtain the perturbed semantic vector space representation. The perturbated semantic vector space representation is input into the review decision tree to obtain the judgment results of each decision node in the review decision tree; The perturbation amplitude is gradually increased along the direction of the adversarial perturbation vector, and the change of the decision result of the decision node is monitored in real time. When the decision result changes from a non-risk state to a risk state or from a risk state to a non-risk state, the perturbation amplitude and perturbation direction at the time of the state change are recorded. Based on the recorded perturbation amplitude and perturbation direction, a critical semantic boundary that causes a sudden change in the judgment result is determined in the semantic vector space.

7. The method according to claim 1, characterized in that, Under the constraint of keeping the direction of the semantically reconstructed text topic vector unchanged, the semantic units located at the critical semantic boundary are used to determine and replace compliant replacement words through nearest neighbor search in the semantic vector space to obtain compliant SMS content, including: Extract semantic units located at the critical semantic boundary from the semantically reconstructed text, and mark the semantic units as semantic units to be replaced; The vector representations of all semantic units in the semantically reconstructed text are weighted and aggregated to calculate the topic vector of the semantically reconstructed text, and the direction of the topic vector in the semantic vector space is determined as the topic constraint direction. The vector representation of the semantic unit to be replaced is decomposed into components along the topic constraint direction and components perpendicular to the topic constraint direction. In the semantic vector space, a nearest neighbor search is performed with the vector representation of the semantic unit to be replaced as the center. Candidate replacement word vectors with a distance less than a preset distance threshold from the vector representation of the semantic unit to be replaced are selected. The deviation value between the projection component of each candidate replacement word vector along the topic constraint direction and the component of the semantic unit to be replaced along the topic constraint direction is calculated. Candidate replacement words corresponding to candidate replacement word vectors whose deviation values ​​are less than a preset deviation threshold are selected as compliant replacement words; Replace the semantic unit to be replaced with the compliant replacement word, and update the semantically reconstructed text; The updated semantically reconstructed text is then restored to obtain compliant SMS content.

8. A multi-level SMS content review and intelligent sensitive word filtering system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The multimodal recognition unit is used to perform multimodal deformation recognition on the SMS messages to be reviewed, and obtain semantically reconstructed text. The semantic reconstruction unit is used to construct an audit decision tree based on the semantic density distribution features of the semantically reconstructed text and the historical reputation weight of the SMS sender. A decision tree construction unit is used to calculate the risk propagation coefficient of each semantic unit in the semantically reconstructed text. In the review decision tree, a corresponding decision node is matched for each semantic unit. The review level corresponding to the current semantic unit is dynamically determined along the branch path of the review decision tree, and the review path of the SMS to be reviewed is generated. The risk propagation unit is used to perform corresponding audit operations according to the audit path, output the local risk label and credibility measurement value of the semantic unit, and summarize the global audit results after traversing all decision nodes of the audit decision tree. The review execution unit is used to construct the semantic vector space representation of the semantically reconstructed text when the global review result is risky, generate multiple adversarial perturbation vectors, perform gradient detection along the direction of the adversarial perturbation vectors, and identify the critical semantic boundary that causes the decision result of the decision node to change abruptly. A critical semantic unit is used to determine compliant replacement words and replace semantic units located at the critical semantic boundary by means of nearest neighbor search in the semantic vector space, while keeping the direction of the semantically reconstructed text topic vector unchanged, so as to obtain compliant SMS content.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images