Method and system for monitoring compliance of marketing mobile terminal dynamic behavior

Abstract

The application discloses a marketing mobile terminal dynamic behavior compliance monitoring method and system, belongs to the technical field of information transmission monitoring, and aims to solve the problems of insufficient non-text information processing, sensitive word determination redundancy, incomplete coverage of illegal expression variants and low data reuse rate of existing monitoring. The method comprises the following steps: capturing dynamic behavior information, constructing a composite retrieval code retrieval history library, and judging whether sensitive word identification has been performed; if not, a standardized word segmentation sequence is generated; word segmentation comparison is performed to determine matched word segmentation and unmatched word segmentation, the matched word segmentation history synonym set is called, the unmatched word segmentation core semantic screening avoidance expression is positioned, the synonym set thereof is generated, and the synonym data set is constructed; the matched word segmentation history judgment result is called, the synonym sequence algorithm probability is established, the sensitive word is judged in combination with the significant proportion, the processing flow is executed, and the history library and the word segmentation library are updated. The application realizes efficient and accurate compliance monitoring, improves the data reuse rate, and guarantees the effectiveness of the monitoring closed loop.

Description

Technical Field

[0001] This invention relates to the field of information transmission monitoring technology, and more specifically to a method and system for monitoring the compliance of dynamic behavior of marketing mobile terminals. Background Technology

[0002] With the rapid development of the mobile internet, marketing mobile terminals have become the core carrier for enterprises to reach users. Whether the dynamic behavioral information they send complies with regulatory standards is directly related to the protection of consumer rights and the maintenance of market competition order. Therefore, the need for compliance monitoring of such information is becoming increasingly urgent.

[0003] However, current technologies for monitoring the dynamic behavior compliance of marketing mobile terminals still have many key shortcomings, making it difficult to meet actual monitoring needs. Existing monitoring methods often perform redundant processing on the same or repeated information multiple times, which not only consumes a lot of computing resources but also leads to slow monitoring response speeds, making it difficult to adapt to the rapid compliance verification requirements in real-time scenarios. Furthermore, violations often exist in flexible variations, and existing monitoring technologies mostly rely on fixed sensitive word libraries for mechanical matching, resulting in limited coverage of these variations and difficulty in capturing complex and ever-changing violation expressions, easily leading to missed detections. In addition, the judgment results and processing records generated during the monitoring process are not effectively linked and updated with relevant databases. When encountering the same or similar dynamic behavior information later, the entire monitoring process still needs to be re-executed, preventing data reuse and further exacerbating the problem of low monitoring efficiency. It also makes it difficult to form a coherent compliance monitoring data link, failing to provide effective support for long-term compliance management. Therefore, to overcome these limitations, this invention proposes a method and system for monitoring the dynamic behavior compliance of marketing mobile terminals. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for monitoring the compliance of dynamic behavior on marketing mobile terminals. It solves the problems of how to reduce repetitive operations on processed information to improve efficiency, effectively process non-text dynamic behavior information so that it can be used for compliance analysis, comprehensively identify various variants of illegal expressions in target scenarios to avoid missed judgments, and accurately judge sensitive content in dynamic behavior information to reduce the risk of misjudgment.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Methods for monitoring the compliance of dynamic behavior on marketing mobile devices include:

[0007] The system captures dynamic behavioral information sent by marketing mobile terminals, performs matching and retrieval in the historical information database by constructing composite retrieval codes, determines whether the current dynamic behavioral information has been identified as sensitive words, and if not, generates a standardized word segmentation sequence of dynamic behavioral information through text processing to construct a text dataset of information interaction behavior of marketing mobile terminals.

[0008] By comparing and identifying matched and unmatched words in a standardized word segmentation sequence, a set of historical synonyms for matched words is retrieved from a word segmentation verification library. By locating the core semantics of unmatched words, avoidance expressions are filtered out, and a set of synonyms for unmatched words is generated to construct a synonym dataset for the standardized word segmentation sequence.

[0009] For matched word segments, retrieve the historical sensitive word judgment results. If the word is determined to be non-sensitive, merge it with the unmatched word segments to form a word segment to be judged. Establish a synonym sequence for the word segment to be judged, and calculate the sensitive word probability of each synonym in the synonym sequence through bidirectional matching. Combine the significant probability ratio to obtain the sensitive word probability of each word segment to be judged, so as to determine whether the word segment to be judged is a sensitive word.

[0010] Based on the sensitive word determination results, execute the sensitive word processing flow and update the historical information database and word segmentation verification database.

[0011] Specifically, the steps for constructing a text dataset of information interaction behavior on marketing mobile terminals include:

[0012] Capture dynamic behavioral information sent by marketing mobile terminals, classify and store the information according to its type in the raw information pool, and bind it to the metadata fields of the marketing mobile terminals.

[0013] Metadata fields are structured data fields used to uniquely identify the associated subject, interaction scenario, and processing status of dynamic behavior information;

[0014] For dynamic behavior information stored in the original information pool, the information feature value comparison method is used to compare the information with the historical information database to determine whether the current dynamic behavior information has been identified for sensitive words.

[0015] The historical information database refers to a structured database that stores dynamic behavioral information of all marketing mobile terminals in the current target scenario and the results of compliance processing of associated sensitive words;

[0016] If it is determined that sensitive word identification has been performed, the sensitive word identification results and sensitive word compliance processing methods corresponding to the matching historical dynamic behavior information are retrieved through the associated index of the historical information database.

[0017] If it is determined that no sensitive word recognition has been performed, the current dynamic behavior information and its metadata fields will be stored in the historical information database, and text processing will be started, that is, the dynamic behavior information will be converted into text according to information type and scenario.

[0018] Specifically, the steps for constructing a text dataset of information interaction behaviors on marketing mobile terminals also include:

[0019] The original text information that was not identified for sensitive words was extracted and the corresponding non-text text generated by textualization were summarized to form a set of texts to be standardized.

[0020] Perform text consistency checks and format standardization on the standardized text set to obtain a standardized text set;

[0021] Configure a behavior-specific thesaurus. A behavior-specific thesaurus is a structured thesaurus that is customized for the current target scenario and contains scenario-specific terms and compliance boundary words. It consists of core scenario terms, industry compliance sensitive boundary words, and an exclusion list of invalid function words.

[0022] Based on a behavior-specific lexicon, a standardized text set is segmented to generate a standardized segmentation sequence.

[0023] By structurally associating standardized text sets, standardized word segmentation sequences, and corresponding metadata fields, a marketing mobile terminal information interaction behavior text dataset is constructed.

[0024] Specifically, the steps for determining whether the current dynamic behavior information has undergone sensitive word identification include:

[0025] For different information types of dynamic behavior information, an appropriate feature extraction algorithm is adopted and multimodal feature normalization is performed to obtain the unique feature values ​​of the current dynamic behavior information.

[0026] Extract the core retrieval factors of the current dynamic behavior information, including the target scene identifier, time partition code, terminal account hash fragment, and device identifier hash fragment; combine them with the exclusive feature values, concatenate them in a fixed order to form an initial retrieval string, and compress the initial retrieval string through a hash compression algorithm to generate a composite retrieval code of fixed length;

[0027] A two-tiered index system consisting of a composite retrieval code main index and dimension auxiliary indexes is constructed in the historical information database. The main index adopts a B+ tree structure and uses the composite retrieval code as the index key. The dimension auxiliary indexes include a joint index of target scene identifier, time partition, terminal account hash fragment, and device identifier hash fragment.

[0028] During retrieval, the physical storage partition is located through the dimension-assisted index, and a subset of historical records is constructed. Within the subset of historical records, matching is performed through the composite retrieval encoded main index. If a potential matching record is found, it is determined that the current dynamic behavior information has been identified as a sensitive word; otherwise, it is determined that the current dynamic behavior information has not been identified as a sensitive word.

[0029] Specifically, the steps for constructing a standardized word segmentation sequence thesaurus include:

[0030] From the text dataset, retrieve the standardized word segmentation sequence. For each word in the standardized word segmentation sequence, perform a search and match in the word segmentation verification library based on the hash value of the word segmentation string to identify the matched and unmatched words.

[0031] For the matched words in the standardized word segmentation sequence, retrieve the historical synonym set from the word segmentation verification library; and match the synonyms of the unmatched words in the standardized word segmentation sequence to establish a synonym set, which is then added to the word segmentation verification library.

[0032] The steps to build a thesaurus include:

[0033] By using a semantic matching algorithm, unmatched word segments are compared with core terms in the scenario, and semantic boundaries are defined by combining compliance-sensitive boundary words to locate the core semantics of unmatched word segments.

[0034] The evasion expression corpus is invoked. The evasion expression corpus is a structured database that stores variants of illegal expressions in the target scenario. It consists of homophonic substitution expressions, character rewriting expressions, and metaphorical association expressions.

[0035] Calculate the semantic similarity between the core semantics of the unmatched word segmentation and the evasion expressions in the corpus, and filter the evasion expressions based on the semantic similarity and include them in the synonym set;

[0036] Retrieve the text context of the unmatched word segment, parse the grammatical dependency relationship between the unmatched word segment and the text context, extract contextual collocation words and variant expressions, and add them to the synonym set;

[0037] By integrating the historical synonym set of matched words with the synonym set of unmatched words, a synonym dataset for each word in the standardized word segmentation sequence is formed.

[0038] Specifically, the steps to locate the core semantics of unmatched word segments include:

[0039] Based on the target scenario in the metadata associated with the unmatched word segmentation, call the behavior-specific word sub-library of the corresponding target scenario tag in the behavior-specific word library;

[0040] Extract the core terminology set for specific behaviors and the set of industry compliance-sensitive boundary words;

[0041] Standardization processing is performed on unmatched word segments, the core terminology set of the scenario, and the industry compliance-sensitive boundary word set, and they are converted into semantic vectors of fixed dimensions.

[0042] Calculate the similarity value between the semantic vector of the unmatched word segmentation and the semantic vector of each scene's core term; set a similarity threshold to filter scene core terms and form a list of candidate related terms;

[0043] The semantic vectors of unmatched words are compared with the semantic vectors of the industry compliance sensitive boundary word set. The intersection ratio of the semantic vectors of unmatched words and the semantic vectors of industry compliance sensitive boundary words is calculated to determine whether there is a non-compliant semantic association between the unmatched words. If so, the semantic boundary is defined.

[0044] Based on the similarity values ​​of each candidate related term in the candidate related term list, normalization processing is performed, and a weighted average algorithm is used to fuse the semantic vectors of each candidate related term to generate the core semantic vector of the unmatched word segmentation. Combined with the semantic boundary, a structured core semantic is generated.

[0045] Specifically, the steps for determining whether a word segment to be judged is a sensitive word include:

[0046] The expression collection sub-library is invoked, and the word segmentation verification library is retrieved based on the matched word segmentation. The historical sensitive word judgment results of the matched word segmentation are retrieved: if it is judged as a sensitive word, the sensitive word processing process is triggered; otherwise, it is marked as a word segmentation to be judged, its historical synonym set is marked as a historical set to be optimized, and the unmatched word segmentation is marked as a word segmentation to be judged, and the synonym set of the unmatched word segmentation is marked as a historical set to be optimized.

[0047] For each historical set to be optimized, filter and add avoidable expressions with similarity higher than the preset similarity threshold in the expression collection sub-library, and construct an optimized synonym set;

[0048] Based on the synonyms of each optimized synonym set, a synonym sequence is constructed, and deduplication is performed. A unique code is assigned to each synonym in the synonym sequence, and the probability of sensitive words for each synonym in the synonym sequence is calculated.

[0049] By mapping the unique encoding of synonyms to the words to be judged, the probability set of sensitive words corresponding to each word segment is extracted and bound;

[0050] Calculate the basic statistics of the probability distribution of sensitive words, and count the number of elements in the sensitive word probability set whose sensitive word probability is greater than the preset significant probability threshold, so as to calculate the proportion of significant probability.

[0051] When the proportion of significant probability is less than the preset proportion threshold, the median of the sensitive word probability set is used as the sensitive word probability of the word to be segmented; otherwise, the maximum value of the sensitive word probability set is used as the sensitive word probability of the word to be segmented.

[0052] Configure a probability threshold. If the probability of a sensitive word in the segment to be judged is greater than the probability threshold, then the segment to be judged is determined to be a sensitive word; otherwise, the segment to be judged is determined to be a non-sensitive word.

[0053] Specifically, the steps for calculating the probability of sensitive words for each synonym in a synonym sequence include:

[0054] Configure a sensitive word database and perform bidirectional matching between each synonym in the synonym sequence and the preset sensitive word database. This includes: if the synonym matches a sensitive word in the sensitive word database, the sensitive word probability of the synonym is set to 1; otherwise, through fuzzy matching, the similarity of the character shape, homophony, and semantics between the synonym and each sensitive word in the sensitive word database is calculated.

[0055] Character shape similarity refers to the degree of similarity between two word segments in terms of stroke structure and component composition, which is calculated by an edit distance algorithm based on stroke features;

[0056] Homophonic similarity refers to the degree of similarity in pronunciation between two word segments, which is calculated using a pinyin string similarity algorithm;

[0057] Semantic similarity refers to the degree of overlap in the semantic connotations of two word segments in a target scene, which is obtained by calculating the cosine similarity of the semantic vectors of the word segments;

[0058] The maximum value of the similarity in character shape, homophony, and semantic similarity between the synonym and sensitive word database is selected as the probability of the synonym being a sensitive word.

[0059] Specifically, the steps of executing the sensitive word processing flow based on the sensitive word determination results and updating the historical information database and word segmentation verification database include:

[0060] If the word to be judged is determined to be a sensitive word, the sensitive word handling process will be initiated, the sensitive word and its corresponding dynamic behavior information will be marked as a violation, and a real-time interception mechanism will be initiated for the marketing mobile terminal; otherwise, no action will be taken.

[0061] Record the execution information of the sensitive word handling process and build a structured violation processing log;

[0062] By using the unique feature values ​​of dynamic behavior information associated with the words to be judged, the historical information database is retrieved, the corresponding dynamic behavior information records are located, and the sensitive word recognition results, processing status, and processing log index are updated.

[0063] By using the hash value of the segmented string to be judged, the corresponding segmented entry in the segmentation verification library is retrieved, and the sensitive word status, judgment time, and sensitive word probability of the segmented entry are updated; and the processing log index is associated.

[0064] The marketing mobile terminal dynamic behavior compliance monitoring system includes an information capture module, a synonym matching module, a sensitive word determination module, and a sensitive word processing module.

[0065] The information capture module captures dynamic behavioral information sent by marketing mobile terminals, generates standardized word segmentation sequences of dynamic behavioral information through text processing, and constructs a text dataset of information interaction behavior of marketing mobile terminals; the synonym matching module constructs a synonym dataset of standardized word segmentation sequences; the sensitive word determination module calculates the sensitive word probability of each synonym in the synonym sequence, and obtains the sensitive word probability of each word to be judged by combining the significant probability ratio, so as to determine whether the word to be judged is a sensitive word; the sensitive word processing module executes the sensitive word processing process according to the sensitive word determination results, and updates the historical information database and word segmentation verification database.

[0066] The beneficial effects of this invention are:

[0067] This application effectively reduces resource consumption from repetitive processing and significantly improves monitoring efficiency by capturing real-time dynamic behavioral information of all types of marketing mobile terminals and reusing historical sensitive word identification results through an information comparison mechanism. It transforms non-textual information into structured, analyzable text through multi-technology fusion, eliminating monitoring blind spots for non-textual information and ensuring that all types of information are included in the compliance monitoring scope. It constructs and optimizes a thesaurus based on a behavior-specific lexicon, a corpus of avoidance expressions, and a sub-library of expression collection, comprehensively covering variants of illegal expressions in the target scenario and providing sufficient semantic support for sensitive word identification. It determines sensitive words through bidirectional matching combined with the probability distribution characteristics of synonyms, balancing scenario adaptability and judgment accuracy, effectively reducing the risk of misjudgment and missed judgment. Finally, it updates the historical information database and word segmentation verification database through feedback, achieving long-term reuse of monitoring data and forming a complete monitoring closed loop from information capture, analysis and judgment to data update, ensuring the continuity and accuracy of subsequent monitoring and providing reliable support for the compliance monitoring of dynamic behavior of marketing mobile terminals. Attached Figure Description

[0068] Figure 1 This is a flowchart of the marketing mobile terminal dynamic behavior compliance monitoring method of the present invention;

[0069] Figure 2 A flowchart illustrating the construction of a text dataset of information interaction behaviors for a marketing mobile terminal for this invention;

[0070] Figure 3 This is a flowchart illustrating the process of determining whether sensitive word identification has been performed on the current dynamic behavior information according to the present invention.

[0071] Figure 4 A flowchart illustrating the construction of a synonym dataset for a standardized word segmentation sequence in this invention;

[0072] Figure 5 This is a flowchart for determining whether a word segment to be judged is a sensitive word according to the present invention. Detailed Implementation

[0073] Please see Figure 1 This embodiment introduces a method for monitoring the compliance of dynamic behavior of marketing mobile terminals, including:

[0074] Step S1: Real-time capture of dynamic behavioral information sent by marketing mobile terminals, including text information, image files, video clips, link content, and terminal operation behavior logs. The dynamic behavioral information is compared to determine if it has been verified. If verified, the sensitive word recognition results of the matching historical dynamic behavioral information are obtained. If not, the non-text information in the dynamic behavioral information is converted into text format using multiple technologies and standardized to construct a text dataset of information interaction behavior of the marketing mobile terminal. The marketing mobile terminal refers to a mobile device used by enterprises to conduct targeted activities and reach target users. It serves as the core carrier for information interaction between enterprises and users, sending various dynamic behavioral information such as text information, image files, video clips, and links to target users to achieve goals such as information promotion, user reach, and product or service promotion. Specifically, text information collection covers the text content sent by the marketing mobile terminal; image and video information collection synchronously records file metadata fields, and video captures keyframes at fixed time intervals; link information collection automatically extracts addresses and preloads page content; operation behavior logs are associated with information interaction actions and terminal accounts and device identifiers to form a traceability link. Building upon this foundation, non-textual information is transformed into text using a multi-technology fusion approach: images are extracted using optical character recognition (OCR) to extract explicit text, and combined with image semantic analysis to identify homophones and metaphorical patterns, which are then converted into descriptive text; videos are extracted using speech-to-text conversion to extract audio content, supplemented with keyframe image recognition results to form complete text; links are parsed to extract core text from page structures and eliminate redundant information. Finally, all text undergoes format standardization, with unified encoding and word segmentation rules to ensure consistent formatting.

[0075] In this embodiment, real-time capture of all types of information comprehensively covers information sent by marketing mobile terminals, synchronously linking terminal accounts and device identifiers to form a complete traceability link, avoiding blind spots in information collection and providing data support for tracing the source of violations. Utilizing an information comparison mechanism, it quickly identifies repetitive dynamic behavior information from multiple terminals, reuses historical sensitive word identification results, reduces redundant processing, and improves monitoring efficiency. Through the integration of multiple technologies such as optical character recognition, image semantic analysis, and speech-to-text, it overcomes the analysis bottleneck of non-textual information, transforming it into analyzable descriptive text, ensuring that all types of information are included in a unified monitoring system. Standardized processing using unified encoding and word segmentation rules eliminates text format differences, constructing a structured information interaction behavior text dataset, providing a consistent data foundation for subsequent word segmentation comparison, sensitive word identification, and other stages, ensuring a coherent monitoring process and accurate analysis.

[0076] Please see Figure 2 Preferably, the steps for constructing a text dataset of information interaction behaviors on marketing mobile terminals include:

[0077] By configuring the application interface of the marketing mobile terminal, dynamic behavior information sent by the marketing mobile terminal is captured. The information is classified and stored in the raw information pool according to the information type of the dynamic behavior information, and bound to the metadata field of the marketing mobile terminal. The metadata field refers to the structured data field used to uniquely identify the associated subject, interaction scenario and processing status of the dynamic behavior information, including terminal account, device identifier, sending time and information source application.

[0078] For dynamic behavior information stored in the original information pool, the information feature value comparison method is used to compare the information with the historical information database to determine whether the current dynamic behavior information has been identified as sensitive words. The historical information database refers to a structured database that stores all processed and deduplicated dynamic behavior information of marketing mobile terminals in the current target scenario and the results of associated sensitive word compliance processing. It consists of dynamic behavior information feature values, sensitive word identification results, sensitive word compliance processing methods, and associated metadata fields.

[0079] If sensitive word identification is determined, the system retrieves the corresponding sensitive word identification results and compliant handling methods for the matched historical dynamic behavior information through the associated index in the historical information database. If sensitive word identification is determined not to have been performed, the system stores the current dynamic behavior information and its metadata fields in the historical information database and initiates text processing, i.e., performs text conversion on the dynamic behavior information according to information type and scenario, including:

[0080] When processing image files, first extract the explicit text in the image through optical character recognition technology, and then use an image semantic analysis model based on a convolutional neural network to identify homophonic images and metaphorical patterns in the image, and convert visual information into descriptive text. When processing video clips, first convert the video audio into audio text through automatic speech recognition technology, then intercept key frames of the video at fixed time intervals, perform image recognition on the key frames, extract the text and semantic information in the pictures, and supplement them to the audio text to form a complete video text. When processing link content, parse the HTML structure corresponding to the link through a page parsing engine, extract the core text in the page, and then use a redundancy filtering algorithm to剔除 page codes, advertising plug-ins, and irrelevant navigation text, and retain the valid content to form link text. During the conversion process, establish a quality verification mechanism to check the semantic integrity of the converted text to ensure the conversion effect of non-text information.

[0081] Extract the original text information that has not undergone sensitive word recognition, that is, the text information content directly sent by the marketing mobile terminal, such as SMS copywriting and social software push话术, and summarize it with the non-text corresponding text generated by text processing to form a set of texts to be standardized.

[0082] Perform text consistency verification on the set of texts to be standardized. Exclude duplicate texts caused by capture errors through semantic similarity comparison, and screen out garbled and null texts through encoding format detection and perform supplementary recording or deletion to ensure the data validity of the set of texts to be standardized.

[0083] Perform format standardization processing on the set of texts to be standardized that has undergone text consistency verification, unify the encoding format of the texts, and eliminate encoding incompatibility problems caused by terminal application differences to obtain a set of standardized texts.

[0084] Configure a behavior-specific word library. The behavior-specific word library refers to a structured word library customized for the current target scenario and containing scenario-specific terms and compliance boundary vocabulary. The target scenarios include financial product promotion and fast-moving consumer goods promotion. It consists of scenario core terms, such as the yield rate, risk reminder, and minimum investment amount in the financial scenario; industry compliance sensitive boundary words, such as the prefixes and suffixes of information expressions that need to be key monitored, such as variants of absolute and most-like words, and a list of excluded meaningless function words, such as meaningless modal particles and auxiliary words like "de", "le", "o", "ne", etc., which is used to provide a basis for scenario adaptation for subsequent word segmentation processing, avoid mis-segmentation of core vocabulary by general word segmentation tools, and ensure that the word segmentation results meet the semantic analysis requirements of the current target scenario. Perform word segmentation processing on the set of standardized texts based on the behavior-specific word library, retain the core vocabulary, exclude meaningless function words, and at the same time correct word segmentation deviations through semantic verification of the word segmentation results, such as avoiding unreasonable combinations like splitting "time-limited discount" into "time-limited", "you", "hui", etc., and generate a standardized word segmentation sequence.

[0085] The standardized text set, standardized word segmentation sequence, and corresponding metadata fields are structurally associated, and all three-dimensional data entries are integrated in a dual-dimensional manner: in the time dimension, the data entries are sorted according to the chronological order of information sending time to ensure that the data entries can be traced back to the interaction sequence; in the terminal dimension, the data entries are grouped by terminal account, and all text data entries of the same marketing mobile terminal are collected, ultimately constructing a marketing mobile terminal information interaction behavior text dataset; to improve the retrieval and subsequent analysis efficiency of the dataset, an inverted index technology is used to construct a multi-dimensional index system for the text dataset: a terminal index is built based on the terminal account, a time index is built based on the sending time, and a word segmentation index is built based on the standardized word segmentation sequence; a caching mechanism is used to store frequently accessed index entries to ensure the retrieval response efficiency of the dataset in subsequent sensitive word identification, word segmentation comparison, and other stages.

[0086] Please see Figure 3 Preferably, the steps for determining whether the current dynamic behavior information has been identified as having sensitive words include:

[0087] For different information types of dynamic behavior information, an appropriate feature extraction algorithm is adopted, and multimodal feature normalization is used to unify the feature value format, so as to obtain the unique feature value of the current dynamic behavior information. For example, for text information, a text fingerprint algorithm is used to calculate its unique text fingerprint, eliminating the interference of meaningless characters such as spaces and punctuation marks; for image files, a visual feature extraction algorithm is used to extract the core visual features of the image, such as texture and color distribution, to generate a fixed-length visual feature value; for video clips, audio feature extraction and keyframe visual feature extraction are combined to fuse the audio feature value and the keyframe visual feature value to generate a comprehensive video feature value; for link content, URL hash calculation is combined with the core text fingerprint of the page to generate a joint feature value containing the link address and page content; ensuring that the feature value format of different types of information is unified provides a consistent data foundation for subsequent comparisons.

[0088] The core retrieval factors for extracting current dynamic behavior information include target scene identifier, unique code of current target activity used to limit the scope of target scene for retrieval, time partition code, code generated according to time period used to divide the storage partition of historical information database according to time dimension, terminal account hash fragment, partial characters extracted after hashing terminal account, taking into account terminal correlation and data security, and device identifier hash fragment.

[0089] The initial search string is formed by concatenating the above search factors in a fixed order: target scene identifier, time partition code, terminal account hash fragment, device identifier hash fragment, and unique feature value. Finally, the initial search string is compressed using a hash compression algorithm to generate a fixed-length composite search code, realizing the mapping of multi-dimensional search information to a single search identifier and avoiding the time-consuming associated queries when searching multiple fields.

[0090] A two-tiered index system is constructed in the historical information database, consisting of a composite retrieval code primary index and dimension auxiliary indexes. The primary index adopts a B+ tree structure, using the composite retrieval code as the index key to associate the storage address of the corresponding historical dynamic behavior information record in the historical information database. The leaf nodes of the B+ tree are arranged in order according to the composite retrieval code, supporting efficient exact matching and range queries. The dimension auxiliary indexes include a target scene identifier, a time partition combined index, and a terminal account hash fragment and device identifier hash fragment combined index. The former divides the historical information database into multiple physical storage partitions according to the combination of the target scene identifier and the time partition code, while the latter is used to filter out a subset of historical records from the same terminal or device associated with the current information within a specific storage partition. During retrieval, the physical storage partition to which the current dynamic behavior information belongs is first located through the information reach behavior identifier time partition combined partition index; then, within this partition, the historical record subset of the same terminal or device is filtered out through the terminal account hash fragment and device identifier hash fragment combined index; finally, within this historical record subset, exact matching is performed through the composite retrieval encoding main index. If a potential matching record is found, it is determined that the current dynamic behavior information has undergone sensitive word identification; otherwise, it is determined that the current dynamic behavior information has not undergone sensitive word identification.

[0091] Step S2: Based on the marketing mobile terminal information interaction behavior text dataset, obtain the standardized word segmentation sequence of the current dynamic behavior information. By comparing the word segments, match each word in the standardized word segmentation sequence with the word segmentation verification library to distinguish between matched and unmatched words. For matched words, directly retrieve their historical synonym set from the word segmentation verification library. For unmatched words, add them to the word segmentation verification library. Then, combine the behavior-specific word library and the avoidance expression corpus to perform synonym matching, construct an initial synonym set and store it in the word segmentation verification library. Finally, complete the construction of the synonym dataset of the words in the standardized word segmentation sequence, providing semantic reference for subsequent sensitive word identification. The word segmentation comparison uses word segmentation string features as the basis. The word segmentation string features are generated by a hash algorithm after format normalization to represent the uniqueness of the string. If the match is successful, it is determined to be a matched word segment, and its historical synonym set is directly retrieved. If the match fails, it is determined to be an unmatched word segment, and it is marked as a newly added unverified word segment. Its semantic vector, hash value and associated metadata fields are stored in the word segmentation verification library. Then, relying on the behavior-specific lexicon to extract the core semantics, associate it with the avoidance expression corpus, and combine it with the word segmentation context in the text dataset to supplement the contextual content, construct the initial synonym set and bind it to the word segment for storage.

[0092] In this embodiment, by prioritizing the identification of matched word segments through word segmentation comparison and reusing historical synonym sets, the repeated processing of identical word segments can be avoided, reducing redundant calculations and improving preprocessing efficiency. For unmatched word segments, a synonym set is constructed by combining a target scenario-specific lexicon and an avoidance expression corpus. This set can cover basic expressions, avoidance forms, and scenario-specific variations, ensuring that the synonym set fits the actual needs of the target scenario and provides comprehensive semantic support for sensitive word identification. Newly added word segments and their synonym sets are simultaneously stored in the word segmentation verification library, which not only improves the library data but also lays the foundation for the matching and reuse of identical word segments on other terminals in the future. This enables long-term recycling of word segmentation data and ensures the continuity and efficiency of the sensitive word identification process.

[0093] Please see Figure 4 Preferably, the steps for constructing a synonym dataset for segmented words in a standardized word segmentation sequence include:

[0094] From the marketing mobile terminal information interaction behavior text dataset, standardized word segmentation sequences are retrieved. For each word in the standardized word segmentation sequence, based on the hash value of the word segmentation string, a search and matching process is performed in the word segmentation verification library to identify matched and unmatched word segments. The word segmentation verification library is a structured database that supports word segmentation matching and synonym reuse in the target scenario. It is used to store the associated data of all matched word segments in the current target scenario and the set of synonyms. The word segmentation string hash value is a fixed-length character sequence obtained by first performing format normalization on the current word segmentation string and then calculating it through a hash algorithm. It is used to quickly filter processed word segments in the word segmentation verification library that are completely consistent with the current word segmentation at the character level, avoiding misjudgments caused by differences in character format and improving the efficiency of search and matching.

[0095] For the matched words in the standardized word segmentation sequence, retrieve the historical synonym set from the word segmentation verification library; and match the synonyms of the unmatched words in the standardized word segmentation sequence to establish a synonym set, which is then added to the word segmentation verification library.

[0096] Specifically, the steps for establishing a synonym set by matching unmatched synonyms of the standardized word segmentation sequence include:

[0097] By comparing unmatched word segments with core terms in the scenario using a semantic matching algorithm, and combining this with compliance-sensitive boundary words to define semantic boundaries, the core semantics of the unmatched word segments are located, namely:

[0098] Retrieve the behavior-specific terminology library. First, based on the target scenario in the metadata associated with the unmatched word segmentation, call the behavior-specific terminology sub-library of the corresponding target scenario tag in the behavior-specific terminology library. Extract the core terminology set of the scenario from the behavior-specific terminology sub-library, such as the yield, risk warning, and minimum investment amount in the financial sub-library, and the set of industry compliance sensitive boundary words, such as absolute and superlative word variant roots that need to be monitored.

[0099] Perform standardization processing on the unmatched word segments, the set of scene core terms, and the set of industry compliance sensitive boundary words, unify the encoding format, and avoid format differences from interfering with the semantic comparison results. Call the BERT pre-trained model fine-tuned with the target domain corpus to convert the pre-processed unmatched word segments, the set of scene core terms, and the set of industry compliance sensitive boundary words into semantic vectors of a fixed dimension;

[0100] Use the cosine similarity algorithm to calculate the similarity values between the semantic vectors of the unmatched word segments and the semantic vectors of each scene core term; set a similarity threshold, filter out the scene core terms with similarity values higher than the preset similarity threshold, and form a candidate associated term list, and the candidate associated term list is sorted in descending order of similarity values.

[0101] Compare the semantic vectors of the unmatched word segments with the semantic vectors of the set of industry compliance sensitive boundary words. Through semantic inclusion analysis, calculate the intersection ratio of the semantic vectors of the unmatched word segments and the semantic vectors of the industry compliance sensitive boundary words. Preset an intersection ratio threshold. If the intersection ratio of a certain industry compliance sensitive boundary word and the unmatched word segment is higher than the preset intersection ratio threshold, determine whether there is a violation semantic association for the unmatched word segment. If so, define the semantic boundary, and clarify the violation category to be excluded from the core semantics. For example, if the unmatched word segment is most optimal full reduction associated with the most category boundary word, the boundary is defined as excluding the most absolute expression of the most category and retaining the core semantics of full reduction.

[0102] Based on the similarity values of each candidate associated term in the candidate associated term list, perform normalization processing, and use the weighted average algorithm to fuse the semantic vectors of each candidate associated term to generate the core semantic vector of the unmatched word segment; then map the core semantic vector to the semantic space of the behavior-specific word library, identify the scene core terms through the K-nearest neighbor algorithm, and combine the semantic boundary to generate structured core semantics.

[0103] Construct an avoidance expression corpus, which refers to a structured database storing variant forms of violation expressions in the target scene, consisting of homophone replacement expressions, glyph rewriting expressions, and metaphorical association expressions. Calculate the semantic similarity between the core semantics of the unmatched word segments and the avoidance expressions in the avoidance expression corpus, and filter out the avoidance expressions that meet the similarity standard. For example, when the unmatched word segment is most, filter out glyph rewriting expressions such as 蕞 and 罪; when it is cashback, filter out homophone expressions such as 返限, and include them in the synonym set to ensure coverage of the violation avoidance forms in the target scene.

[0104] Retrieve the text context of unmatched words from the marketing mobile terminal information interaction behavior text dataset, analyze the grammatical dependency relationship between the unmatched words and the text context, such as the modification relationship between discount and cross-store, extract scenario-based collocation words and variant expressions, such as cross-store discount and limited-time discount, and add them to the synonym set to make the synonym set fit the actual usage scenario of the unmatched words and avoid the disconnect between general synonyms and target scenarios.

[0105] The generated synonym set is bound to the hash value of the segmented string of the unmatched segment and associated metadata, and stored in the segmentation verification library. At the same time, an initial version tag and a generation timestamp are added to the synonym set to facilitate the tracking of subsequent version iterations.

[0106] By integrating the historical synonym set of matched words with the synonym set of unmatched words, a synonym dataset for each word in the standardized word segmentation sequence is formed.

[0107] Step S3: Construct a statement collection sub-library based on the compliance monitoring requirements of the target scenario. This statement collection sub-library is a structured database that stores manually annotated violation avoidance statements and their corresponding semantic tags, used to provide compliance-oriented statement supplements for the optimization of the synonym set. Subsequently, for the historical synonym set that has been matched and segmented, retrieve the historical sensitive word judgment results associated with it from the word segmentation verification library: if it is determined to be a sensitive word, execute the preset sensitive word processing flow; if it is not determined to be a sensitive word, then the historical synonym set and the unmatched synonym set are used together as the set to be optimized, and optimization is performed based on the statement collection sub-library. Manually annotated avoidance statements in the library are retrieved and processed through... The semantic similarity algorithm calculates the similarity between the expression and the words in the set to be optimized, and filters the expressions with similarity higher than a preset threshold to supplement the corresponding set. Then, a synonym sequence is constructed, and overlapping synonyms in the historical synonym set to be optimized and the unmatched segmented synonym set are removed. Each synonym in the synonym sequence is bound with a unique code to ensure that the code can be reversed to the corresponding original segment. Finally, a matching weight and semantic importance weighted model is constructed to perform bidirectional matching between the synonym sequence and the preset sensitive word library. Finally, based on the probability distribution of synonyms, it is determined whether the segment to be judged is a sensitive word. The segment to be judged includes the matched segment that was not originally judged as a sensitive word and the unmatched segment that was originally not judged as a sensitive word.

[0108] In this embodiment, by constructing a sub-library of expressions that store manually annotated violation avoidance expressions, a compliance-oriented and precise supplement to the synonym set optimization is provided, effectively covering easily overlooked violation forms such as homophones, rewriting, and metaphors, thus solving the problem of insufficient coverage of violation variants in the target scenario by traditional synonym sets. By retrieving the historical sensitive word judgment results of matched word segments, a preset processing flow can be directly executed on the identified sensitive words, avoiding repeated judgments and significantly improving the efficiency of sensitive word processing. For sets not judged as sensitive words, the avoidance expressions in the expression collection sub-library are combined with semantics... Similarity screening further enriches the coverage of synonyms, ensuring that the set not only fits the actual expression of the target scenario but also matches the compliance monitoring requirements. When constructing the synonym sequence, duplicate synonyms are removed to reduce data redundancy. Finally, combined with bidirectional matching, the probability calculation of sensitive words takes into account both the matching accuracy and the semantic influence of word segmentation in the information, effectively reducing the risk of misjudgment and omission. Ultimately, it can accurately determine whether the original word segmentation is a sensitive word, providing an accurate basis for the compliance monitoring of dynamic behavior of marketing mobile terminals, and ensuring the closed-loop effectiveness of the monitoring process from information preprocessing, synonym construction to sensitive word identification.

[0109] Please see Figure 5 Preferably, the steps for determining whether a word segment to be judged is a sensitive word include:

[0110] By combining evasive expressions from historical violation cases, typical violation variations in regulatory notices, and metaphorical violation expressions manually annotated, a sub-database for expression collection is constructed.

[0111] For matched words, the word segmentation verification library is retrieved by searching the hash value of the segmented string, and the historical sensitive word judgment results are called: if it is judged as a sensitive word, the sensitive word processing process is directly triggered, such as marking the violation level, pushing it to the review terminal, and recording the number of violations of the terminal account; otherwise, it is marked as a word segment to be judged, its historical synonym set is marked as a historical set to be optimized, and the unmatched words are marked as words to be judged, and the synonym set of the unmatched words is marked as a historical set to be optimized.

[0112] The expression collection sub-library is invoked, and the cosine similarity algorithm is used to calculate the semantic vector similarity between each synonym in each historical set to be optimized and the avoidance expression in the expression collection sub-library. Avoidance expressions with similarity higher than the preset similarity threshold are filtered out and added to the corresponding historical set to be optimized according to the violation type label, forming the optimized synonym set.

[0113] A synonym sequence is constructed based on the synonyms of each optimized synonym set, and deduplication is performed to remove duplicate words. A unique code is assigned to each synonym in the sequence to associate it with its optimized synonym set and the corresponding segmented word to be judged, including matched and unmatched segments that were not judged as sensitive words. The probability of each synonym in the sequence being a sensitive word is calculated, specifically including:

[0114] Configure a sensitive word library, which refers to a structured summary library of sensitive words constructed according to the compliance monitoring requirements of the target scenario and is used as a reference benchmark for sensitive word determination. It consists of a list of prohibited words issued by the official regulatory agency, sensitive words defined in industry self-discipline norms, sensitive words and their variants manually confirmed in historical violation cases.

[0115] Perform a two-way match between each synonym in the synonym sequence and the preset sensitive word library, including: if the synonym is the same as the sensitive word in the sensitive word library, set the sensitive word probability of the synonym to 1; otherwise, through fuzzy matching, calculate the glyph similarity, homophone similarity, and semantic similarity between the synonym and each sensitive word in the sensitive word library. The glyph similarity refers to the similarity degree between two segmented words in terms of stroke structure and component composition, such as "蕞" and "最", "账" and "帐", which is calculated based on the edit distance algorithm of stroke features: disassemble the segmented words into basic stroke sequences, calculate the edit distance of the two stroke sequences, and then map it to the 0-1 interval after normalization. The closer the value is to 1, the more similar the glyphs are. The homophone similarity refers to the similarity degree between two segmented words in terms of pronunciation, such as "返限" and "返现", "利熄" and "利息", which is calculated through the pinyin string similarity algorithm: convert the words into pinyin strings with tones, use the Levenshtein edit distance to calculate the difference degree of the two pinyin strings, and take 1 minus the difference degree after normalization as the homophone similarity. The semantic similarity refers to the overlapping degree of the semantic connotations of two segmented words in the target scenario, such as "收益无上限" and "最高收益", which is calculated through the cosine similarity algorithm. Call the BERT model fine-tuned with the target domain corpus, convert the synonym and the sensitive word into semantic vectors respectively, and calculate the cosine value of the two semantic vectors. The closer the value is to 1, the more similar the semantics are.

[0116] Select the maximum value of the glyph similarity, homophone similarity, and semantic similarity between the synonym and each sensitive word in the sensitive word library as the sensitive word probability of the synonym.

[0117] Through the associated mapping between the synonym unique encoding and the to-be-determined segmented word, extract and bind the set of sensitive word probabilities corresponding to each segmented word; calculate the basic statistical quantities of the sensitive word probability distribution, and count the number of elements in the sensitive word probability set where the sensitive word probability is greater than the preset significant probability threshold to calculate the significant probability ratio.

[0118] When the percentage of significant probabilities is less than the preset percentage threshold, it indicates that the proportion of words with strong sensitivity tendencies among the synonyms to be segmented is low, the overall sensitivity signal is weak and relatively scattered, and the reference value of extremely high-sensitivity probability words for overall judgment is limited. In this case, the median of the set of sensitivity probability is used as the sensitivity probability of the word to be segmented. The median can effectively avoid the interference of a few high-probability extreme values ​​and is more in line with the sensitivity tendency level of most synonyms. Otherwise, it indicates that the proportion of words with strong sensitivity tendencies among the synonyms to be segmented is high, the overall sensitivity signal is significant and the high-risk characteristics are prominent. The sensitivity orientation of high-probability words is more valuable for judgment. In this case, the maximum value of the set of sensitivity probability is used as the sensitivity probability of the word to be segmented. The maximum value can accurately capture the most significant sensitivity signal and meets the need to prioritize the identification of high-risk expressions in compliance monitoring.

[0119] Based on the differences in compliance and regulatory intensity of target scenarios and the historical accuracy data of sensitive word determination, a probability threshold is configured to distinguish the sensitivity risk level of the words to be judged under different target scenarios, so as to avoid missed judgment in strict regulatory scenarios or misjudgment in weak regulatory scenarios due to a uniform threshold. If the probability of the sensitive word to be judged is greater than the probability threshold, the words to be judged are judged as sensitive words; otherwise, the words to be judged are judged as non-sensitive words.

[0120] Step S4: Execute the sensitive word processing flow based on the sensitive word determination results, and simultaneously record the violation processing log, including processing time, processing personnel, associated terminal account and device identifier, and update the historical information database and word segmentation verification database to facilitate the direct retrieval and reuse of matched words in the future; finally, complete the processing and implementation of the determination results and the data link closure, provide updated historical data support for subsequent compliance monitoring of dynamic behavior of marketing mobile terminals, and reduce the cost of repeated determinations.

[0121] In this embodiment, sensitive and non-sensitive words are handled accordingly to avoid indiscriminate operations; the synchronously recorded structured violation processing log enables full-link traceability of violations from judgment to processing, facilitating subsequent compliance verification; and the real-time feedback updates of the historical information database and word segmentation verification database directly provide reusable judgment results for subsequently matched words, significantly reducing the computation time and data redundancy of repeated judgments; ensuring the continuity of compliance monitoring data for marketing mobile terminals, further improving the efficiency and accuracy of subsequent monitoring.

[0122] Preferably, the steps for executing the sensitive word processing flow based on the sensitive word determination results include:

[0123] The results of the word segmentation sensitivity assessment are categorized as follows: If the result indicates a sensitive word, the sensitive word handling process is initiated; the sensitive word and its corresponding dynamic behavior information are marked as in violation status, and the sensitivity probability, target scenario tag, and corresponding terminal account and device identifier of the word segmentation are associated to form a violation-related data group; a real-time interception mechanism is activated for the marketing mobile terminal to prevent further spread of the violation information. If the result indicates a non-sensitive word, no action is taken.

[0124] Record the execution information of the sensitive word handling process, including processing time, word segmentation text to be judged, judgment result, and processing operation content. Associate terminal account and device identifier to build a structured violation handling log record and ensure that the entire chain of violation behavior is traceable from judgment to handling.

[0125] Preferably, the steps for updating the historical information database and the word segmentation verification database include:

[0126] By using the unique feature values ​​of the dynamic behavior information associated with the words to be judged, the historical information database is retrieved to locate the corresponding dynamic behavior information records; for the located dynamic behavior information records, the sensitive word recognition results, processing status, and processing log index are updated.

[0127] By using the hash value of the segmented string to be judged, the corresponding segmented entry in the segmentation verification library is retrieved, and the sensitive word status, judgment time, and sensitive word probability of the segmented entry are updated; and the processing log index is associated; to ensure that subsequent matched segmented words can directly retrieve the judgment result without repeated calculation;

[0128] The sensitive word status index and word segmentation string hash index of the word segmentation verification library are updated synchronously to ensure that the identified word segmentation entries can be quickly located during subsequent searches, thereby improving matching efficiency.

[0129] This embodiment introduces a marketing mobile terminal dynamic behavior compliance monitoring system, including an information capture module, a synonym matching module, a sensitive word determination module, and a sensitive word processing module;

[0130] The information capture module captures dynamic behavioral information sent by marketing mobile terminals, generates standardized word segmentation sequences of dynamic behavioral information through text processing, and constructs a text dataset of information interaction behavior of marketing mobile terminals; the synonym matching module constructs a synonym dataset of standardized word segmentation sequences; the sensitive word determination module calculates the sensitive word probability of each synonym in the synonym sequence, and obtains the sensitive word probability of each word to be judged by combining the significant probability ratio, so as to determine whether the word to be judged is a sensitive word; the sensitive word processing module executes the sensitive word processing process according to the sensitive word determination results, and updates the historical information database and word segmentation verification database.

[0131] The information capture module, through the configuration of marketing mobile terminal application interfaces, captures text information, image files, video clips, link content, and terminal operation behavior logs in real time. All information is then bound to metadata such as terminal account, device identifier, sending time, and information source application, and categorized and stored in the raw information pool. Subsequently, it calls the historical information database, using information feature value comparison to determine whether the current information has undergone sensitive word identification. If identified, historical results are directly retrieved; otherwise, multi-technology integrated text processing is initiated. After all text undergoes consistency verification and format standardization, it is segmented using a behavior-specific thesaurus to generate standardized word segments, constructing a marketing mobile terminal information interaction behavior text dataset. An inverted index technique is used to build multi-dimensional indexes for terminal, time, and word segmentation, and high-frequency indexes are cached to improve subsequent retrieval efficiency.

[0132] The synonym matching module retrieves standardized word segmentation sequences from the text dataset, searches the word segmentation verification library based on the hash value of the segmented strings, distinguishes between matched and unmatched word segments, and directly retrieves the historical synonym set stored in the library for matched word segments; for unmatched word segments, its semantic vector, hash value, and metadata are first synchronously stored in the word segmentation verification library and marked as newly added unverified, and then the synonym construction process is started: calling the scene sub-library corresponding to the behavior-specific lexicon, extracting the core terminology of the scene and industry compliance-sensitive boundary words, and using a BERT model finely tuned to the target domain corpus to match unmatched word segments and core terminology. Words and boundary words are transformed into semantic vectors. Cosine similarity is used to filter candidate related terms and define semantic boundaries. Then, the avoidance expression corpus is called to filter homophones, word rewritings, and metaphorical expressions that are similar to the core semantics of unmatched words. At the same time, the grammatical dependency relationship of the word segmentation context is analyzed to extract contextual collocation words and integrate them to form an initial synonym set. Finally, the initial synonym set of unmatched words is integrated with the historical synonym set of matched words. After binding the word segmentation hash value and metadata, it is stored in the word segmentation verification library to form a complete synonym dataset, which provides semantic support for subsequent sensitive word recognition.

[0133] The sensitive word determination module constructs a sub-library for expression collection based on the compliance requirements of the target scenario. It then searches the word segmentation verification library, retrieving historical sensitive word determination results of matched segments. If a word is sensitive, a preset processing flow is triggered; otherwise, its historical synonym set and the set of unmatched synonyms are marked as the set to be optimized. The expression collection sub-library is then called, and a cosine similarity algorithm is used to filter high-similarity avoidance expressions to supplement the set to be optimized, forming an optimized synonym set. Based on the optimized set, a deduplicated synonym sequence is constructed, and a unique code is assigned to each word, followed by further processing. A preset sensitive word database is used for bidirectional matching: the probability of sensitive words for exact matching synonyms is set to 1, while for fuzzy matching, the similarity of character shape, homophony, and semantic similarity are calculated and the maximum value is taken as the sensitive word probability; the sensitive word probability set of each word to be judged is extracted, and the significant probability ratio is calculated. If the ratio is lower than the threshold, the median is taken as the sensitive word probability of the word segmentation, otherwise the maximum value is taken; finally, a threshold is configured in combination with the regulatory intensity of the target scenario and the historical accuracy. If the sensitive word probability is higher than the threshold, it is judged as a sensitive word, otherwise it is judged as a non-sensitive word, to ensure that the judgment result meets the compliance requirements of the scenario.

[0134] The sensitive word processing module performs differentiated processing based on the sensitive word determination results: when a word is determined to be sensitive, the sensitive word and its corresponding dynamic behavior information are marked as a violation, and a violation-related data group is formed by associating the sensitive word probability, scene tag, and terminal metadata. Real-time interception is initiated for subsequent messages sent by the terminal containing sensitive words or their synonyms. When a word is determined to be non-sensitive, no processing is performed. A structured violation processing log is recorded synchronously, including processing time, avoidance expression corpus, segmented text to be judged, determination result, operation content, and associated terminal and device identifiers, ensuring full-link traceability. Subsequently, a dual-database update is initiated: the historical information database is called, and the corresponding record is located using the unique feature value of dynamic behavior information, updating the sensitive word recognition result, processing status, and processing log index; the segmentation verification database is called, and the entry is located using the hash value of the segmented string, updating the sensitive word status, determination time, and sensitive word probability, associating the processing log index, and synchronously updating the sensitive word status index and hash index; finally, the processing and data update of the determination results are completed, providing reusable historical data for subsequent terminal dynamic behavior monitoring, reducing the cost of repeated determinations, and ensuring the continuity and efficiency of the system monitoring process.

[0135] Working principle and its effects:

[0136] This invention focuses on the compliance monitoring of dynamic behavior on marketing mobile terminals. Through targeted technical design, it overcomes the pain points of existing monitoring, achieving comprehensive information coverage while taking into account monitoring efficiency and judgment accuracy, thus forming a reusable compliance monitoring system.

[0137] Specifically, in the information input and standardization stage, all types of dynamic behavioral information, including text, images, videos, and links, sent by marketing mobile terminals are captured in real time. By constructing composite retrieval codes and comparing them with historical information databases, information that has already undergone sensitive word identification is quickly identified and the results are reused, reducing redundant calculations. For unidentified information, multiple technologies such as optical character recognition, speech-to-text conversion, and image semantic analysis are integrated to transform non-text information into structured text. Then, after format standardization and contextualized word segmentation, a text dataset is generated. This process not only avoids blind spots in the monitoring of non-text information but also improves information processing efficiency by reusing historical results, laying a consistent data foundation for subsequent analysis. In the semantic reference construction stage, based on the standardized word segmentation sequence of the text dataset, the word segmentation verification library is compared to distinguish between matched and unmatched word segments: matched word segments directly retrieve the historical synonym set, while unmatched word segments are combined with the behavior-specific lexicon to locate the core semantics, and the association avoidance expression corpus is used to screen illegal variants and supplement contextual vocabulary to build a complete synonym dataset. This not only avoids the repeated processing of the same word segments, but also provides comprehensive semantic support for sensitive word identification by covering illegal expression variants, reducing missed judgments caused by semantic missing. In the sensitivity determination and data closure stage, the synonym set is first optimized by collecting a sub-library of expressions, a deduplicated synonym sequence is constructed and the sensitivity probability of each word is calculated, and then the sensitivity probability of the word to be judged is determined according to the proportion of significant probability. Sensitive words are determined in combination with scenario-based thresholds. After the judgment, interception or archiving is performed according to the results, and structured logs are recorded simultaneously and the historical information library and word segmentation verification library are updated to ensure that the judgment results can be directly reused in subsequent monitoring. This process reduces the risk of misjudgment and missed judgment through multi-dimensional probability analysis, and at the same time, the long-term circulation of monitoring data is realized through dual-library updates, ensuring the continuity of the monitoring process and subsequent efficiency.

[0138] In summary, this invention not only solves the problems of repetitive processing, non-text blind spots, insufficient coverage of violation variants, and low judgment accuracy in existing monitoring technologies, but also achieves a dual improvement in monitoring efficiency and compliance accuracy through data reuse and dynamic updates. Ultimately, it forms a feasible and cyclical dynamic behavior compliance monitoring solution for marketing mobile terminals, providing reliable support for tracing violations and long-term compliance management.

[0139] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for monitoring the compliance of dynamic behavior on marketing mobile terminals, characterized in that, include: The system captures dynamic behavioral information sent by marketing mobile terminals, performs matching and retrieval in the historical information database by constructing a composite retrieval code, determines whether the current dynamic behavioral information has been identified as a sensitive word, and if not, generates a standardized word segmentation sequence of the dynamic behavioral information through text processing to construct a text dataset of information interaction behavior of marketing mobile terminals. By comparing word segmentation, the matched and unmatched word segments of the standardized word segmentation sequence are identified. The historical synonym set of the matched word segments is retrieved from the word segmentation verification library. By locating the core semantics of the unmatched word segments, the expression to be avoided is filtered, and a synonym set of the unmatched word segments is generated to construct a synonym dataset of the standardized word segmentation sequence. For matched word segments, retrieve the historical sensitive word judgment results. If the word is determined to be non-sensitive, merge it with the unmatched word segments to form a word segment to be judged. Establish a synonym sequence for the word segment to be judged, and calculate the sensitive word probability of each synonym in the synonym sequence through bidirectional matching. Combine the significant probability ratio to obtain the sensitive word probability of each word segment to be judged, so as to determine whether the word segment to be judged is a sensitive word. Based on the sensitive word determination results, execute the sensitive word processing flow and update the historical information database and word segmentation verification database accordingly; The steps for constructing a standardized word segmentation sequence synonym dataset include: From the text dataset, retrieve the standardized word segmentation sequence. For each word in the standardized word segmentation sequence, perform a search and match in the word segmentation verification library based on the hash value of the word segmentation string to identify the matched and unmatched words. For the matched words in the standardized word segmentation sequence, retrieve the historical synonym set from the word segmentation verification library; and match the synonyms of the unmatched words in the standardized word segmentation sequence to establish a synonym set, which is then added to the word segmentation verification library. The steps for establishing a synonym set include: By using a semantic matching algorithm, unmatched word segments are compared with core terms in the scenario, and semantic boundaries are defined by combining compliance-sensitive boundary words to locate the core semantics of unmatched word segments. The evasion expression corpus is invoked. The evasion expression corpus is a structured database that stores variants of illegal expressions in the target scenario. It consists of homophonic substitution expressions, character rewriting expressions, and metaphorical association expressions. Calculate the semantic similarity between the core semantics of the unmatched word segmentation and the evasion expressions in the corpus, and filter the evasion expressions based on the semantic similarity and include them in the synonym set; Retrieve the text context of the unmatched word segment, parse the grammatical dependency relationship between the unmatched word segment and the text context, extract contextual collocation words and variant expressions, and add them to the synonym set; Integrate the historical synonym set of matched words with the synonym set of unmatched words to form a synonym dataset for each word in the standardized word segmentation sequence; The steps for determining whether the segmented word to be judged is a sensitive word include: The expression collection sub-library is invoked, and the word segmentation verification library is retrieved based on the matched word segmentation. The historical sensitive word judgment results of the matched word segmentation are retrieved: if it is judged as a sensitive word, the sensitive word processing process is triggered; otherwise, it is marked as a word segmentation to be judged, its historical synonym set is marked as a historical set to be optimized, and the unmatched word segmentation is marked as a word segmentation to be judged, and the synonym set of the unmatched word segmentation is marked as a historical set to be optimized. For each historical set to be optimized, filter and add avoidable expressions with similarity higher than the preset similarity threshold in the expression collection sub-library, and construct an optimized synonym set; Based on the synonyms in each optimized synonym set, a synonym sequence is constructed, and deduplication is performed. A unique code is assigned to each synonym in the synonym sequence, and the probability of sensitive words for each synonym in the synonym sequence is calculated. By mapping the unique encoding of synonyms to the words to be judged, the probability set of sensitive words corresponding to each word segment is extracted and bound; Calculate the basic statistics of the probability distribution of sensitive words, and count the number of elements in the sensitive word probability set whose sensitive word probability is greater than the preset significant probability threshold, so as to calculate the proportion of significant probability. When the proportion of significant probability is less than the preset proportion threshold, the median of the sensitive word probability set is used as the sensitive word probability of the word to be segmented; otherwise, the maximum value of the sensitive word probability set is used as the sensitive word probability of the word to be segmented. Configure a probability threshold. If the probability of a sensitive word in the segment to be judged is greater than the probability threshold, then the segment to be judged is determined to be a sensitive word; otherwise, the segment to be judged is determined to be a non-sensitive word. The steps for calculating the probability of sensitive words for each synonym in a synonym sequence include: Configure a sensitive word database and perform bidirectional matching between each synonym in the synonym sequence and the preset sensitive word database. This includes: if the synonym matches a sensitive word in the sensitive word database, the sensitive word probability of the synonym is set to 1; otherwise, through fuzzy matching, the similarity of the character shape, homophony, and semantics between the synonym and each sensitive word in the sensitive word database is calculated. The similarity of the character shape refers to the degree of similarity between two word segments in terms of stroke structure and component composition, which is calculated by an edit distance algorithm based on stroke features; The homophonic similarity refers to the degree of similarity in pronunciation between two word segments, which is calculated using a pinyin string similarity algorithm; The semantic similarity refers to the degree of overlap in the semantic connotation of two word segments in the target scene, which is obtained by calculating the cosine similarity of the semantic vectors of the word segments; The maximum value of the similarity in character shape, homophony, and semantic similarity between the synonym and sensitive word database is selected as the probability of the synonym being a sensitive word.

2. The marketing mobile terminal dynamic behavior compliance monitoring method as described in claim 1, characterized in that, The steps for constructing the text dataset of information interaction behavior on marketing mobile terminals include: Capture dynamic behavioral information sent by marketing mobile terminals, classify and store the information according to its type in the raw information pool, and bind it to the metadata fields of the marketing mobile terminals. The metadata field refers to a structured data field used to uniquely identify the associated subject, interaction scenario, and processing status of dynamic behavior information; For dynamic behavior information stored in the original information pool, the information feature value comparison method is used to compare the information with the historical information database to determine whether the current dynamic behavior information has been identified for sensitive words. The historical information database refers to a structured database that stores dynamic behavioral information of all marketing mobile terminals in the current target scenario and the results of compliance processing of associated sensitive words; If it is determined that sensitive word identification has been performed, the sensitive word identification results and sensitive word compliance processing methods corresponding to the matching historical dynamic behavior information are retrieved through the associated index of the historical information database. If it is determined that no sensitive word recognition has been performed, the current dynamic behavior information and its metadata fields will be stored in the historical information database, and text processing will be started, that is, the dynamic behavior information will be converted into text according to information type and scenario.

3. The marketing mobile terminal dynamic behavior compliance monitoring method as described in claim 2, characterized in that, The steps for constructing the text dataset of information interaction behavior for marketing mobile terminals also include: The original text information that was not identified for sensitive words was extracted and the corresponding non-text text generated by textualization were summarized to form a set of texts to be standardized. Perform text consistency checks and format standardization on the standardized text set to obtain a standardized text set; Configure a behavior-specific thesaurus. The behavior-specific thesaurus refers to a structured thesaurus that is customized for the current target scenario and contains scenario-specific terms and compliance boundary words. It consists of core scenario terms, industry compliance sensitive boundary words, and an exclusion list of invalid function words. Based on a behavior-specific lexicon, a standardized text set is segmented to generate a standardized segmentation sequence. By structurally associating standardized text sets, standardized word segmentation sequences, and corresponding metadata fields, a marketing mobile terminal information interaction behavior text dataset is constructed.

4. The method for monitoring the dynamic behavior compliance of marketing mobile terminals as described in claim 2, characterized in that, The step of determining whether the current dynamic behavior information has been identified for sensitive words includes: For different information types of dynamic behavior information, an appropriate feature extraction algorithm is adopted and multimodal feature normalization is performed to obtain the unique feature values ​​of the current dynamic behavior information. Extract the core retrieval factors of the current dynamic behavior information, including the target scene identifier, time partition code, terminal account hash fragment, and device identifier hash fragment; combine them with the exclusive feature values, concatenate them in a fixed order to form an initial retrieval string, and compress the initial retrieval string through a hash compression algorithm to generate a composite retrieval code of fixed length; A two-tiered index system consisting of a composite retrieval code main index and dimension auxiliary indexes is constructed in the historical information database. The main index adopts a B+ tree structure and uses the composite retrieval code as the index key. The dimension auxiliary indexes include a joint index of target scene identifier, time partition, terminal account hash fragment, and device identifier hash fragment. During retrieval, the physical storage partition is located through the dimension-assisted index, and a subset of historical records is constructed. Within the subset of historical records, matching is performed through the composite retrieval encoded main index. If a potential matching record is found, it is determined that the current dynamic behavior information has been identified as a sensitive word; otherwise, it is determined that the current dynamic behavior information has not been identified as a sensitive word.

5. The method for monitoring the dynamic behavior compliance of marketing mobile terminals as described in claim 1, characterized in that, The steps for locating the core semantics of unmatched word segments include: Based on the target scenario in the metadata associated with the unmatched word segmentation, call the behavior-specific word sub-library of the corresponding target scenario tag in the behavior-specific word library; Extract the core terminology set for specific behaviors and the set of industry compliance-sensitive boundary words; Standardization processing is performed on unmatched word segments, the core terminology set of the scenario, and the industry compliance-sensitive boundary word set, and they are converted into semantic vectors of fixed dimensions. Calculate the similarity value between the semantic vector of the unmatched word segmentation and the semantic vector of each scene's core term; set a similarity threshold to filter scene core terms and form a list of candidate related terms; The semantic vectors of unmatched words are compared with the semantic vectors of the industry compliance sensitive boundary word set. The intersection ratio of the semantic vectors of unmatched words and the semantic vectors of industry compliance sensitive boundary words is calculated to determine whether there is a non-compliant semantic association between the unmatched words. If so, the semantic boundary is defined. Based on the similarity values ​​of each candidate related term in the candidate related term list, normalization processing is performed, and a weighted average algorithm is used to fuse the semantic vectors of each candidate related term to generate the core semantic vector of the unmatched word segmentation. Combined with the semantic boundary, a structured core semantic is generated.

6. The marketing mobile terminal dynamic behavior compliance monitoring method as described in claim 1, characterized in that, The steps of executing the sensitive word processing flow based on the sensitive word determination result and updating the historical information database and word segmentation verification database include: If the word to be judged is determined to be a sensitive word, the sensitive word handling process will be initiated, the sensitive word and its corresponding dynamic behavior information will be marked as a violation, and a real-time interception mechanism will be initiated for the marketing mobile terminal; otherwise, no action will be taken. Record the execution information of the sensitive word handling process and build a structured violation processing log; By using the unique feature values ​​of dynamic behavior information associated with the words to be judged, the historical information database is retrieved, the corresponding dynamic behavior information records are located, and the sensitive word recognition results, processing status, and processing log index are updated. By using the hash value of the segmented string to be judged, the corresponding segmented entry in the segmentation verification library is retrieved, and the sensitive word status, judgment time, and sensitive word probability of the segmented entry are updated; and the processing log index is associated.

7. A marketing mobile terminal dynamic behavior compliance monitoring system, used to implement the marketing mobile terminal dynamic behavior compliance monitoring method according to any one of claims 1-6, characterized in that, It includes an information capture module, a synonym matching module, a sensitive word determination module, and a sensitive word processing module: The information capture module is used to capture dynamic behavioral information sent by the marketing mobile terminal, and generate a standardized word segmentation sequence of the dynamic behavioral information through text processing to construct a text dataset of information interaction behavior of the marketing mobile terminal; the synonym matching module is used to construct a synonym dataset of the standardized word segmentation sequence; the sensitive word determination module is used to calculate the sensitive word probability of each synonym in the synonym sequence, and obtain the sensitive word probability of each word to be judged by combining the significant probability ratio, so as to determine whether the word to be judged is a sensitive word; the sensitive word processing module is used to execute the sensitive word processing process according to the sensitive word determination result, and update the historical information database and the word segmentation verification database.

Images