A multi-dimensional recognition network chat sensitive word auditing method and system

Abstract

The application belongs to the technical field of network chat auditing, and provides a network chat sensitive word auditing method and system based on multi-dimensional identification, which comprises the following steps: collecting multi-source data in real time and extracting core keywords, obtaining text with core keywords and interactive data corresponding to the text, identifying hot events by analyzing the interactive data, and extracting temporary sensitive words based on the hot events. The application collects multi-source data in real time and dynamically extracts temporary sensitive words related to hot events, optimizes the problem of response lag of emerging sensitive words generated by hot events in the traditional static sensitive word library, calculates the correlation score by means of semantic similarity and co-occurrence frequency, can cover explicit associated words and identify implicit reference words, reduces the missed judgment caused by the variable form of words, and improves the timeliness and comprehensiveness of sensitive word identification.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of network chat auditing, and in particular relates to a multi-dimensional recognition-based network chat sensitive word auditing method and system. BACKGROUND

[0002] Network chat sensitive word auditing refers to the process of monitoring, identifying and processing the chat content of users on a network platform (such as social media, instant messaging tools, forums, etc.) through technical means and rule systems, screening out words or expressions related to specific sensitive information, to regulate network communication behavior and prevent the spread of risks, with the core goal of balancing network freedom of speech and information security, and avoiding the spread of sensitive content (such as illegal and irregular information, privacy leaks, etc.) to cause adverse effects on society or individuals.

[0003] However, multi-dimensional recognition often relies on the completeness of the sensitive word library, but the dynamic nature and boundary definition of the sensitive word library are easily overlooked, including temporary sensitive words generated by hot events, i.e. social hot events quickly generate new sensitive words, which are time-sensitive, such as the names and codes of specific events in sudden events and public opinion disputes, which may only be sensitive during the event fermentation period. If the word library update relies on manual annotation, it will lag behind the speed of network dissemination, leading to missed judgments.

[0004] Therefore, the application provides a multi-dimensional recognition-based network chat sensitive word auditing method and system. SUMMARY

[0005] In order to make up for the shortcomings of the prior art and solve at least one technical problem raised in the background art.

[0006] The technical solution adopted by the application to solve its technical problems is: a multi-dimensional recognition-based network chat sensitive word auditing method, comprising the following steps:

[0007] Real-time collection of multi-source data and core keyword extraction, obtaining text with core keywords and interactive data corresponding to the text, identifying hot events by analyzing interactive data, and extracting temporary sensitive words based on hot events;

[0008] Based on the temporary sensitive word library, scene-based effective rules are set in combination with the temporary word library and hot event labels, and the network chat is audited;

[0009] Real-time monitoring of hotness data of hot events, and analysis and output of hotness comprehensive value, change analysis of the hotness comprehensive value based on two dimensions of regular trend and resurgence trend, and evaluation of the timeliness of temporary sensitive words according to the change analysis result of the hotness comprehensive value;

[0010] According to the timeliness evaluation result of the temporary sensitive word, the review intensity of the temporary sensitive word is adjusted.

[0011] A network chat sensitive word review system based on multi-dimensional identification, the system comprises:

[0012] The temporary sensitive word acquisition module: real-time collection of multi-source data and core keyword extraction, obtaining the text with core keywords and the interactive data corresponding to the text, identifying hot events by analyzing the interactive data, and extracting temporary sensitive words based on hot events;

[0013] The chat review module: based on the temporary sensitive word, a temporary word library is constructed, combined with the temporary word library and the hot event label, a scenario-based effective rule is set, and the network chat is reviewed;

[0014] The timeliness analysis module: real-time monitoring of hotness data of hot events, and analyzing and outputting the hotness comprehensive value, realizing the change analysis of the hotness comprehensive value based on the double dimensions of the regular trend and the relapse trend, and evaluating the timeliness of the temporary sensitive word according to the change analysis result of the hotness comprehensive value;

[0015] The review intensity adjustment module: according to the timeliness evaluation result of the temporary sensitive word, the review intensity of the temporary sensitive word is adjusted.

[0016] The beneficial effects of the present application are as follows:

[0017] The present application extracts the temporary sensitive words related to hot events in real time by collecting multi-source data and dynamically extracting the temporary sensitive words related to hot events, optimizes the problem of the response lag of the traditional static sensitive word library to the emerging sensitive words generated by hot events, calculates the relevance score by means of semantic similarity and co-occurrence times, can cover explicit associated words and identify implicit reference words, reduces the missed judgment caused by the variable form of words, and improves the timeliness and comprehensiveness of sensitive word identification;

[0018] The present application dynamically evaluates the timeliness of the temporary sensitive word by real-time monitoring of the hotness data, adjusts the review intensity according to the trend change and relapse, forms a new whole-cycle management mechanism, which can not only strengthen the risk prevention and control in the active period of the event, but also reduce unnecessary intervention in the recession period, significantly improves the flexibility, accuracy and efficiency of the network chat sensitive word review. BRIEF DESCRIPTION OF DRAWINGS

[0019] The present application will be further described below with reference to the accompanying drawings.

[0020] Figure 1 is a step flow chart of a network chat sensitive word review method based on multi-dimensional identification of the present application;

[0021] Figure 2It is a kind of acquisition step flow chart of hot events in the method for auditing sensitive words in network chat based on multi-dimensional identification network of the application;

[0022] Figure 3 It is an architecture diagram of the system for auditing sensitive words in network chat based on multi-dimensional identification network of the application. DETAILED DESCRIPTION

[0023] In order to make the technical means, creative features, purposes and effects realized by the application easy to understand, the application will be further described below in combination with specific embodiments. Embodiment 1

[0024] Please refer to Figure 1 and Figure 2 The method for auditing sensitive words in network chat based on multi-dimensional identification network of the application includes the following steps:

[0025] Step one: real-time collection of multi-source data and analysis of the multi-source data, extraction of core keywords for identification of hot events, and extraction of temporary sensitive words based on hot events;

[0026] In some embodiments, the first thing that needs to be defined is that a hot event refers to an event or topic that is rapidly spread through a network platform (such as social media, news websites, forums, instant messaging tools, etc.) within a certain period of time, and that attracts high public attention, extensive discussion and a large number of interactions (such as likes, comments, forwards, shares, etc.).

[0027] In step one, the process of real-time collection of multi-source data is as follows:

[0028] Through the use of web crawlers and API interfaces, real-time data grabbing of each platform is carried out, and message queues (such as Kafka) are used to cache the collected data to ensure the orderly processing of data.

[0029] The collection sources of multi-source data include but are not limited to social media platforms, news websites, forums, and instant messaging tools.

[0030] For example, the social media platform: high-frequency words and hot words in the comment area under the event-related topic, such as the name, code and specific expression of a certain event;

[0031] News website: core entities in mainstream media reports, such as names, places and event abbreviations; user-generated referential vocabulary in the comment area, such as "a certain melon" and "that matter" in a certain event;

[0032] Instant messaging tool group chat sample: select event-related chat records in public group chats and extract high-frequency new words or abnormal expressions;

[0033] In step one, the process of extracting core keywords is:

[0034] First, the collected multi-source data is preprocessed, including data cleaning and word segmentation processing;

[0035] Data cleaning includes removing noise such as special symbols, HTML tags;

[0036] Word segmentation processing includes splitting text into individual words and extracting keywords using the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm;

[0037] Specifically, TF-IDF measures the importance of a word by its frequency in a single text and its scarcity in all texts;

[0038] TF is the number of times a word appears in a text divided by the total number of words in the text; IDF is log(total number of documents / number of documents containing the word +1); TF-IDF is the product of TF and IDF;

[0039] Extract the words whose TF-IDF values rank in the top 25% of the ranking as candidate keywords; for batch text, extract candidate keywords with TF-IDF values greater than or equal to the TF-IDF threshold as core keywords;

[0040] The determination of batch text is set by a person skilled in the art according to experience and event characteristics, such as a collection of texts within 1 hour on a certain platform; the TF-IDF threshold can be adjusted according to the field, such as a higher threshold for social events than for entertainment events;

[0041] In step one, the process of identifying hot events is:

[0042] Extract related texts that contain core keywords and obtain interaction data for each text;

[0043] Among them, the interaction data includes: likes, comments, forwards, shares, and collections;

[0044] Sum the likes, comments, forwards, shares, and collections of all texts and perform mean processing to obtain absolute interaction volume;

[0045] The ratio of the current period's interaction volume to the previous period's interaction volume is the relative interaction volume;

[0046] Set absolute interaction volume threshold and relative interaction volume threshold, extract texts with absolute interaction volume greater than absolute interaction volume threshold and relative interaction volume greater than relative interaction volume threshold as hot events;

[0047] Wherein, the absolute interaction amount threshold and the relative interaction amount threshold are set by the person skilled in the art according to experience and event characteristics; the period is preset by the person skilled in the art;

[0048] For example, taking a social event as an example:

[0049] Cleaning HTML tags in news comments , special symbol "#", after word segmentation, get "event, involved enterprise, security risks, investigation" and other keywords;

[0050] The TF-IDF values of "involved enterprise" and "security risks" are 0.45 and 0.42, respectively, and the core keywords are selected;

[0051] Extract all texts that exist core keywords, the total sum of likes: 325678; The total sum of comments: 128905; The total sum of forwarding: 876543; The total sum of sharing: 456789; The total sum of collection: 98765;

[0052] The absolute interaction amount is: 377336; If the average interaction amount of the same type of event in the last period (yesterday) is: 150000; The relative interaction amount is: 2.5156; The absolute interaction amount (377336) > absolute threshold (100000) and the relative interaction amount (2.5156) > relative threshold (2.0); At the same time, the absolute interaction amount and the relative interaction amount threshold requirements are met, and it is recognized as a hot event;

[0053] In step one, the process of extracting temporary sensitive words based on hot events is:

[0054] Taking the identified hot event as the core, extract all texts containing the core keywords of the hot event;

[0055] Pretreatment of the selected text, including data cleaning and word segmentation processing;

[0056] Data cleaning: remove noise in the text, such as special symbols, HTML tags, and ensure the purity of the text.

[0057] Word segmentation processing: split the text into individual words for subsequent extraction of sensitive words;

[0058] In the pretreated text, extract the words highly related to the hot event as temporary sensitive words; Temporary sensitive words include but are not limited to: specific names, codes, controversial expressions, implicit reference words generated in the event fermentation process;

[0059] For example, in a hot event of a certain star's rumor, the attack nicknames and event-related implicit reference words extracted from the related text may be determined as temporary sensitive words;

[0060] For temporary sensitive words, screen them by relevance, the process is:

[0061] For any core keyword, use a pre-trained word vector model to convert the core keyword and the candidate words in the pretreated text into a multi-dimensional vector, and calculate the semantic similarity between the vector of all words in the pretreated text and the vector of the core keyword by cosine similarity.

[0062] The pre-trained word vector model includes but is not limited to Word2Vec, BERT;

[0063] Statistics of the words appearing together with the core keywords in the preprocessed text, and the number of co-occurrences is counted;

[0064] For any candidate word in the preprocessed text, the semantic similarity and the number of co-occurrences are summed and averaged to output the correlation score;

[0065] Extract the candidate word whose correlation score is greater than the correlation score limit as the temporary sensitive word;

[0066] The reason for extracting the temporary sensitive word through the above process is as follows: first, the temporary sensitive word has strong relevance to the hot event, reducing irrelevant word interference; second, semantic similarity can identify implicit reference words, and co-occurrence frequency can strengthen explicit association words, covering both explicit expressions and implicit references, adapting to the diversified expression in network chat and reducing the missed judgment caused by the variable form of words; third, the sensitive words of the hot event will evolve with the fermentation of the event, and the real-time calculation of the co-occurrence frequency and the semantic similarity can capture these newly generated words, when the co-occurrence frequency of the new word and the core keyword increases, the semantic association is enhanced, the correlation score will be dynamically improved, so that it is timely included in the temporary sensitive word library, and the problem of lag of traditional static word library to emerging sensitive words is optimized;

[0067] Step two: based on the obtained temporary sensitive word, construct a temporary word library, and combine the temporary word library with the hot event label to set a scenario-based effective rule to audit the network chat;

[0068] In some embodiments, the temporary sensitive word is obtained, and the temporary word library is constructed as the core audit object;

[0069] Add multi-dimensional attribute labels to each temporary sensitive word, including: associated event label, risk type label;

[0070] Associated event label: clearly indicates the hot event to which the word belongs, for example, "# certain food safety event #", "# certain star gossip #";

[0071] Risk type label: mark the sensitive type of the word, for example, "privacy type", "dispute type", "violation type", corresponding to "unpublished address", "aggressive nickname", "homonym violation word";

[0072] Add multi-dimensional attribute labels to the hot event, including: event type label, event type label, and sensitive trigger point label;

[0073] The event type label is used to distinguish the nature of the event: such as social livelihood events, entertainment gossip events, and emergencies;

[0074] The transmission platform label is used to mark the main platform of the event transmission: such as microblogging, group chat diffusion, and news comment area fermentation;

[0075] The sensitive trigger point label is used to record the core contradiction in the event that is most likely to cause controversy: such as privacy leakage risk;

[0076] Setting scenario-based effective rules for network chat auditing, specifically:

[0077] When temporary sensitive words appear in network chat content, match the associated event label, event type label, and transmission platform label of the hot event, and when the three are consistent, trigger specific audit rules;

[0078] Among them, by matching the consistency of the three, the possibility of cross-event and cross-platform misjudgment can be reduced;

[0079] Exemplary: Temporary sensitive word "certain melon", associated event label "#certain star rumor#", only in the group chat scenario corresponding to the hot event "#certain star rumor#", event type label "entertainment gossip", and transmission platform label "group chat diffusion", trigger the audit;

[0080] Differential auditing based on event type labels can include: when the event type label is social livelihood or emergency, and when the event type label is entertainment gossip;

[0081] When the event type label is social livelihood or emergency, trigger strict audit mode:

[0082] Including but not limited to: if the transmission platform is all platforms (including microblogging, news comment area, group chat, etc.), as long as the temporary sensitive word (such as the location code involved, unpublicized casualty data) appears, immediate real-time interception is performed, and the sender's IP and account information are pushed to the artificial audit queue (reviewed within 3 minutes) for tracing;

[0083] Exemplary, temporary sensitive word certain chemical plant leakage (associated event label "#certain chemical safety event#", event type label "emergency", transmission platform label "all-platform diffusion"), when appearing in microblogging, group chat, and news comments, are automatically intercepted, and after artificial review confirms whether it is false information, decide to release or permanently shield;

[0084] When the event type label is "entertainment gossip", trigger the lenient audit mode:

[0085] Including but not limited to: if the transmission platform is microblog or social media, the temporary sensitive word (such as offensive nickname) appears in a single content >= 3 times to intercept, appears 1-2 times only pop-up window to remind "please pay attention to the expression of civilization";

[0086] If the transmission platform is a group chat, only when the word is reported by 2 or more group members, the delayed interception (shielded within 30 minutes without withdrawal) is triggered;

[0087] The core of the scene is to let the audit rules be in place, limit the scene range of the audit by matching the event type label, the event type label and the transmission platform label of the temporary sensitive word, and reduce the impact of "one size fits all"; Different scenes have different communication needs and different transmission risks, and the rules adapt to the core characteristics of the scene through the combination of event type and transmission platform;

[0088] This embodiment has at least the following effects:

[0089] First, the problem of response lag of traditional static word library to temporary sensitive words is optimized, and the words with strong association with hot events are included in the audit range in time through real-time correlation score calculation;

[0090] Second, both explicit associated words and implicit reference words are covered, reducing the missed judgment caused by the diversification of network expression, ensuring the strong association of temporary sensitive words and hot events, and laying a foundation for subsequent accurate audit;

[0091] Third, the effective scene of the audit is limited, and the misjudgment across events and platforms is reduced;

[0092] Fourth, different rules are formulated for different event types and transmission scenes, which effectively prevent risks while reducing the excessive intervention of one-size-fits-all audit on normal communication of users, balancing the needs of information security control and network freedom of speech.

[0093] Embodiment 2

[0094] Please refer to Figure 1 The multi-dimensional recognition network chat sensitive word audit method provided by the embodiment of the application further includes the following steps:

[0095] Since the core feature of temporary sensitive words is to spread rapidly with the outbreak of hot events, as the hotness of hot events decreases, temporary sensitive words have a certain timeliness, so that the temporary word library is dynamically updated;

[0096] Step 3: Real-time monitoring of hotness data of hot events, analyzing the change trend of hotness data, and evaluating the timeliness of temporary sensitive words based on the change trend of hotness data;

[0097] Since the time limit of the temporary sensitive word is strongly related to the event heat and public opinion cycle, but the heat decay can be dynamic and nonlinear, for example, a certain event can cool down after a week, but suddenly rekindle due to new developments, so the analysis of the trend of heat data is particularly important. If the trend of heat data is not analyzed, it may lead to the loss of powerful basis for the invalid management of temporary sensitive words, and increase the possibility of sensitive word misjudgment and excessive review;

[0098] In some embodiments, the heat data of the hot event is monitored in real time; wherein the monitoring object is centered on the identified hot event, and the related heat data is collected in real time, the heat data including: absolute interaction amount, relative interaction amount and interaction quality value;

[0099] Wherein, the calculation method of absolute interaction amount and relative interaction amount is the same as that in step one, which is not repeated here;

[0100] The calculation process of the interaction quality value is: extracting the comments whose word count exceeds the average word length, and counting the number as the comment word count exceeding number, and calculating the ratio of the comment word count exceeding number to the total comment number as the comment depth;

[0101] Extracting the comments replied by other users and counting the number as the secondary interaction comment number, and calculating the ratio of the secondary interaction comment number to the total comment number as the secondary interaction proportion;

[0102] The comment depth and the secondary interaction proportion are weighted and fused to output the interaction quality value;

[0103] Wherein, as known by those skilled in the art, the weight of the comment depth is generally higher than that of the secondary interaction proportion, the weight of the comment depth is generally set to 0.6, and the weight of the secondary interaction proportion is generally set to 0.4,

[0104] After the absolute interaction amount, the relative interaction amount and the interaction quality value are normalized, they are weighted and fused to output the heat comprehensive value;

[0105] Wherein, as known by those skilled in the art, the weight of the normalized absolute interaction amount is generally 0.4, the weight of the normalized relative interaction amount is generally 0.3, and the weight of the interaction quality value is generally 0.3;

[0106] For any hot event, the double-dimensional analysis of the conventional trend and the rekindling trend is carried out, and the specific process is:

[0107] The conventional trend is analyzed by trend direction, cycle division and conventional prediction: based on the heat comprehensive value of 3-5 consecutive time periods, the natural trend type is identified;

[0108] The time period is preset by a person skilled in the art;

[0109] The deviation value of the heat comprehensive value of the adjacent time period is calculated to obtain a heat comprehensive value deviation value sequence;

[0110] The positive heat comprehensive value deviation value is extracted, and the number thereof is counted, and then the number is calculated by ratio with the total heat comprehensive value deviation value to obtain a direction consistency index;

[0111] The heat comprehensive value deviation value is weighted according to time, and weighted summation is performed to obtain an amplitude index;

[0112] The weight according to time increases with the period, that is, the recent deviation value has a greater impact on the trend, and the weight of the Kth heat comprehensive value deviation value is K / (n-1), and n is the total number of heat comprehensive value deviation values;

[0113] The maximum number of continuous same-direction deviation of the heat comprehensive value deviation value sequence is counted as a continuous trend index;

[0114] The direction consistency index, the amplitude index and the continuous trend index are standardized and then weighted and fused to obtain a first comprehensive trend value;

[0115] The standardization process is as follows: the direction consistency index C is standardized as C'=2C-1; C' represents the standardized direction consistency index; the amplitude index W is standardized as W'=W / |max(W)|; W' represents the standardized amplitude index; the continuous trend index S is standardized as S'=S / max(S), and S' represents the standardized continuous trend index; weights are assigned according to the importance of the indexes, and the direction consistency index (0.4)>amplitude index (0.3)>continuous trend index (0.3);

[0116] When the first comprehensive trend value is greater than or equal to a first judgment value (0.2), it is determined to be an upward trend; when the first comprehensive trend value is greater than a second judgment value (-0.2) and less than the first judgment value (0.2), it is determined to be a fluctuation trend; and when the first comprehensive trend value is less than or equal to the second judgment value (-0.2), it is determined to be a downward trend;

[0117] When the upward trend is presented, the temporary sensitive word is strong timeliness; when the downward trend is presented, the temporary sensitive word is weak timeliness, and when the fluctuation trend is presented, the temporary sensitive word is stable timeliness;

[0118] First of all, it is necessary to define that 50% of the initial heat comprehensive value of the hot event is the recession baseline, and when the heat comprehensive value is lower than the recession baseline, it is determined that the heat trend enters the recession period;

[0119] Take 30% of the recession baseline as the relapse trigger baseline, that is, when the heat comprehensive value rises to the relapse trigger baseline, start the relapse analysis; the relapse analysis includes verifying the relapse trend:

[0120] Monitor whether the heat comprehensive value continuously maintains at least 2 time periods not lower than the baseline after rising to the relapse trigger baseline, if only 1 period meets the standard, it is determined as short-term fluctuation and invalid relapse; if 2 or more consecutive periods meet the standard, the next step verification is entered;

[0121] Calculate the interaction quality value of the relapse period and compare it with the average interaction quality value of the recession period;

[0122] The calculation method of the interaction quality value is the same as the aforementioned calculation method, which is not repeated here;

[0123] If the interaction quality value of the relapse period is greater than or equal to 1.2 times the average value of the recession period, it means that the rise is accompanied by high-quality discussion rather than low-quality screen brushing, which is valid relapse; otherwise, it is invalid relapse;

[0124] Calculate the second comprehensive trend value after relapse, the comprehensive trend value is calculated in the same way as the first comprehensive trend value in the aforementioned regular trend analysis, which is not repeated here;

[0125] If it is valid relapse and the second comprehensive trend value is greater than or equal to the first judgment value (0.2), it is determined as complete relapse, and the time effectiveness of the temporary sensitive word is adjusted from the original weak time effectiveness (recession period) to strong time effectiveness;

[0126] If it is valid relapse and the second comprehensive trend value is less than or equal to the second judgment value (-0.2), it is determined as recession relapse, and the time effectiveness of the temporary sensitive word remains weak time effectiveness (recession period);

[0127] If it is invalid relapse, the second comprehensive trend value is greater than the second judgment value (-0.2) and less than the first judgment value (0.2), it is determined as fluctuation relapse, and the time effectiveness of the temporary sensitive word is adjusted from the original weak time effectiveness (recession period) to stable time effectiveness;

[0128] By monitoring the heat data of the hot event in real time, calculating the heat comprehensive value and analyzing the regular trend, matching the time effectiveness of the temporary sensitive word, the problem that the traditional static time effectiveness judgment cannot adapt to the dynamic changes of the hot event is optimized;

[0129] At the same time, by setting the recession baseline and the relapse trigger baseline, combining the continuous period stability, interaction quality comparison and comprehensive trend value to verify the relapse trend, the short-term fluctuation and substantial relapse are effectively distinguished, the accidental heat rise in the recession period is avoided from being misjudged as relapse, or the real relapse is ignored, the time effectiveness evaluation of the temporary sensitive word is consistent with the actual propagation state of the event, and a scientific basis is provided for subsequent rule adjustment.

[0130] Indirectly, the dynamic allocation of review resources concentrates resources to strengthen review during the escalation phase of an event (high timeliness) and reduces unnecessary intervention during the decline phase (low timeliness), thereby reducing the waste of computing power and manpower.

[0131] Step 4: Based on the timeliness assessment results of temporary sensitive words, the review intensity of temporary sensitive words will be adjusted in stages.

[0132] In some embodiments, when the temporary sensitive words are highly time-sensitive; highly time-sensitive means that the hot events associated with the temporary sensitive words are in a highly active period or have rapidly intensified after a resurgence, requiring the highest level of review, including but not limited to:

[0133] For social and livelihood issues or emergencies: If an incident occurs even once across the entire platform, it will be blocked, manually reviewed within 5 minutes, and the source tracing information will be recorded.

[0134] For example: After a chemical safety incident reignites, the temporary sensitive term "new leak in a certain factory area" (highly time-sensitive) is blocked if it appears once in Weibo or group chat. It is manually reviewed to confirm whether it is false information. If it is true, it is permanently blocked and traced back to its source.

[0135] Entertainment gossip event types: ≥2 occurrences of a single event trigger a notification; ≥4 occurrences of a single event trigger a blocking message.

[0136] When temporary sensitive words are of stable validity, a medium-level review needs to be initiated, including but not limited to:

[0137] Social and livelihood issues or emergencies: If there are ≥2 records across the entire platform, and the same record occurs within 1 hour, an alert will be sent;

[0138] Entertainment gossip: 3 or more occurrences of a single post trigger alerts; 6 or more occurrences result in blocking.

[0139] When temporary sensitive words have weak timeliness, a low-level review should be initiated, including but not limited to:

[0140] All types of events: Only events that occur frequently and in a concentrated manner are recorded, such as ≥10 times within 12 hours, and are manually reviewed within 24 hours; private chat scenarios are not interfered with.

[0141] Example 3

[0142] Based on the same inventive concept as the multi-dimensional identification method for sensitive words in online chat described in the foregoing embodiments, such as... Figure 3 As shown, this application provides a multi-dimensional identification system for sensitive words in online chat, wherein the system specifically includes:

[0143] Temporary sensitive word acquisition module: collects and analyzes multi-source data in real time, extracts core keywords to identify hot events, and extracts temporary sensitive words based on hot events;

[0144] Real-time collection of multi-source data such as social media, news websites, forums and instant messaging tools, after cleaning and word segmentation processing, the core keywords are extracted by using TF-IDF algorithm; by analyzing the interactive data containing the core keywords, the hot events meeting the absolute and relative interaction threshold are identified; based on the hot events, the semantic similarity and co-occurrence frequency are calculated by combining the pre-trained word vector model, and the temporary sensitive words meeting the correlation standard are screened out, which optimizes the lag problem of traditional word library to emerging sensitive words;

[0145] Chat auditing module: based on the obtained temporary sensitive words, a temporary word library is constructed, and scene-based effective rules are set for network chat auditing by combining the temporary word library and hot event labels;

[0146] Based on the extracted temporary sensitive words, a temporary word library is constructed, and for each sensitive word, an associated event label and a risk type label are added, and for a hot event, an event type, a propagation platform and other labels are added; by matching the associated event label of the temporary sensitive word with the type and propagation platform label of the hot event, scene-based auditing rules are set, such as strict interception triggered by social livelihood events, and relaxed reminder for entertainment gossip events, to reduce cross-scene misjudgment;

[0147] Time effectiveness analysis module: real-time monitoring of hotness data of hot events, analysis of the change trend of hotness data, and evaluation of the time effectiveness of temporary sensitive words based on the change trend of hotness data;

[0148] Real-time monitoring of hotness data of hot events, calculation of hotness comprehensive value and analysis of trend: through the first comprehensive trend value, the event is determined to be in rising, fluctuating or declining trend, corresponding to strong, stable and weak time effectiveness of temporary sensitive words; when the hotness is lower than the decline baseline (initial value 50%) and enters the decline period, and rises to the rekindling trigger baseline (decline baseline 30%), rekindling analysis is started, and through continuous period stability, interactive quality comparison and second comprehensive trend value verification, it is determined that the event is completely rekindled, declined rekindled or fluctuated rekindled, and the time effectiveness of sensitive words is dynamically adjusted;

[0149] Auditing intensity adjustment module: based on the evaluation results of the time effectiveness of temporary sensitive words, the auditing intensity of temporary sensitive words is adjusted;

[0150] According to the evaluation results of the time effectiveness of temporary sensitive words, the auditing intensity is adjusted: for strong time effectiveness, social livelihood events are intercepted and manually reviewed once on all platforms, and entertainment gossip events are reminded twice and intercepted four times; for stable time effectiveness (such as fluctuating trend), social livelihood events are recorded twice and sent reminders repeatedly, and entertainment gossip events are reminded three times and intercepted six times; for weak time effectiveness (such as decline period), only high-frequency concentrated events (such as 10 times in 12 hours) are recorded, and private scenes are not intervened, to balance the auditing accuracy and user experience.

[0151] The foregoing is considered as illustrative only of the principles of the application. Further, since numerous modifications and changes will readily occur to those skilled in the art, it is not desired to limit the application to the exact construction and practice described. Accordingly, all such variations and modifications are intended to be included within the scope of the application as defined in the following claims and the equivalents thereof.

Claims

1. A method for reviewing sensitive words in online chat based on multi-dimensional identification, characterized in that: Includes the following steps: Real-time collection of multi-source data and extraction of core keywords; acquisition of text containing core keywords and corresponding interactive data; identification of hot events by analyzing interactive data; extraction of temporary sensitive words based on hot events. A temporary thesaurus is built based on temporary sensitive words. The temporary thesaurus is combined with hot event tags to set scenario-based rules for effectiveness and to review online chat. Real-time monitoring of the popularity data of trending events, analysis and output of comprehensive popularity value, based on the dual dimensions of regular trend and resurgence trend to realize the change analysis of comprehensive popularity value, and evaluate the timeliness of temporary sensitive words based on the change analysis results of comprehensive popularity value; The process of conventional trend analysis is as follows: Extract the comprehensive heat index value for N consecutive time periods, calculate the deviation value of the comprehensive heat index value between adjacent time periods, obtain the sequence of comprehensive heat index value deviation values, analyze and calculate the sequence of comprehensive heat index value deviation values, and obtain the first comprehensive trend value. When the first comprehensive trend value is greater than or equal to the first judgment value, it is determined to be an upward trend, and the temporary sensitive words are highly time-sensitive. When the first comprehensive trend value is greater than the second judgment value and the first comprehensive trend value is less than the first judgment value, it is judged as a fluctuating trend, and the temporary sensitive word is stable timeliness; When the first comprehensive trend value is less than or equal to the second judgment value, it is judged as a downward trend, and the temporary sensitive words are weak in timeliness; The process of the first comprehensive trend value is as follows: Extract the deviation values ​​of the positive heat index composite value, count their number, and then calculate the ratio with the total number of deviation values ​​of the heat index composite value to obtain the directional consistency index. The deviation values ​​of the comprehensive heat index are weighted according to time and then weighted and summed to obtain the amplitude index; The weight of the Kth comprehensive heat index deviation value is K / (n-1), where n is the total number of comprehensive heat index deviation values; The maximum value of consecutive unidirectional deviations in the statistical heat index comprehensive value deviation sequence is used as an indicator of continuous trend. The directional consistency indicator, amplitude indicator, and continuous trend indicator are standardized and then fused to output the first comprehensive trend value. Based on the timeliness assessment results of temporary sensitive words, the review intensity of temporary sensitive words will be adjusted in stages.

2. The method for reviewing sensitive words in online chat based on multi-dimensional identification as described in claim 1, characterized in that: The process of extracting core keywords is as follows: The collected multi-source data was segmented into words; and keywords were extracted using the TF-IDF algorithm. TF is: the number of times a word appears in the current text / the total number of words in the text; IDF is: log(total number of documents / number of documents containing the word + 1); TF-IDF is: the product of TF and IDF; Extract words whose TF-IDF value ranks before the ranking percentage threshold as candidate keywords; for batch text, extract candidate keywords whose TF-IDF value is greater than or equal to the TF-IDF threshold as core keywords.

3. The method for reviewing sensitive words in online chat based on multi-dimensional identification as described in claim 1, characterized in that: The process of identifying hot events is as follows: Interaction data includes: number of likes, number of comments, number of reposts, number of shares, and number of favorites; Sum the number of likes, comments, reposts, shares, and favorites for all texts, and then average them to get the absolute interaction count. The ratio of the current period's interaction volume to the previous period's interaction volume is used as the relative interaction volume. Texts with both absolute and relative interaction volumes exceeding the limit are extracted and designated as trending events.

4. The method for reviewing sensitive words in online chat based on multi-dimensional identification as described in claim 1, characterized in that: The process of extracting temporary sensitive words is as follows: Extract the text of the core keywords corresponding to each hot event. For any core keyword, use a pre-trained word vector model to transform the core keyword and candidate words in the pre-processed text into multi-dimensional vectors. Calculate the semantic similarity between the vectors of all words in the pre-processed text and the vector of the core keyword using cosine similarity. Analyze the words that co-occur with the core keywords in the preprocessed text and count the number of times they co-occur. For any candidate word in the preprocessed text, the semantic similarity and the number of times they co-occur are summed and averaged to output a comprehensive relevance score. Candidate words with relevance scores exceeding the threshold are extracted and designated as temporary sensitive words.

5. The method for reviewing sensitive words in online chat based on multi-dimensional identification as described in claim 1, characterized in that: The process for obtaining the comprehensive heat value is as follows: Real-time monitoring of trending events' popularity data; popularity data includes: absolute interaction volume, relative interaction volume, and interaction quality score; After normalizing and standardizing the absolute interaction volume, relative interaction volume, and interaction quality value, the overall popularity value is integrated and output.

6. The method for reviewing sensitive words in online chat based on multi-dimensional identification as described in claim 5, characterized in that: The calculation process for the interaction quality value is as follows: Extract comments that exceed the average word count and count their number as the number of comments exceeding the word count limit. Calculate the ratio of the number of comments exceeding the word count limit to the total number of comments as the comment depth. Extract comments that have been replied to by other users and count their number as the number of secondary interaction comments. Calculate the ratio of the number of secondary interaction comments to the total number of comments as the secondary interaction percentage. The interaction quality score is output by combining the depth of comments with the proportion of secondary interactions.

7. The method for reviewing sensitive words in online chat based on multi-dimensional identification as described in claim 1, characterized in that: The analysis process for the reignition trend is as follows: When the overall heat index falls below the decline baseline, the heat trend is determined to have entered a decline phase; when the overall heat index rises back to the reignition trigger baseline, reignition analysis is initiated. After the overall heat index rises back to the re-ignition trigger baseline, it is determined whether it remains above the baseline for at least two consecutive time periods. If it only meets the standard for one period, it is judged as a short-term fluctuation and an invalid re-ignition. If it meets the standard for two or more consecutive periods, it proceeds to the next step of verification. Calculate the interaction quality value during the reignition period and compare it with the mean interaction quality value during the decay period; If the interaction quality value during the reignition period is greater than or equal to 1.2 times the average value during the decline period, it is considered a valid reignition; otherwise, it is considered an invalid reignition. Calculate the second comprehensive trend value after reignition; If it is an effective resurgence, and the second comprehensive trend value is greater than or equal to the first judgment value, then it is judged as a complete resurgence, and the temporary sensitive words are highly time-sensitive; If it is an effective resurgence, and the second comprehensive trend value is less than or equal to the second judgment value, it is judged as a recession resurgence, and the temporary sensitive words still maintain weak timeliness; If it is an invalid resurgence, and the second comprehensive trend value is greater than the second judgment value and the second comprehensive trend value is less than the first judgment value, then it is judged as a fluctuating resurgence, and the temporary sensitive word is stable timeliness.

8. A multi-dimensional identification system for sensitive words in online chat, characterized in that, The system is used to perform the method according to any one of claims 1-7, the system comprising: Temporary sensitive word acquisition module: Real-time collection of multi-source data and extraction of core keywords, acquisition of text containing core keywords and corresponding interactive data, identification of hot events by analyzing interactive data, and extraction of temporary sensitive words based on hot events; Chat moderation module: A temporary thesaurus is built based on temporary sensitive words. The temporary thesaurus is combined with hot event tags to set scenario-based rules for effect and to review online chat. Timeliness Analysis Module: Monitors the popularity data of trending events in real time, analyzes and outputs a comprehensive popularity value, realizes the change analysis of the comprehensive popularity value based on the dual dimensions of regular trend and resurgence trend, and evaluates the timeliness of temporary sensitive words based on the change analysis results of the comprehensive popularity value; Review intensity adjustment module: Based on the timeliness assessment results of temporary sensitive words, the review intensity of temporary sensitive words is adjusted in stages.

Images