A Model-Based Intelligent Monitoring Method and System for Online Live E-commerce

Abstract

The present invention discloses a model-based intelligent monitoring method and system for network live e-commerce. By integrating a method for text recognition in video images based on an improved CRNN algorithm, a speech recognition method based on a Hidden Markov Model (HMM) and a Gaussian probability intensity function, and a semantic analysis method for large language models based on the Transformer algorithm, the characteristics of illegal, disciplinary, and irregular behaviors in network live broadcasts are summarized and an operation supervision / monitoring model is constructed. Through model matching, illegal, disciplinary, and irregular behaviors in network live broadcasts are identified in a timely manner, thereby providing an intelligent and automated monitoring method and system for the market supervision and management department for network live e-commerce.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent computing technology, and specifically relates to a model-based online live broadcast e-commerce intelligent monitoring method and system. Background Art

[0002] While the booming development of live e-commerce has brought huge social and economic benefits, it has also brought unprecedented new challenges to government governance and market monitoring / regulation. Cases of illegal, disciplinary and irregular publicity and advertising on online platforms and online operations (the term "violation" is used uniformly in this paper) and consumer rights protection cases have increased year by year. However, due to the real-time, virtual, cross-regional and diverse characteristics of online live broadcasting, as well as its extremely fast transmission speed and wide coverage, traditional monitoring methods are difficult to meet the growing monitoring needs.

[0003] On the other hand, my country's artificial intelligence technologies, including machine vision technology, speech recognition technology, and natural language processing technology, are booming. Intelligent analysis technologies represented by face recognition, vehicle recognition, event monitoring, image retrieval, binocular stereo vision technology, and video splicing technology are widely used in government monitoring and security industries. Speech recognition technology is fully applied in life services, intelligent customer service and other fields. Key technologies such as semantic analysis and large language models are also booming in government affairs and digital economy. However, the monitoring / regulatory applications of technologies such as machine vision text extraction, speech recognition text conversion, semantic analysis, and large language models in the fields of live e-commerce and online advertising have not been fully developed. Problems such as the difficulty in identifying illegal, disciplinary, and regulatory content, the difficulty in obtaining evidence, and the insufficient construction of monitoring / regulatory business models have not been resolved, and cannot meet the growing monitoring needs. Summary of the invention

[0004] In response to the above problems, the present invention constructs a reasonable monitoring model and conducts real-time or quasi-real-time analysis of live network video information to promptly discover or verify illegal, disciplinary and regulatory violations, providing technical and information support for timely disposal.

[0005] To achieve the above purpose, the present invention designs a model-based online live broadcast e-commerce intelligent monitoring method, which monitors online live broadcast behaviors by building a reasonable monitoring model and using intelligent recognition technology. The intelligent monitoring method includes the following steps:

[0006] S1. Collection of basic data. The basic data includes live broadcast platform data, live broadcast organization data or online e-commerce data, anchor data, complaint and reporting data and / or public opinion monitoring data. The collection of basic data of monitored objects is a basic and common task for intelligent monitoring of online live broadcast e-commerce.

[0007] S2. Monitor model construction and maintenance: Based on the behaviors explicitly prohibited by national or local government laws, regulations, and rules, convert the rule descriptions in natural language into mathematical logic expressions, and establish a monitor model that can be automatically executed by a computer, that is, the matching rules for illegal and disciplinary (collectively referred to as violations in this invention) live streaming behaviors, and maintain the monitor model in real time according to new rules or public order and good customs, that is, add new matching rules or modify and improve the existing matching rules;

[0008] The monitor model includes at least one of the illegal promotion model (B-1), the prohibited and restricted sales model (B-2), the counterfeit and shoddy model (B-3), the unfair price behavior model (B-4, also known as the price violation model), the false publicity model (B-5), etc.;

[0009] There is no order of precedence between steps S1 and S2;

[0010] S3. Sample the live streaming video, and the sampling results include video frames and / or audio frames. The sampling methods include random sampling, directional sampling, patrol sampling, or daily comprehensive sampling, and the daily comprehensive sampling includes the combination of random sampling and directional sampling;

[0011] S4. Intelligent identification: Through model matching and calculation of image text recognition, speech recognition, and semantic analysis in the live streaming video, complete the identification of abnormal behaviors in the live streaming e-commerce; The abnormal behaviors include illegal promotion behaviors, sales of prohibited goods behaviors, counterfeit and shoddy behaviors, unfair price competition behaviors, false publicity behaviors, etc.

[0012] S5. Evidence preservation and fixation: Include video recording for archiving or image capture, record information such as the live streaming room name, live streaming video address, capture time, the number of online people at the capture time, and illegal content, etc., for archiving, and form a violation record;

[0013] S6. Clue handling: Include recording, summarizing, and displaying violation information, sending warning signals, reminding for manual handling, etc.

[0014] Furthermore, the live streaming platform data includes the unified social credit code, enterprise name, platform name, etc.;

[0015] The live streaming agency data includes the unified social credit code, enterprise name, etc.;

[0016] The online e-commerce data includes the unified social credit code, enterprise name, online business filing information, online store address, names of mandatory certification products, certification information, etc., as well as geographical indication product information, geographical indication product certification information, etc.;

[0017] The anchor data includes the live streaming platform name, live streaming room information, personal information of the anchor (including information such as name, ID number, permanent residence, etc.), etc.

[0018] The complaint and reporting data includes information such as the online live broadcast address of the complaint target and the complaint content information;

[0019] The public opinion monitoring data includes information such as the live broadcast room information and the public opinion content information;

[0020] The monitoring model further includes an obviously unfair model and / or an abnormal order placement model.

[0021] Furthermore, the illegal promotion model includes at least one of the following models:

[0022] (1) Identification formula for marketing food to minors without a bottom line: (There exists a violation word in the live broadcast promotion audio and video that belongs to the sample package of minor characteristic words) ∩ (There exists a violation word in the live broadcast promotion audio and video that belongs to the sample package of spoof and vulgar characteristic words), where ∃ represents existence, ∈ represents belonging to, and ∩ represents the intersection (referring to the data set, the same below);

[0023] (2) Identification formula for marketing using major political events and activities: (There exists a violation word in the live broadcast promotion audio and video that belongs to the sample package of major event characteristic words) ∨ (The product picture logo = the official logo of the major event); where ∨ represents the union;

[0024] (3) Identification formula for marketing that violates public order and good customs: There exists a violation word in the live broadcast promotion audio and video that belongs to {x|x is a public order and good customs characteristic word};

[0025] (4) Identification formula for violating the seven-day no-reason return and exchange policy: (There exists a violation word in the live broadcast promotion audio and video that belongs to {x|x is a seven-day no-reason return and exchange characteristic word}) ∨ (There exists a violation word in the live broadcast promotion audio and video that belongs to {x|x is a non-returnable and non-exchangeable characteristic word});

[0026] (5) Identification formula for illegal final interpretation right: There exists a violation word in the live broadcast promotion audio and video that belongs to {x|x is an interpretation right characteristic word}.

[0027] Furthermore, the prohibited and restricted sales model includes at least one of the following models:

[0028] (1) Identification formula for the sales behavior of products that do not meet the national standards and industry standards for ensuring human health and personal and property safety: The product quality standards shown in the live broadcast ∉ {x|x is a national mandatory standard, a national recommended standard, an industry standard}, where ∉ represents not belonging to;

[0029] (2) Identification formula for the sales behavior of adulterating, passing off fakes as genuine, passing off inferior goods as good ones, or passing off unqualified products as qualified products in products: (The product logo = the brand registered trademark) ∩ (The actual selling price of the product < the official price of the same product brand * k1); where the price comparison coefficient k1 of the same product ranges from 10% to 50%;

[0030] (3)Identification formula for the sales behavior of products explicitly prohibited by the state: (There exists a violation word in the live promotion audio / video ∈ {x|x is a characteristic word of prohibited or restricted sales commodities}) ∨ (There exists a violation word in the live promotion audio / video ∈ {x|x is a characteristic word of phased-out products});

[0031] (4)Identification formula for the sales behavior of products with forged place of origin, products with forged or misappropriated factory name and address of others, and products with forged or misappropriated quality marks such as certification marks: (The region of product promotion ∉ the region approved by the mark) ∨ (The actually sold product ≠ the product filed for the certification mark).

[0032] Furthermore, the counterfeit and shoddy model includes at least one of the following models:

[0033] (1)Identification formula for the behavior of fabricating the place of origin of products protected by geographical indication: The region of product promotion ∉ the region approved by the mark;

[0034] (2)Identification formula for the sales behavior of products infringing the exclusive right to use a registered trademark: (The product LOGO ≈ the brand registered trademark) ∩ (The product category ∈ the categories approved for the registered trademark).

[0035] Furthermore, the unfair price behavior model (also known as the price violation model) includes at least one of the following models:

[0036] (1)Core formula for price violations in e-commerce live streaming: The deviation degree F of the live streaming sales price from the fair price = (Current price - Fair price) / Fair price. Further judgment based on the formula is made to determine whether it belongs to price gouging, reasonable price, or unfair competition using a monopoly position;

[0037] (2)Data determination for "special price, wholesale price, factory price, market lowest price" in live streaming: The original price and comparison price data volume = 0;

[0038] (3)Live streaming limited-time promotion and limited-time special price: The limited-time data = 0;

[0039] (4)Live streaming buy-one-get-one-free and complimentary gifts: The gift data = 0;

[0040] (5)Improper formula for marking the original price in live streaming: The actual price of the product < the original price & the sales volume data in the past seven days = 0;

[0041] (6)Formula for live streaming promotion activities:

[0042] Price increase formula of first raising and then raising: The live streaming activity price ≥ the platform's re-purchase price > the platform's first-purchase price;

[0043] Price increase formula of first raising and then decreasing: The platform's re-purchase price > the live streaming activity price > the platform's first-purchase price;

[0044] First decrease then increase price formula: Initial purchase price > Re - purchase price and promotional price > Re - purchase price;

[0045] False price decrease formula: Re - purchase price > Promotional price = Initial purchase price;

[0046] False discount formula: Re - purchase price > Initial purchase price > Promotional price;

[0047] Lottery - type prize - giving sales formula: Price of the gift ≥ k2, where k2 is the price threshold of the gift.

[0048] Furthermore, the false publicity model at least includes one of the following models:

[0049] (1) Origin - fiction identification formula: Live - broadcast - claimed origin data > Known origin data of the commodity;

[0050] (2) Performance - fiction identification formula: Live - broadcast - publicized performance data > Known performance standard of the commodity;

[0051] (3) Use - fiction identification formula: Live - broadcast - publicized use of the commodity > Known use of the commodity or content of government - approved certification;

[0052] (4) Honor - fiction identification formula: Live - broadcast - claimed honor of the commodity > Data in the relevant honor database;

[0053] (5) False promotion identification formula, live - broadcast false "special price, wholesale price, factory price, market lowest price" formula: Quantity of compared price data = 0, or price of the live - broadcast commodity ≥ Compared price data;

[0054] (6) Live - broadcast false time - limited promotion, false time - limited special price: Time - limited period data = 0, or time - limited period > Publicized period;

[0055] (7) Original - price - fiction identification formula: Live - broadcast - claimed original price > Historical price of the same commodity on the e - commerce platform;

[0056] (8) License - fiction identification formula: Live - broadcast - claimed approved commodity > Data in the administrative license database;

[0057] (9) Patent - fiction identification formula: Live - broadcast - claimed patent content > Data in the patent database.

[0058] Furthermore, the random sampling includes grabbing a certain number of video frames from past videos or real - time broadcast videos according to random rules; the directional sampling includes sampling by searching and matching video sources according to market or public complaint and reporting information or abnormal information in public opinion monitoring; the patrol sampling includes randomly checking a certain number or fully checking all video frame information of each subject in sequence according to the registered information of live - broadcast platforms, live - broadcast agencies, online e - commerce platforms, or live - streamers, etc.

[0059] Furthermore, step S4 includes:

[0060] S41. Text recognition in video images, including a method for text recognition in video images based on an improved CRNN algorithm;

[0061] S42. Speech recognition, including a speech recognition method based on a hidden Markov model and a Gaussian probability intensity function;

[0062] S43. Semantic analysis, which performs semantic analysis on the text recognition results and / or speech recognition results in video images. The semantic analysis method includes a large language model semantic analysis method based on the Transformer algorithm, using word-character joint embedding vectors to establish connections between words sharing the same characters; adding an Attention structure at the word vector level (the Attention structure mainly refers to the attention mechanism in deep learning, which allows the model to dynamically focus on the most relevant information part for the current task when processing sequential data, and this mechanism has been widely applied in fields such as natural language processing and machine translation), adding context relationships at the word granularity while capturing context relationships at the character granularity; after encoding each Token after fusion, performing mean pooling (Pooling-Mean), and then connecting to a classifier for prediction;

[0063] S44. Model matching and calculation (C-4), which performs a correlation match between the semantic analysis results and the model constructed in step S2, or a correlation match with the dataset constructed for the corresponding model, or a conditional match with the training evaluation function constructed for the model, so as to calculate the corresponding match result, that is, to give a conclusion on whether there is a violation and which rule is violated.

[0064] Further, the evidence preservation and fixation in step S5 includes the record storage of the violating entity, violating keywords, and violating matching model information; the violating entity includes at least one of a live broadcast agency, an online e-commerce, and a live streamer.

[0065] Furthermore, the intelligent monitoring method includes a method for identifying similar illegal, disciplinary, and irregular behaviors based on dynamic association (which can be denoted as step S5a); the dynamic association includes analyzing whether there is an association with past violation records for each new violation record. Generally, only association analysis is performed on similar types of violation records (referring to the same violation record matching model), and there is no need to associate different types of violation records; the objects of the dynamic association include at least one of violation subject association, video source association, commodity association, and video content association. The video content association includes at least one of image matching association, text matching association, and voice matching association. The commodity association refers to the same commodity or essentially the same commodity or associated commodities sold in a package, etc. The associated commodities sold in a package include the situation where if any one or several key commodities are missing, the functions of other commodities cannot be exerted or reflected; the calculation techniques for the matching associations are all existing technologies, and the present invention uses the correlation coefficient (or degree of association) to characterize the strength of the association characteristics;

[0066] The method for identifying similar illegal, disciplinary, and irregular behaviors includes a suspected similar identification method (i.e., suspected similar illegal, disciplinary, and irregular behaviors). The suspected similar identification method includes that the correlation coefficient of at least one of the violation subject, the involved commodity (according to the calculation result of the commodity association), or the video source is not lower than the third threshold K3, where K3≥90%, and the correlation coefficient of at least one of the image matching association, text matching association, and voice matching association is not lower than the fourth threshold K4, where K4≥80%.

[0067] Furthermore, the method for identifying similar illegal, disciplinary, and irregular behaviors also includes a confirmed similar identification method. The confirmed similar identification method includes that the product of the correlation coefficients of at least 2 of the image matching association, text matching association, and voice matching association is greater than the fifth threshold K5, or the product of the three correlation coefficients is greater than the sixth threshold K6, where K5>K6. At this time, it is determined that two videos from different sources are closely related, that is, they belong to similar illegal, disciplinary, and irregular behaviors. Of course, the matching degree of videos from the same source is the highest, and it is not very meaningful to perform a correlation analysis because it is essentially the same thing.

[0068] On the other hand, a model-based intelligent monitoring system for live streaming e-commerce includes a sampling module, a model set, and an evidence preservation and fixation module that are respectively connected to an intelligent recognition module, and a clue handling module connected to the evidence preservation and fixation module. The sampling module is respectively connected to an external input module and a sampling rule module; the intelligent recognition module includes a video slicing unit, a picture and text recognition unit, a semantic recognition unit, and a model matching unit that are sequentially connected, and another branch including a voice recognition unit, a text extraction unit, a semantic recognition unit, and a model matching unit that are sequentially connected. Among them, the semantic recognition unit and the model matching unit are shared; the clue handling module includes a clue preliminary review unit, a handover supervision unit, and a handling result unit; the model set includes at least one of a violation promotion model, a prohibited and restricted sales model, a counterfeit and shoddy model, an unfair price behavior model, and a false publicity model; the external input module includes at least one of live streaming platform data, live streaming agency data, online e-commerce data, and anchor data, and also includes complaint and reporting data and / or public opinion monitoring data.

[0069] Further, the monitoring system implements the model-based intelligent monitoring method for live streaming e-commerce described in any one of the above.

[0070] The advantages and beneficial effects of the present invention are as follows: By integrating the method for text recognition in video images based on the improved CRNN algorithm, the speech recognition method based on the Hidden Markov Model (HMM) and the Gaussian probability intensity function, and the semantic analysis method of the large language model based on the Transformer algorithm, combined with the newly constructed business supervision / monitoring model for matching, illegal and disciplinary violations are identified, solving problems such as difficult identification and difficult evidence fixation of illegal content, significantly reducing the workload and difficulty of manual supervision; at the same time, through the design and application of the grossly unfair model and the abnormal order placement model, the universality of the monitoring method and system is improved, and the fair business environment of online e-commerce is further reasonably maintained; moreover, the distributed architecture design and the productized functional module design can adapt to the supervision needs of different scales and different scenarios. Description of the Drawings

[0071] Figure 1 It is a schematic diagram of the principle process of the model-based intelligent monitoring method and system for live streaming e-commerce;

[0072] Figure 2 It is a schematic diagram of an example of live streaming video picture and voice text conversion and semantic analysis. Detailed Embodiments

[0073] The following combines the drawings and embodiments to further describe the specific embodiments of the present invention. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and cannot be used to limit the protection scope of the present invention.

[0074] Example 1:

[0075] As Figure 1 shown, the present invention designs a model - based intelligent monitoring method for online live - streaming e - commerce. By constructing a reasonable monitoring model and using intelligent recognition technology to monitor online live - streaming behaviors, the intelligent monitoring method includes the following steps:

[0076] S1. Basic data collection. The basic data includes live - streaming platform data, live - streaming agency data, or online e - commerce data, and also includes anchor data, as well as complaint and reporting data and / or public opinion monitoring data, etc. Collecting the basic data of monitoring objects is a basic and common task for the intelligent monitoring of online live - streaming e - commerce.

[0077] The basic data collection includes data source management and data governance.

[0078] The data governance includes data cleaning, data comparison, and data association based on the collected data source data. Data governance technologies are all existing technologies.

[0079] S2. Monitoring model construction and maintenance. According to the behaviors explicitly prohibited by national or local government laws, regulations, and rules, convert the rule descriptions in natural language into mathematical logic expressions, and establish a monitoring model that can be automatically executed by a computer, that is, the matching rules for illegal, disciplinary, and irregular (collectively referred to as irregular in the present invention) online live - streaming behaviors, and the monitoring model can be maintained in real - time according to new rules or public order and good customs, that is, adding new matching rules or modifying and improving the existing matching rules.

[0080] The monitoring model includes at least one of an irregular promotion model (B - 1), a prohibited and restricted sales model (B - 2), a counterfeit and shoddy model (B - 3), an unfair price behavior model (B - 4, also known as a price violation model), a false publicity model (B - 5), etc. The naming of different models is only based on the customary naming of illegal, disciplinary, and irregular behaviors. In fact, there may be overlaps in behaviors for different names, that is, a certain behavior may simultaneously meet the recognition rules of two or more models.

[0081] The present invention converts the traditional natural language descriptions of violations of laws, disciplines, and regulations into strict mathematical logic expressions, making it possible to transform manual analysis and judgment into direct computer analysis and judgment, which is a prerequisite for realizing intelligent monitoring. The proposal of such a model and the subsequent model establishment method are both original to the present invention. The model architecture is very convenient for the later maintenance of the monitoring system because it is closer to the form of natural language expression and mainly conducts analysis and judgment from the perspective of established rules or habitual cognitions. Other existing live data monitoring methods mainly conduct analysis and judgment based on the results of learning, which requires a sufficient number of learning cases. Generally, these learning cases need to be manually intervened, and the learning system needs to be informed of its probable judgment results so as to improve the accuracy of system monitoring. Its adaptability to new rules is not strong. Every time a new rule appears, the system needs to have a sufficient number of cases for learning. Compared with the model establishment method of the present invention that only needs to convert the rules into logical expressions, the workload and complexity of the present invention are significantly lower, that is, the system is easier to understand and maintain.

[0082] There is no order of precedence between step S1 and step S2;

[0083] S3. Sample the network live video. The sampling results include video frames and / or audio frames. The sampling methods include random sampling, directional sampling, patrol sampling, or daily comprehensive sampling. The daily comprehensive sampling includes the combination of random sampling and directional sampling. In this embodiment, the daily comprehensive sampling method combining random sampling and directional sampling is adopted.

[0084] S4. Intelligent identification. Through model matching and calculation by means of image text recognition, speech recognition, and semantic analysis in the live video, the screening of abnormal behaviors in the network live e-commerce is completed. The abnormal behaviors include illegal promotion behaviors, behaviors of selling prohibited goods, counterfeiting and shoddy behaviors, unfair price competition behaviors, false publicity behaviors, etc.

[0085] This embodiment mainly integrates the method of text recognition in video images based on the improved CRNN algorithm, the speech recognition method based on the hidden Markov model (HMM) and the Gaussian probability intensity function, and the semantic analysis method of the large language model based on the Transformer algorithm to conduct video live content recognition and model calculation.

[0086] S5. Evidence preservation and fixation, including video recording and archiving or image capture, and forming a violation record;

[0087] Record, archive or download the live video suspected of violating laws, disciplines or regulations, or capture images. The violation record also includes associated information such as the subject of the violation, violation keywords, violation matching model, video source, network transfer information, download volume, click volume, etc., to solidify evidence. This embodiment mainly records information such as the live room name, live video address, capture time, number of online people at the capture time, and violation content, and archives it.

[0088] S6. Clue disposal, including recording, summarizing, and displaying violation information, sending warning signals, and reminding manual disposal, etc.

[0089] According to the governing unit of the regulatory object, classify and grade the clues for assignment, and give early warnings, guidance, interviews and case-filing investigations respectively in accordance with the law. The disposal results are fed back to the data source department in real time and recorded on file.

[0090] Preferably, the live platform data includes the unified social credit code, enterprise name, platform name, etc.;

[0091] The live agency data includes the unified social credit code, enterprise name, etc.;

[0092] The online e-commerce data includes the unified social credit code, enterprise name, online business record-filing information, online store address, names of mandatory certification products, certification information, etc., as well as geographical indication product information, geographical indication product certification information, etc.;

[0093] The anchor data includes the live platform name, live room information, personal information of the anchor (including information such as name, ID number, permanent residence, etc.);

[0094] The complaint and report data includes the online live address information of the complaint object, complaint content information, etc.;

[0095] The public opinion monitoring data includes the live room information, public opinion content information, etc.;

[0096] The monitoring model further includes an unconscionable model and / or an abnormal order placement model. The formula of the unconscionable model includes: ((the actual selling price of the product < n1 * the cost price of the product) & (the quantity of orders placed ≥ the lower limit of order placement n2) & (the activity has no time limit or the purchase quantity is unlimited)), and / or ((the cost price - the actual selling price) * the quantity of orders placed ≥ the platform profit of the merchant for the whole year); the formula of the abnormal order placement model includes: (the total quantity of orders placed within the unit time interval t1 ≥ the upper limit of order placement), and the upper limit of order placement can be set as n3 * the average daily order quantity, such as (the total quantity of orders placed within 1 minute ≥ 10 * the average daily order quantity), or (the total quantity of orders placed within 10 minutes ≥ 30 * the average daily order quantity), or (the total quantity of orders placed within 60 minutes ≥ 60 * the average daily order quantity), or the combined use of multiple similar formulas above, etc.; where n1 is the unconscionable coefficient, and in this embodiment, n1 = 0.1 is set, n2 is the lower limit of order placement, and in this embodiment, n2 = 5 is taken, n3 is the order quantity limit coefficient, t1 is the order quantity statistics interval, as shown in the above example, different combinations can be taken for t1 and n3, and specific system maintenance personnel can limit them according to factors such as regional income level, consumption ability, or product category.

[0097] In this embodiment, the unconscionable model and the abnormal order placement model are set mainly to address the possible large losses beyond the merchant's tolerance that may be caused by incorrect operations in merchant promotions, or the losses are significantly too large and far exceed the penalties that should be borne by the mistakes, that is, it is obviously unfair. The monitoring system needs to be able to detect and prove in a timely manner to maintain social fairness. This is the first creation of the present invention, and the current existing monitoring systems neither have this function nor effective analysis and judgment models and methods.

[0098] In this embodiment, through summarizing and generalizing the characteristics of current online live broadcast illegal, disciplinary, and irregular behaviors, a series of business supervision / monitoring models are constructed.

[0099] Preferably, the illegal promotion model (B-1, the program design number, the same below) includes at least one of the following models:

[0100] (1) The identification formula for bottomless marketing of food to minors: (the live broadcast promotional audio and video ∃ illegal words ∈ the sample package of minor characteristic words) ∩ (the live broadcast promotional audio and video ∃ illegal words ∈ the sample package of spoof and vulgar characteristic words), where ∃ means there exists, ∈ means belongs to, and ∩ means intersection (referring to the data set, the same below); the minor characteristic words can be dynamically updated;

[0101] (2) The identification formula for marketing using major political events and activities: (the live broadcast promotional audio and video ∃ illegal words ∈ the sample package of major event characteristic words) ∨ (the product picture LOGO = the official logo of the major event); where ∨ means union; the major event characteristic words can be dynamically updated;

[0102] (3)Marketing identification formula for violating public order and good customs: Live broadcast promotional audio and video ∃ illegal words ∈ {x|x is a characteristic word of public order and good customs}; the characteristic words of public order and good customs can be dynamically updated;

[0103] (4)Marketing identification formula for violating the seven-day no-reason return and exchange: (Live broadcast promotional audio and video ∃ illegal words ∈ {x|x is a characteristic word of seven-day no-reason return and exchange}) ∨ (Live broadcast promotional audio and video ∃ illegal words ∈ {x|x is a characteristic word of non-return and non-exchange}); the characteristic words of seven-day no-reason return and exchange include once sold, no return, no return, non-refundable deposit, pre-sale products not applicable to seven-day no-reason return, etc., and the characteristic words of non-return and non-exchange include non-return and non-exchange, non-refundable, non-refundable or non-exchangeable, prizes and gifts are not subject to three guarantees, etc., and the characteristic words can be dynamically updated;

[0104] (5)Marketing identification formula for illegal final interpretation right: Live broadcast promotional audio and video ∃ illegal words ∈ {x|x is a characteristic word of interpretation right}; the characteristic words of interpretation right include the final interpretation right belongs to XX, etc.

[0105] Preferably, the prohibited and restricted sales model (B-2) includes at least one of the following models:

[0106] (1)Sales behavior identification formula for products that do not meet national standards and industry standards for ensuring human health and personal and property safety: Live broadcast display of product quality standards ∉ {x|x is a national mandatory standard, a national recommended standard, an industry standard}, where ∉ means not belonging to;

[0107] (2)Sales behavior identification formula for adulterating, passing off fakes as genuine, passing off inferior goods as good ones, or passing off unqualified products as qualified products in products: (Product LOGO = brand registered trademark) ∩ (Actual selling price of the product < Brand official price of the same product * k1); where the price comparison coefficient k1 of the same product ranges from 10% to 50%, and in this embodiment, k1 = 20%;

[0108] (3)Sales behavior identification formula for products explicitly prohibited by the state: (Live broadcast promotional audio and video ∃ illegal words ∈ {x|x is a characteristic word of prohibited and restricted sales goods}) ∨ (Live broadcast promotional audio and video ∃ illegal words ∈ {x|x is a characteristic word of eliminated products}); the characteristic words of eliminated products include clearance sales, company bankruptcy clearance sales, clearance special offers, etc.; the characteristic words can be dynamically updated;

[0109] (4)Sales behavior identification formula for products with forged product origin, products with forged or misappropriated factory names and addresses of others, and products with forged or misappropriated quality marks such as certification marks: (Product publicity area ∉ mark approval area) ∨ (Actual sold product ≠ certified mark record-filing product).

[0110] Preferably, the counterfeit and shoddy model (B-3) includes at least one of the following models:

[0111] (1) Identification formula for the act of fabricating the place of origin of goods protected by geographical indication: The region of origin publicized for the goods ∉ the region approved for the indication;

[0112] (2) Identification formula for the act of selling goods infringing the exclusive right to use a registered trademark: (The goods' LOGO ≈ the brand's registered trademark) ∩ (The goods' category ∈ the categories approved for the registered trademark); where the symbol "≈" can be judged by methods such as the image comparison matching degree exceeding 90% or 95%, etc.

[0113] Preferably, the unfair price behavior model (B-4, also known as the price violation model) includes at least one of the following models:

[0114] (1) Core formula for price violations in e-commerce live streaming: The deviation degree F of the live streaming sales price from the fair price = (Current price - Fair price) / Fair price. According to the formula, further judge whether it belongs to price gouging, reasonable price, or unfair competition using a monopoly position. In this embodiment, when F ≥ 3, it is determined as price gouging, and when F ≤ 0.2, it is determined as unfair competition using a monopoly position;

[0115] (2) Judgment of data on "special price, wholesale price, factory price, market lowest price" in live streaming: The original price and comparison price data volume = 0;

[0116] (3) Live streaming limited-time promotions and limited-time special offers: The limited-time data = 0;

[0117] (4) Live streaming buy-one-get-one-free and complimentary gifts: The gift data = 0;

[0118] (5) Unfair formula for marking the original price in live streaming: The actual price of the goods < original price & the sales volume data in the past seven days = 0;

[0119] (6) Live streaming promotion activity formula:

[0120] Price increase formula of first raising and then raising: The live streaming activity price ≥ the platform's re-purchasing price > the platform's first-purchasing price; The platform's first-purchasing price refers to the normal sales price before the activity, the activity price is the actual sales price collected during the activity, and the platform's re-purchasing price refers to the price of the goods collected again from the online store after the activity; The so-called first raising and then raising means that the merchant first raises the sales price label before holding the live streaming sales activity, and then raises the price again through bidding and other means during the activity to achieve the purpose of abnormal price increase and profit;

[0121] Price increase formula of first raising and then lowering: The platform's re-purchasing price > the live streaming activity price > the platform's first-purchasing price; That is, the product price is raised before the activity and lowered during and after the activity;

[0122] Price increase formula of first lowering and then raising: The first-purchasing price > the re-purchasing price and the activity price > the re-purchasing price; That is, the product price is lowered before the activity and raised during and after the activity;

[0123] False price reduction formula: Re-purchase price > activity price = first purchase price; that is, the price is the same before and during the event, and the price is increased after the event.

[0124] False discount formula: Re-purchase price > first purchase price > activity price; that is, after the marketing event ends, the marked price of the product is increased to achieve the purpose of false discount.

[0125] Lottery-based prize sales formula: Gift price ≥ k2, where k2 is the gift price threshold. In this embodiment, the gift price threshold k2 is set to 50,000 yuan.

[0126] Among them, items (2) to (6) are derivative formulas for price violations in e-commerce live broadcasts.

[0127] Preferably, the false publicity model (B-5) includes at least one of the following models:

[0128] (1) Origin fiction identification formula: Live broadcast claimed origin data > known origin data of the product.

[0129] (2) Performance fiction identification formula: Live broadcast promoted performance data > known performance standard of the product.

[0130] (3) Use fiction identification formula: Live broadcast product use promotion > known use of the product or government-approved certification content.

[0131] (4) Honor fiction identification formula: Live broadcast product claimed honor > data in the relevant honor database.

[0132] (5) False promotion identification formula, live broadcast false "special price, wholesale price, factory price, market lowest price" formula: Comparison price data volume = 0, or live broadcast product price ≥ comparison price data.

[0133] (6) Live broadcast false limited-time promotion, limited-time special price: Limited time period data = 0, or limited time period > publicized time period.

[0134] (7) Original price fiction identification formula: Live broadcast claimed original price > historical price of the same product on the e-commerce platform.

[0135] (8) License fiction identification formula: Live broadcast claimed approved product > data in the administrative license database.

[0136] (9) Patent fiction identification formula: Live broadcast claimed patent content > data in the patent database.

[0137] Preferably, the random sampling includes capturing a certain number of video frames from past videos or real-time transmitted videos according to random rules; the directional sampling includes sampling by searching for matching video sources based on market or public complaint and reporting information or abnormal information in public opinion monitoring; the patrol sampling includes sampling a certain number or all video frame information of each subject in sequence according to the registered information of live streaming platforms, live streaming agencies, online e-commerce or live streamers, etc., such as sampling or inspecting all in turn in alphabetical order; the random sampling, directional sampling and daily comprehensive sampling are essentially an active monitoring sampling method, while the directional sampling is a passive monitoring sampling method; in actual application, the ideal situation is to monitor all live videos, that is, to conduct a full inspection sampling analysis. Considering that it is difficult to achieve real-time sampling analysis of all network live streaming behaviors with the existing computing power, the necessity of full inspection is not very great, and the monitoring is mainly carried out by sampling; in engineering practice, the method of combining active monitoring sampling inspection and passive monitoring sampling is usually adopted, that is, daily sampling monitoring according to certain rules and directional sampling monitoring immediately when there is a report.

[0138] Preferably, step S4 includes:

[0139] S41. Text recognition in video images (C-1, program design number, the same below), including a method for text recognition in video images based on an improved CRNN algorithm;

[0140] Whether it is live streaming monitoring or advertising monitoring, text recognition in the video must be completed.

[0141] Based on the CRNN (Convolution Recurrent Neural Network) algorithm, text recognition of video graphics can be referred to the paper "An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition". It proposed a new neural network architecture that integrates feature extraction, sequence modeling and transcription into a unified framework. Tested on standard benchmark datasets including IIIT-5K, Street View Text and ICDAR datasets, it is state-of-the-art (optimal).

[0142] In this embodiment, a method for text recognition in video images based on an improved CRNN algorithm is adopted for video text recognition: Specifically, the video stream is processed into frames of pictures through OpenCV technology, and then a target detection model is used to detect a series of texts. For the detected texts, the texts are scaled to a unified size and normalized. Based on the principle of convolutional neural network, the stride parameter is optimized and adjusted to solve the problems of large width differences and dispersion of Chinese characters, and easy occurrence of text recognition errors, so as to accurately extract data features. Through the recurrent neural network, the traditional CRNN (Convolution Recurrent Neural Network) algorithm is improved into a bidirectional LSTM (Long Short-Term Memory, a common recurrent neural network Recurrent Neural Network, RNN) algorithm to solve the problems of gradient disappearance and long-term dependence existing in the traditional RNN algorithm.

[0143] S42. Speech recognition (C-2), including a speech recognition method based on a hidden Markov model and a Gaussian probability intensity function;

[0144] Both live broadcast or advertising videos contain a large amount of voice information. In live broadcasts, it is even more reliant on the continuous incitement of the host's emotions to achieve the effects of live e-commerce or user rewards. Therefore, speech recognition technology occupies an important position in this video monitoring. Generally, a speech recognition system is divided into three stages, namely the signal processing stage, the phoneme stage, and the decoding stage.

[0145] This embodiment adopts a speech recognition method based on a hidden Markov model (HMM) and a Gaussian probability intensity function:

[0146] Signal preprocessing: Noise suppression and noise removal are performed on the speech signal. Through methods such as energy analysis and zero-crossing rate detection, the effective speech segment of the speech signal is identified. First, the acoustic waveform in the video is segmented into phonetic frame (usually 10ms, 15ms or 20ms), and the phonetic frame is converted into a spectrogram feature.

[0147] Acoustic Modeling: The Hidden Markov Model (HMM) is used to describe the relationship between the hidden state sequence and the observable acoustic feature sequence in the speech signal. The hidden Markov model is a probability model for time series, which describes the process of randomly generating an unobservable state sequence (StateSequence) by a hidden Markov chain, and then generating an observation sequence (Observation Sequence) by each state. Each position in the sequence can be regarded as a moment. In this information processing stage, it can be used to model the acoustic observations (feature vectors) at the sub-word level (such as Chinese phonemes). Usually, 3 states are modeled for each phoneme, corresponding to the beginning, middle, and end of the phoneme respectively. Each state has a self-transition and a transition to the next state.

[0148] The commonly used method in this stage is to calculate the probability intensity function in a continuous space. The acoustic signal of speech is transformed through a series of signal processing transformations into a series of feature vectors, and each vector represents a time slice of the speech input. In this embodiment, statistical techniques such as the Gaussian probability intensity function are used to perform the above process calculation and attempt to identify individual phonemes.

[0149] Language Modeling: Language modeling is to deal with the problems of vocabulary and grammar structure in speech recognition. By establishing a language model, the rationality of different vocabulary sequences can be evaluated, so as to guide the decoding process to select a recognition result that is more in line with language habits. In the recognition stage, the feature sequence of the input speech signal is matched with the established HMM template, and metric indicators such as likelihood or distance are calculated.

[0150] Decoding and Output: Using the word pronunciation dictionary and the language model, the VITERBI algorithm is used to determine the word sequence with the highest probability for the given acoustic event. First, the Viterbi table is initialized, and a two-dimensional table is used to store the intermediate results, where the rows of the table correspond to different time steps (or observations), and the columns correspond to different hidden states. Then, the speech signal to be recognized is converted into a series of observations to form the input sequence processed by the algorithm. Starting from the second column, the language model is traversed for each slice, the cumulative probability is calculated, and the optimal matching result is retained as the final state, which is mapped to the corresponding output symbol sequence (usually words or phonemes in the vocabulary), and the final output is the recognition result of the speech recognition system.

[0151] S43. Semantic analysis (C-3), which performs semantic analysis on the text recognition result and / or speech recognition result in the video image. The semantic analysis method includes a large language model semantic analysis method based on the Transformer algorithm. On the basis of using traditional word vectors and / or character vectors, word-character joint embedding vectors are used simultaneously to establish connections between words sharing the same character; an Attention structure is added at the word vector level (the Attention structure mainly refers to the attention mechanism in deep learning, which allows the model to dynamically focus on the most relevant information part when processing sequential data, and this mechanism has been widely used in fields such as natural language processing and machine translation). While capturing the character-level context relationship, the word-level context relationship is added, and the establishment of the context relationship here is not only for the traditional adjacent context relationship, but the system first performs a correlation detection within the same limited-time video segment. When there are significantly relevant (such as a correlation coefficient greater than 90%) characters or words, the character-level context relationship and / or word-level context relationship across time intervals are captured and established; after encoding each Token after word-character fusion, mean pooling (Pooling-Mean) is performed, and then a classifier is connected for prediction; the present invention makes an application improvement to the large language model semantic analysis method based on the Transformer algorithm, which is more suitable for the phenomena of short and hasty tones and frequent semantic jumps existing in current webcasting (that is, the semantics of adjacent utterances are not necessarily closely related, and the semantics after a period of time may be closely related), which is one of the innovative points of the present invention.

[0152] The video scope in webcasting is relatively wide, and the models for different scenarios are different, and an effective model needs to be quickly constructed for screening. Especially for some sudden popular words of illegal advertisements, the model production speed is particularly crucial. Compared with other conventional algorithms, the model trained by the Transformer algorithm can be trained using a distributed GPU, which can improve the training efficiency. In addition, the entire live video is the host speaking, and the front-back relevance is not necessarily very strong. Sometimes the topic is digressed and then comes back for introduction, and it is necessary to capture the semantics with a longer interval for association. In this regard, that is, when analyzing and predicting longer texts, the algorithm based on Transformer captures the semantic association with a longer interval better.

[0153] Therefore, in this embodiment, the Transformer algorithm is selected for semantic analysis, as Figure 2 shown.

[0154] Text preprocessing: Convert the original text data into a format suitable for model processing, including removing irrelevant characters, stop words, and noisy data through data cleaning, splitting the text into smaller units through word segmentation, reducing the lexical diversity by lemmatization to the basic format, and achieving a broader morphological merger through stemming.

[0155] CLS token addition: Based on the CLS (Concept Learning System) specification, for each Token formed after the word integration of the text recognition result and / or the speech recognition result.

[0156] Text vectorization: Through the word embedding technology (Word Embedding), convert words or sentences into vector representations to capture their semantic information. Then encode each Embedding (encoder), and the encoding results form several related sequences (real-time feature words).

[0157] Feature extraction: Add an Attention structure (model library structure, including model feature words) at the word vector level. While capturing the character-level context relationship, add the word-level context relationship, and then use a classifier (Softmax Classifier, a general classifier) for classification, thus forming a semantic expression that is easy for a computer to recognize.

[0158] S44. Model matching and calculation (C-4): Match the extracted features with the model constructed in step S2, or perform a correlation match with the dataset constructed by the corresponding model, or perform a conditional match with the training evaluation function constructed by the model, so as to calculate the corresponding matching result, that is, give a conclusion on whether there is a violation and which rule is violated. The specific matching method is as Figure 2 shown. Perform a model matching operation on the encoding results forming several related sequences and the model feature words within the Attention structure (the ⊕ symbol in the figure represents the model matching operation). In this embodiment, the deep semantic matching model DSSM (Deep Structured Semantic Models) method is adopted. Use the extracted feature vector as a query, and calculate the cosine similarity with the keyword vectors in the model library constructed in B-1, B-2, B-3, B-4, B-5. If the chord value is closer to 1, that is, the included angle tends to 0°, it means the greater the similarity between the two vectors; the closer the cosine value is to 0, that is, the included angle tends to 90°, it indicates that the two vectors are less similar. In this embodiment, it is set that a chord value higher than 0.8 is used as a basic match, and this record is used as a clue for evidence preservation and fixation.

[0159] Preferably, the evidence storage in step S5 includes record storage of the violating subject, violating keywords and violating matching model information; the violating subject includes at least one of a live broadcast organization, an online e-commerce company and an anchor.

[0160] Embodiment 2:

[0161] The difference from Example 1 is that the intelligent monitoring method described in this embodiment also includes a method for identifying similar illegal, disciplinary and regulatory violations based on dynamic association (which can be recorded as step S5a). The similar illegal, disciplinary and regulatory violations include repeated or multiple similar illegal, disciplinary and regulatory violations, such as one person exaggerating the same product in different occasions, illegally promoting, or selling the same restricted product on different platforms, etc., which is equivalent to multiple and repeated violations, and the punishment should be increased, but the premise is that the monitoring system needs to grasp accurate information in a timely manner; the dynamic association includes an analysis of whether each new violation record is associated with a past violation record, and generally only the same type of violation records ( The dynamic association object includes at least one of the association of the violating subject, the association of the video source, the association of the commodity, and the association of the video content. The association of the video content includes at least one of the association of the image matching, the association of the text matching, and the association of the voice matching. The commodity association refers to the same commodity or essentially the same commodity or associated commodities sold in a package, etc. The associated commodities sold in a package include the lack of any one of the commodities or several related commodities, and the role of other commodities cannot be played or reflected. The matching association technologies are all existing technologies. The present invention uses the correlation coefficient (or correlation degree) to characterize the strength of the association characteristics.

[0162] The method for identifying similar illegal, disciplinary and regulatory violations includes a suspected similar identification method (i.e. suspected similar illegal, disciplinary and regulatory violations). The suspected similar identification method includes that at least one correlation coefficient among the violating entity, the involved commodities (based on the commodity association calculation results) or the video source is not lower than the third threshold K3, K3≥90%, and at least one correlation coefficient among the image matching association, text matching association and voice matching association is not lower than the fourth threshold K4, K4≥80%. In this embodiment, K3=98% and K4=90% are set; wherein, the correlation coefficient requirement of K3=98% is set mainly because it is determined that the violation is repeated multiple times and at least one of the anchor, live broadcast platform, online e-commerce, live broadcast organization, and product is the same or essentially the same; the correlation coefficient requirement of K4=90% is set mainly because it is considered that the video content association does not need to be expressed in the same or almost the same form from the perspective of image, text, voice, etc., because the form of expressing the same semantics in Chinese may vary greatly. Therefore, a requirement lower than the third threshold is set, but considering that only one of the image, text, and voice is required, it is reasonable to have a higher standard, that is, if the correlation coefficients of the three items of image, text, and voice are not high enough, it is reasonable to consider them as irrelevant. Therefore, in this embodiment, K4=90% is set.

[0163] Preferably, the method for identifying similar illegal, disciplinary and regulatory behaviors also includes a method for confirming similar identification, which includes that the product of at least two correlation coefficients of image matching association, text matching association and voice matching association is greater than the fifth threshold K5, or the product of three correlation coefficients is greater than the sixth threshold K6, K5>K6. Considering that any correlation coefficient of image content association should not be lower than 60%, and at least one correlation coefficient should not be lower than 90%, it is more conservative and generally K5≥0.6 and K6≥0.4 are taken. In this embodiment, K5=0.8 and K6=0.6 are taken. Of course, the larger the values ​​of the two parameters K5 and K6, the higher the credibility of the identification of similar illegal, disciplinary and regulatory behaviors, but this will also increase the probability of missed identification. If the values ​​of the two parameters are too small, the recognition success rate will increase (that is, the missed identification rate will decrease), but the probability of misidentification, i.e., false alarm, will also increase, which needs to be determined by the user according to the actual situation. At this time, it is judged that the two videos from different sources are closely related, that is, they belong to the same type of illegal, disciplinary and regulatory behaviors. Of course, the same-source video has the highest matching degree, and it is not meaningful to do correlation analysis because they are the same thing.

[0164] Although the calculation method of the correlation coefficient can adopt a relevance calculation method similar to that of a general search engine, establishing different correlation judgment criteria is still the first of its kind in the present invention. It converts some of people's ideas into strict mathematical expressions, which is convenient for computer execution, making the monitoring of online live e-commerce more intelligent and the warning information more timely and effective.

[0165] Identifying similar illegal, disciplinary and irregular behaviors mainly provides basis and information for the disposal and punishment of irregular behaviors, enabling timely detection and disposal of multiple violators and their behaviors; while storing and fixing evidence, the present invention also records information such as the number of downloads or views, the types of download or view subjects, the sales volume of related products, transaction prices, consumer groups, etc., and then analyzes the degree of network influence, the affected population, the amount of illegal benefits, etc., which also provides basis and information for disposal and punishment.

[0166] Embodiment 3:

[0167] The difference from Embodiment 1 is that in this embodiment, only targeted sampling is performed, that is, when there is a report or abnormal public opinion, targeted videos are sampled.

[0168] Embodiment 4:

[0169] A model-based intelligent monitoring system for online live e-commerce, the system includes a sampling module, a model set and an evidence storage and fixation module respectively connected to the intelligent recognition module, and a clue disposal module connected to the evidence storage and fixation module, the sampling module is respectively connected to an external input module and a sampling rule module; the intelligent recognition module includes a video slicing unit, a picture and text recognition unit, a semantic recognition unit and a model matching unit connected in sequence, and another branch voice recognition unit, a text extraction unit, a semantic recognition unit and a model matching unit connected in sequence, wherein the semantic recognition unit and the model matching unit are shared; the clue disposal module includes a clue preliminary review unit, a transfer to supervision unit and a disposal result unit connected in sequence; the model set includes at least one of an illegal promotion model, a prohibited and restricted sales model, a counterfeit and shoddy model, an unfair price behavior model, a false publicity model; the external input module includes at least one of live platform data, live agency data, online e-commerce data, and anchor data, and also includes complaint and report data and / or public opinion monitoring data.

[0170] Preferably, the model-based intelligent monitoring system for online live e-commerce implements at least one of Embodiments 1 to 3 and their preferred methods.

[0171] Constructing a suitable supervision / monitoring model is the key to the implementation of the present invention. It is necessary to summarize and generalize the characteristics of current online live illegal, disciplinary and irregular behaviors in combination with various laws, regulations, rules and public order and good customs to form an operation model that can be recognized by a computer. And with the development of society and the change of the situation, the monitoring model can be maintained in real time according to new rules or public order and good customs, that is, adding new matching rules or modifying and improving existing matching rules, so that the system has higher practical value.

[0172] The basic principle of the model-based online live e-commerce intelligent monitoring method designed by the present invention is to form a monitoring model that can be recognized by a computer by reasonably summarizing and generalizing the characteristics of current online live illegal, disciplinary and regulatory violations, and utilize a video image Chinese character recognition method based on an improved CRNN algorithm, a speech recognition method based on a hidden Markov model (HMM) and a Gaussian probability intensity function, and a large language model semantic analysis method based on a Transformer algorithm to parse and convert the real-time online, dynamically transformed video images and video speech of the live e-commerce into text that can be formatted and stored, perform semantic analysis through a large language model, and match the business monitoring model to identify illegal, disciplinary and regulatory violations, thereby solving the problems of difficulty in identifying illegal, disciplinary and regulatory content and difficulty in obtaining evidence for the content, and significantly reducing the workload and difficulty of manual supervision. At the same time, through distributed architecture design and productized functional design, it can adapt to monitoring / regulatory needs of different scales and scenarios.

[0173] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical principles of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention. For example, the number and combination of monitoring models can form several different embodiment combinations, and the third to sixth threshold setting selection in the dynamic association-based similar illegal, disciplinary and regulatory behavior identification method can also form several new embodiment combinations, and so on.

Claims

1. A model-based online live e-commerce intelligent monitoring method, characterized in that: By building a monitoring model and using intelligent recognition technology to monitor live broadcasting behavior on the Internet, the intelligent monitoring method includes the following steps: S1. Collection of basic data, including live broadcast platform data, live broadcast organization data or online e-commerce data, anchor data, complaint and reporting data and / or public opinion monitoring data; S2. Construction and maintenance of monitoring models. Based on the behaviors expressly prohibited by national or local government laws, regulations, rules and regulations, the rules description in natural language is converted into mathematical logic expressions to establish a monitoring model for automatic execution by computers; The monitoring model includes at least one of an illegal promotion model, a sales restriction model, a counterfeit and shoddy model, an unfair pricing behavior model, and a false advertising model; S3. Sampling live webcast videos, wherein the sampling method includes random sampling, directional sampling, inspection sampling or daily comprehensive sampling, wherein the daily comprehensive sampling includes a combination of random sampling and directional sampling; S4, intelligent recognition, through model matching and calculation of image and text recognition, voice recognition, and semantic analysis in live video, to identify abnormal behaviors of online live e-commerce; S5. Preserve evidence, including recording and archiving or capturing images, and form a record of violations; S6. Clue handling, including recording, summarizing, and displaying violation information, issuing warning signals, and reminding manual handling; There is no order of precedence between steps S1 and S2; Step S4 includes: S41, video image Chinese character recognition, including video image Chinese character recognition method based on CRNN algorithm; S42. Speech recognition, including speech recognition methods based on hidden Markov models and Gaussian probability intensity functions; S43, semantic analysis, performing semantic analysis on the text recognition results and / or speech recognition results in the video image, wherein the semantic analysis method includes a large language model semantic analysis method based on the Transformer algorithm, using word joint embedding vectors to create connections between words that share characters; adding an Attention structure at the word vector level to capture the word granularity context relationship while adding the word granularity context relationship; after encoding each Token after word fusion, performing mean pooling, and then connecting to a classifier for prediction; S44, model matching and calculation, matching the semantic analysis result with the model constructed in step S2 for relevance, thereby calculating the corresponding matching result; The intelligent monitoring method includes a method for identifying similar illegal, disciplinary and regulatory violations based on dynamic association, wherein the dynamic association includes, for each new violation record, an analysis of whether it is associated with a previous violation record; the object of the dynamic association includes at least one of violation subject association, video source association, commodity association and video content association, wherein the video content association includes at least one of image matching association, text matching association and voice matching association; The method for identifying similar illegal, disciplinary and regulatory violations includes a suspected similar identification method, which includes that at least one correlation coefficient among the violating subject, product or video source is not lower than a third threshold K3, K3≥90%, and at least one correlation coefficient among the image matching association, text matching association and voice matching association is not lower than a fourth threshold K4, K4≥80%.

2. According to the model-based online live e-commerce intelligent monitoring method of claim 1, it is characterized by: The live broadcast platform data includes unified social credit code, enterprise name, and platform name; The live broadcasting organization data includes unified social credit code and company name; The online e-commerce data includes unified social credit code, enterprise name, online business registration information, online store address, compulsory certification product name, and certification information; The anchor data includes the name of the live broadcast platform, live broadcast room information, and anchor personal information; The complaint report data includes the webcast address information and complaint content information of the complained party; The public opinion monitoring data includes live broadcast room information and public opinion content information; The monitoring model also includes an unfairness model and / or an abnormal order model.

3. According to the model-based online live e-commerce intelligent monitoring method of claim 1, it is characterized in that: The illegal promotion model includes at least one of the following models: (1) The formula for identifying food marketing with no bottom line for minors is: (live broadcast promotional audio and video ∃ illegal words ∈ minors’ characteristic word sample package) ∩ (live broadcast promotional audio and video ∃ illegal words ∈ spoof and vulgar characteristic word sample package), where ∃ means existence, ∈ means belonging, and ∩ means intersection; (2) Using the formula for identifying major political events and activities: (live broadcast promotional audio and video ∃ illegal words ∈ major event feature word sample package) ∨ (product image LOGO = official logo of major event); where ∨ represents a union; (3) Identification formula for marketing that violates public order and good morals: live broadcast promotional audio and video ∃violation word ∈ {x|x is a public order and good morals characteristic word}; (4) Violation of the seven-day unconditional return and exchange marketing identification formula: (live broadcast promotion audio and video ∃ illegal word ∈ {x|x is the seven-day unconditional return and exchange feature word}) ∨ (live broadcast promotion audio and video ∃ illegal word ∈ {x|x is the no-refund and no-exchange feature word}); (5) The formula for identifying marketing violations with the final right of interpretation is: live broadcast promotional audio and video ∃violation word ∈ {x|x is a characteristic word of the right of interpretation}.

4. According to the model-based online live e-commerce intelligent monitoring method of claim 1, it is characterized in that: The sales restriction model includes at least one of the following models: (1) The sales behavior identification formula of products that do not meet the national standards and industry standards for protecting human health and personal and property safety is: Live broadcast product quality standard ∉{x|x is a national mandatory standard, a national recommended standard, or an industry standard}, where ∉ means it does not belong; (2) The formula for identifying the sales behavior of adulterating or counterfeiting products, passing off fake products as genuine ones, passing off inferior products as good ones, or passing off substandard products as qualified ones is: (product LOGO = brand registered trademark) ∩ (actual product selling price < official brand price of the same product * k1); where the price comparison coefficient k1 of the same product ranges from 10% to 50%; (3) The formula for identifying sales behaviors of products that have been explicitly eliminated by the state is: (live broadcast promotional audio and video ∃ illegal word ∈ {x|x is a characteristic word for prohibited or restricted sales products}) ∨ (live broadcast promotional audio and video ∃ illegal word ∈ {x|x is a characteristic word for eliminated products}); (4) The sales behavior identification formula for products with falsified origin, products with falsified or counterfeit factory names and addresses, and products with falsified or counterfeit certification marks and other quality marks is as follows: (product advertised origin area ∉ mark approval area) ∨ (actual sales products ≠ certification mark registered products).

5. According to the model-based online live e-commerce intelligent monitoring method of claim 1, it is characterized by: The counterfeit and shoddy models include at least one of the following models: (1) Formula for identifying the fictitious origin of goods protected by geographical indications: the advertised origin of the goods ∉ the area where the geographical indication is approved; (2) The formula for identifying the act of selling goods that infringe the exclusive rights of a registered trademark is: (product LOGO ≈ brand registered trademark) ∩ (product category ∈ registered trademark approved category).

6. According to the model-based online live e-commerce intelligent monitoring method of claim 1, it is characterized in that: The improper price behavior model includes at least one of the following models: (1) The core formula for e-commerce live streaming price violations: The deviation between the live streaming price and the fair price F = (current price - fair price) / fair price. The formula is used to further determine whether it is price gouging, a reasonable price, or unfair competition using a monopoly position; (2) Live broadcast "special price, wholesale price, factory price, market lowest price" data determination: original price, comparison price data volume = 0; (3) Live limited-time promotion, limited-time special offer: limited-time period data = 0; (4) Buy one get one free live broadcast with free gifts: Free gift data = 0; (5) Improper formula for marking original prices in live broadcasts: actual price of the product < original price & sales data for the past seven days = 0; (6) Live broadcast promotion formula: The formula for raising the price first and then raising the price again is: live event price ≥ platform re-purchase price > platform first purchase price; The price increase formula is: platform repurchase price > live event price > platform first purchase price; The formula for price increase after price reduction is: first purchase price > re-purchase price and promotion price > re-purchase price; False price reduction formula: repurchase price > promotion price = first purchase price; False discount formula: repurchase price > first purchase price > promotion price; The formula for lottery-style sales with prizes is: gift price ≥ k2, where k2 is the gift price threshold.

7. According to the model-based online live e-commerce intelligent monitoring method of claim 1, it is characterized by: The false advertising model includes at least one of the following models: (1) Formula for identifying fictitious origin: the origin data claimed in the live broadcast > the known origin data of the product; (2) Formula for identifying fictitious performance: performance data promoted in live broadcast > known performance standards of the product; (3) Formula for identifying fictitious uses: Live broadcast product use promotion > known product uses or government-approved certification content; (4) Formula for identifying fictitious honors: the honors claimed by the live broadcast product > the relevant honor database data; (5) False promotion identification formula, live broadcast false "special price, wholesale price, factory price, market lowest price" formula: comparison price data volume = 0, or live broadcast product price ≥ comparison price data; (6) Live broadcast of false limited-time promotions or limited-time special offers: the limited-time data = 0, or the limited-time data > the promotional period; (7) Formula for identifying fictitious original prices: the original price claimed in the live broadcast > the historical price of the same product on the e-commerce platform; (8) Formula for identifying fictitious licenses: Live broadcast claims of approved products > administrative license database data; (9) Patent fictitious identification formula: patent content claimed in live broadcast > patent database data.

8. According to the model-based online live e-commerce intelligent monitoring method of claim 1, it is characterized by: The random sampling includes capturing a certain number of video frames from past videos or real-time videos according to random rules; the targeted sampling includes searching for matching video sources for sampling based on market or public complaint information or public opinion monitoring abnormal information; the inspection sampling includes spot-checking a certain amount of video frame information for each subject in sequence according to the registered live broadcast platform, live broadcast organization, online e-commerce or anchor information.

9. According to the model-based online live e-commerce intelligent monitoring method of claim 1, it is characterized in that: The evidence storage in step S5 includes the record storage of the violating subject, violating keywords and violating matching model information; the violating subject includes at least one of the live broadcast organization, online e-commerce and anchor.

10. According to the model-based online live e-commerce intelligent monitoring method of claim 1, it is characterized in that: The method for identifying similar illegal, disciplinary and regulatory violations also includes a method for confirming similar identification, which includes that the product of at least two correlation coefficients of image matching association, text matching association and voice matching association is greater than the fifth threshold K5, or the product of three correlation coefficients is greater than the sixth threshold K6, K5>K6. At this time, it is judged that two videos from different sources belong to the same type of illegal, disciplinary and regulatory violations.

11. A model-based online live e-commerce intelligent monitoring system, characterized in that: The monitoring system implements the model-based online live e-commerce intelligent monitoring method described in any one of claims 1 to 10, and the system includes a sampling module, a model set and a evidence storage module respectively connected to the intelligent recognition module, and a clue handling module connected to the evidence storage module, and the sampling module is respectively connected to the external input module and the sampling rule module; the intelligent recognition module includes a video slicing unit, a picture and text recognition unit, a semantic recognition unit and a model matching unit respectively connected in sequence, and another branch voice recognition unit, a text extraction unit, a semantic recognition unit and a model matching unit connected in sequence, wherein the semantic recognition unit and the model matching unit are shared; the clue handling module includes a clue preliminary review unit, a transfer supervision unit and a handling result unit connected in sequence; the model set includes at least one of an illegal promotion model, a prohibited and restricted sales model, a counterfeit and shoddy model, an unfair pricing behavior model, and a false propaganda model; the external input module includes at least one of live broadcast platform data, live broadcast organization data, online e-commerce data, and anchor data, and also includes complaint and reporting data and / or public opinion monitoring data.

Images