Live broadcast information analysis method based on knowledge graph

Abstract

The invention belongs to the field of Internet information processing, and particularly relates to a live broadcast information analysis method based on a knowledge graph, which comprises the following steps: acquiring a certain amount of various types of live broadcast data of a specified platform, and respectively establishing a violation database and a normal database; selecting a certain amount of live broadcast data from the violation database and the normal database, selecting violation and violation keywords, and establishing a knowledge graph of each category according to the violation and violation keywords; according to the method, live broadcast information is analyzed and sorted in a knowledge graph mode, corresponding keyword information is acquired, a target model is established, and illegal information auditing is performed, so that illegal information in daily live broadcast can be audited in real time without too many administrators, the success rate of auditing is far greater than that of manual auditing, and the auditing efficiency is greatly improved. Hidden dangers in the platform can be greatly eliminated, and the live broadcast quality is effectively controlled.

Description

Technical Field

[0001] This application belongs to the field of Internet information processing, and in particular relates to a live broadcast intelligence analysis method based on knowledge graph. Background Art

[0002] Intelligence is information that has been transmitted, organized, and analyzed. It is the premise and foundation for all decision-making. By scientifically and rationally processing intelligence information, we can gain a deep understanding of the current situation, grasp key factors such as our own situation and environment, and thus make decisions that best serve our interests.

[0003] With the continuous development of Internet technology and the rapid growth of live broadcast platforms, more and more people are watching live broadcasts or entering live broadcast-related industries. However, there are currently many illegal or unlawful information in the live broadcast industry, which requires real-time screening by professionals, such as setting up administrators, which increases the burden on practitioners.

[0004] Therefore, how to obtain key intelligence from a large amount of live broadcast data so as to promptly deal with some illegal or unlawful information is a problem that needs to be solved. Summary of the Invention

[0005] The purpose of this application is to provide a live broadcast intelligence analysis method based on knowledge graph to solve the problem of difficulty in efficiently and accurately obtaining violation or illegal information from a large amount of live broadcast data.

[0006] The technical solution of this application is: a live broadcast intelligence analysis method based on knowledge graph, including: Collect a certain amount of live broadcast data of various types from designated platforms and establish violation databases and normal databases respectively; A certain amount of live broadcast data is selected from the violation database and the normal database to select violation and illegal keywords. Based on the violation and illegal keywords, knowledge graphs of different categories are established; different dimensions are set for different categories of knowledge graphs; Establish target models of different categories based on different categories of knowledge graphs; According to different live broadcast types, different types of knowledge graphs are selected respectively. According to the correspondence between each dimension in the knowledge graph and the target model, the corresponding target model is selected to identify illegal information; Report the illegal information, and then verify the illegal information. After the verification is passed, the corresponding illegal information will be processed at different levels.

[0007] Preferably, the live broadcast data is segmented at set time intervals to establish different types of live broadcast databases; live broadcast data that has been determined to be illegal in the live broadcast database is screened out to establish an illegal database; and a certain amount of normal live broadcast data is selected to establish a normal database.

[0008] Preferably, a live broadcast data collection module is set to collect live broadcast data. Different methods are set for collection of different types of live broadcast data, including RTMP stream parsing, WebSocket barrage capture, OCR image recognition and user behavior log collection.

[0009] Preferably, a certain number of keywords with a high frequency of occurrence are selected from the live broadcast data of violations and illegalities as violation and illegality keywords, a certain number of live broadcast data are selected from the violation database and the normal database as input respectively, a recognition algorithm is selected for keyword recognition training, the accuracy of each current violation and illegal keyword in identifying the violation and illegal live broadcast data is counted, and the screened violation and illegal keywords are sorted, and a certain number of violation and illegal keywords with the highest accuracy are selected to establish a knowledge graph of the corresponding category; and the other violation and illegal keywords are trained separately to obtain the knowledge graph of each category.

[0010] Preferably, the number of illegal keywords in different dimensions is different, and the illegal keywords are divided into three levels, namely illegal keywords, illegal keywords and warning keywords, and are displayed in different colors respectively.

[0011] Preferably, the target model includes a speech recognition model, a video image recognition model, a gift recognition model, a scene recognition model and a barrage recognition model; all dimensions of different knowledge graphs are counted respectively, and then the target model is established according to the algorithm selected during the training of the corresponding dimension; then the established target model is associated with the corresponding dimension of the knowledge graph until the correspondence between all dimensions of the knowledge graph and each target model is obtained.

[0012] Preferably, based on the violation information identified by the target model, the violation database is updated at regular intervals, and keywords with higher frequency of occurrence are re-screened. When the keywords with higher frequency of occurrence in a certain dimension change, the changed violation keywords are input into the corresponding target model.

[0013] The knowledge graph-based live broadcast intelligence analysis method of this application uses the knowledge graph to analyze and organize live broadcast intelligence, obtains corresponding keyword information to establish a target model to review illegal information, so that no too many administrators are needed to conduct real-time review of illegal information in daily live broadcasts, and the success rate of the review is much higher than that of manual review, which can greatly eliminate internal hidden dangers of the platform and effectively control the quality of live broadcasts. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions provided by this application, the following is a brief introduction to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of this application.

[0015] Figure 1 This is a schematic diagram of the overall process of this application. DETAILED DESCRIPTION

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0017] A live broadcast intelligence analysis method based on knowledge graph, such as Figure 1 As shown, the following steps are included: Step S100: Collect a certain amount of live broadcast data of various types from a specified platform and establish a violation database and a normal database respectively.

[0018] Set time intervals to segment live broadcast data and establish different types of live broadcast databases. Filter out the live broadcast data identified as illegal to establish a violation database. Select a certain amount of normal live broadcast data to establish a normal database. The time interval is preferably set between 1 and 2 hours. The violation database and normal database can also be directly input using existing data.

[0019] Preferably, a live broadcast data collection module is set to collect live broadcast data. Different methods are set for collection of different types of live broadcast data, including RTMP stream parsing, WebSocket barrage capture, OCR image recognition and user behavior log collection.

[0020] Step S200: Select a certain amount of live broadcast data from the violation database and the normal database, select violation and illegal keywords, and establish knowledge graphs of various categories based on the violation and illegal keywords.

[0021] The types of illegal keywords are determined based on the existing live broadcast regulatory rules. For example, goods or services that are prohibited from production and sale by laws and regulations are not allowed to be sold through online live broadcasts.

[0022] A certain number of frequently occurring keywords are selected from the live broadcast data of illegal activities as illegal activities keywords. A certain amount of live broadcast data is selected from both the illegal and normal databases as input. A recognition algorithm is selected for keyword recognition training. The accuracy of each illegal activity keyword in identifying illegal live broadcast data is calculated. The selected illegal activities keywords are ranked and a certain number of illegal activities keywords with the highest accuracy are selected to create a knowledge graph for the corresponding category. The remaining illegal activities keywords are trained separately to obtain the knowledge graph for each category.

[0023] Different algorithms are used for different types of illegal keyword recognition training, including TF-IDF, TextRank, RNN (recurrent neural network), BERT, etc. For example, RNN is used for training illegal keywords in barrage comments, while TF-IDF is used for training illegal keywords in videos.

[0024] Preferably, different categories of knowledge graphs are set with different dimensions. For example, e-commerce live broadcast is divided into four dimensions: anchor-product-speech-user; game live broadcast includes four dimensions: anchor-gift-lottery-bullet screen; pan-entertainment live broadcast includes five dimensions: anchor-follower-scene-video-bullet screen. Different illegal keywords are selected for each dimension.

[0025] Preferably, the number of illegal keywords in different dimensions is different, and the illegal keywords are divided into three levels, namely illegal keywords, illegal keywords and warning keywords, and are displayed in different colors respectively.

[0026] Step S300: Establish target models of different categories based on knowledge graphs of different categories.

[0027] Target models are used to identify corresponding illegal keywords. Target models include speech recognition models, video image recognition models, gift recognition models, scene recognition models, and bullet comment recognition models. At least one target model is set for each type of target model.

[0028] Specifically, all dimensions of different knowledge graphs are counted separately, and then target models are established according to the algorithms selected during training of the corresponding dimensions; then the established target models are associated with the corresponding dimensions of the knowledge graph until the corresponding relationship between all dimensions of the knowledge graph and each target model is obtained.

[0029] In step S400, different types of knowledge graphs are selected according to different live broadcast types, and corresponding target models are selected according to the correspondence between each dimension in the knowledge graph and the target model to identify illegal information.

[0030] First, different types of live broadcast data are collected through the live broadcast data collection module, and then different types of live broadcast data are sent to the corresponding target model respectively. Based on the target model, illegal information is identified to obtain illegal information.

[0031] Preferably, based on the violations and illegal information identified by the target model, the violation database is updated at regular intervals and frequently occurring keywords are rescreened. When the frequently occurring keywords in a certain dimension change, the changed violations and illegal keywords are input into the corresponding target model. This allows for a consistently high level of accuracy in identifying violations and illegal information even when the means of violation change.

[0032] In step S500, the illegal information is reported and then verified. After the verification is passed, the corresponding illegal information is processed at different levels, such as warning or ban.

[0033] Through the above design, the knowledge graph is used to analyze and organize live broadcast intelligence, and the corresponding keyword information is obtained to establish a target model to review illegal information. This makes it possible to conduct real-time review of illegal information in daily live broadcasts without too many administrators, and the success rate of the review is much higher than that of manual review, which can greatly eliminate internal hidden dangers of the platform and effectively control the quality of live broadcasts.

[0034] Finally, it should be noted that the drawings of the embodiments disclosed in the present invention only involve structures related to the embodiments disclosed in the present invention. Other structures can refer to common designs. In the absence of conflicts, the same embodiment and different embodiments of the present invention can be combined with each other. Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A live broadcast intelligence analysis method based on knowledge graph, characterized in that: include: Collect a certain amount of live broadcast data of various types from designated platforms and establish violation databases and normal databases respectively; A certain amount of live broadcast data is selected from the violation database and the normal database to select violation and illegal keywords. Based on the violation and illegal keywords, knowledge graphs of different categories are established; different dimensions are set for different categories of knowledge graphs; Establish target models of different categories based on different categories of knowledge graphs; According to different types of live broadcasts, different types of knowledge graphs are selected. According to the correspondence between each dimension in the knowledge graph and the target model, the corresponding target model is selected to identify illegal information. Report the illegal information, and then verify the illegal information. After the verification is passed, the corresponding illegal information will be processed at different levels.

2. The live broadcast intelligence analysis method based on knowledge graph according to claim 1, characterized in that: Set time intervals to segment live broadcast data and establish different types of live broadcast databases; filter out live broadcast data that has been identified as illegal in the live broadcast database and establish a violation database; Select a certain amount of normal live broadcast data and establish a normal database.

3. The live broadcast intelligence analysis method based on knowledge graph according to claim 2, characterized in that: Set up the live broadcast data collection module to collect live broadcast data. For different types of live broadcast data, set different methods for collection, including RTMP stream parsing, WebSocket barrage capture, OCR image recognition and user behavior log collection.

4. The live broadcast intelligence analysis method based on knowledge graph according to claim 1, characterized in that: A certain number of keywords with high frequency of occurrence are selected from the live broadcast data of violations and illegal activities as violation and illegal activities keywords. A certain number of live broadcast data are selected from the violation database and the normal database as input respectively. A recognition algorithm is selected for keyword recognition training. The accuracy of each violation and illegal activity keyword in identifying the violation and illegal activity data is calculated. The selected violation and illegal activities keywords are sorted, and a certain number of violation and illegal activities keywords with the highest accuracy are selected to establish a knowledge graph for the corresponding category. Train other illegal keywords separately to obtain knowledge graphs of each category.

5. The live broadcast intelligence analysis method based on knowledge graph according to claim 1, characterized in that: The number of illegal keywords in different dimensions is different. Illegal keywords are divided into three levels, namely illegal keywords, illegal keywords and warning keywords, and are displayed in different colors.

6. The live broadcast intelligence analysis method based on knowledge graph according to claim 5, characterized in that: The target models include speech recognition model, video image recognition model, gift recognition model, scene recognition model and barrage recognition model; all dimensions of different knowledge graphs are counted respectively, and then the target models are established according to the algorithms selected during the training of the corresponding dimensions; then the established target models are associated with the corresponding dimensions of the knowledge graph until the correspondence between all dimensions of the knowledge graph and each target model is obtained.

7. The live broadcast intelligence analysis method based on knowledge graph according to claim 1, characterized in that: Based on the violation information identified by the target model, the violation database is updated at regular intervals, and keywords with higher frequency of occurrence are re-screened. When the keywords with higher frequency of occurrence in a certain dimension change, the changed violation keywords are input into the corresponding target model.

Images