Rumor identification method, device, equipment and storage medium

Through multi-dimensional analysis of the title and text content of rumor information, combined with the preset rumor characteristics, the problem of online rumor identification is solved, the recognition accuracy is improved, and the social information ecology is protected.

CN112069836BActive Publication Date: 2024-12-03TENCENT TECHNOLOGY (SHENZHEN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202010968143.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-09-15
Publication Date
2024-12-03
Estimated Expiration
2040-12-03

AI Technical Summary

Technical Problem

It is difficult for the existing technology to effectively identify and identify rumors spreading on the Internet, resulting in misinformation of information and chaos in social order.

Method used

By extracting features from the title and text content of the information to be identified, combining the preset rumor title type and text content analysis, a multi-dimensional analysis method is used to determine whether the information is a rumor, including title, text content, characteristics of the video playback screen, and audio and video image content and other dimensions.

Benefits of technology

It improves the accuracy and reliability of rumor identification, reduces misjudgment, enhances the ability to identify rumor information, and protects the social information ecology.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112069836B_ABST
    Figure CN112069836B_ABST
Patent Text Reader

Abstract

The embodiment of the present application provides a rumor identification method and device, which at least involves natural language processing technology in artificial intelligence. In the rumor identification method provided by the embodiment of the present application, for the first information to be identified, the first information is analyzed from at least two dimensions, title and text content, to obtain the results corresponding to the two dimensions, which characterize whether the first information is a rumor. That is, judging whether the first information is a rumor from multiple dimensions, the final result obtained based on the results corresponding to the two dimensions, which characterize whether the first information is a rumor, is more accurate.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to rumor identification methods, devices, equipment and storage media. Background Art

[0002] With the development of the Internet, more and more users engage in social activities by sharing information such as videos (especially short videos), audio or text. The information shared by users may be rumors. Rumors shared by users spread quickly and widely among the crowd through the Internet, which may cause panic and affect people's normal life order.

[0003] Therefore, how to identify whether information on the Internet is a rumor is a technical problem that technicians in this field urgently need to solve. Summary of the invention

[0004] In view of this, the present application provides a rumor identification method, device, equipment and storage medium to solve the technical problem of how to identify whether information in the network is a rumor.

[0005] To achieve the above objectives, this application provides the following technical solutions:

[0006] According to a first aspect of an embodiment of the present application, a rumor identification method is provided, including: obtaining a title and text content corresponding to first information to be identified, the text content including content representing the first information obtained directly or indirectly; obtaining a first result representing whether the first information is a rumor based on a relationship between the title and at least one preset rumor title type; determining a second result representing whether the first information is a rumor based on the text content; and obtaining a final result based on the first result and the second result, the final result including information representing whether the first information is a rumor.

[0007] According to the second aspect of an embodiment of the present application, a rumor identification device is provided, including: a first acquisition module, used to acquire a title and text content corresponding to a first information to be identified, the text content including content representing the first information acquired directly or indirectly; a second acquisition module, used to obtain a first result representing whether the first information is a rumor based on a relationship between the title and at least one preset rumor title type; a determination module, used to determine a second result representing whether the first information is a rumor based on the text content; a third acquisition module, used to obtain a final result based on the first result and the second result, the final result including information representing whether the first information is a rumor.

[0008] According to a third aspect of an embodiment of the present application, an electronic device is provided, comprising: a memory for storing a program; a processor for executing the program, wherein the program is specifically used to: obtain a title and text content corresponding to a first information to be identified, wherein the text content includes content representing the first information obtained directly or indirectly; based on a relationship between the title and at least one preset rumor title type, obtain a first result representing whether the first information is a rumor; based on the text content, determine a second result representing whether the first information is a rumor; based on the first result and the second result, obtain a final result, wherein the final result includes information representing whether the first information is a rumor.

[0009] According to a fourth aspect of an embodiment of the present application, a storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps included in the rumor identification method described in the first aspect above are implemented.

[0010] According to a fifth aspect of the embodiments of the present application, a computer program product or a computer program is provided, the computer program product or the computer program including computer instructions, the computer instructions being stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the method provided in various possible implementations of the first aspect above.

[0011] It can be seen from the above technical solutions that in the rumor identification method provided in the embodiment of the present application, for the first information to be identified, the first information is analyzed from at least two dimensions, namely, the title and the text content, to obtain the results corresponding to the two dimensions, which characterize whether the first information is a rumor. That is, judging whether the first information is a rumor from multiple dimensions, the final result obtained based on the results corresponding to the two dimensions, which characterize whether the first information is a rumor, is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0013] Figure 1 An architectural diagram of an implementation method of the implementation environment provided in the embodiments of the present application;

[0014] Figure 2 An architectural diagram of another implementation method of the implementation environment provided in the embodiment of the present application;

[0015] Figure 3 A flowchart of a rumor identification method provided in an embodiment of the present application;

[0016] Figure 4 A schematic diagram of the technical system architecture of the rumor identification method provided in the embodiment of the present application;

[0017] Figure 5 A schematic diagram of another technical system architecture of the rumor identification method provided in an embodiment of the present application;

[0018] Figure 6 A schematic diagram of multiple frames of video images in a video provided by an embodiment of the present application;

[0019] Figure 7 A schematic diagram of a rumor identification device provided in an embodiment of the present application;

[0020] Figure 8 A structural diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0022] The embodiments of the present application disclose a rumor identification method and device. Before introducing the technical solution provided by the embodiments of the present application, the application environment and implementation environment involved in the embodiments of the present application are first introduced.

[0023] First, the application environment involved in the embodiments of the present application is introduced.

[0024] Rumors generally refer to information that is fabricated without any facts. The characteristic of rumors is that they have no corresponding factual basis, but are fabricated and spread between people through certain means. Rumors are an important social phenomenon that concerns fields such as social psychology. In recent years, with the continuous development of science and technology, the form of media has also developed from traditional paper media to electronic media. Rumors spread from person to person through the Internet, making it difficult for people to identify credible information from complex information, which in turn affects people's normal life order, and may even cause panic or lead to unhealthy and unscientific lifestyles, seriously endangering the development of the content ecology. Especially in the face of public emergencies, such as natural disasters, public health events such as COVID-19, and economic crises, widely spread rumors can be extremely destructive.

[0025] Based on this, the present application provides a rumor identification method, which analyzes the first information to be identified from multiple dimensions (for example, title and text content), obtains results representing whether the first information is a rumor from multiple dimensions, and finally determines the final result based on the results corresponding to the multiple dimensions. Compared with identifying the first information from only one dimension, it is more accurate.

[0026] Secondly, the implementation environment involved in the embodiments of the present application is explained. There are multiple implementation environments involved in the embodiments of the present application. The embodiments of the present application provide but are not limited to the following.

[0027] The first implementation environment: Figure 1 As shown, it is an architectural diagram of an implementation method of an implementation environment provided in an embodiment of the present application, and the implementation environment includes: a first server 10, a second server 11 and at least one first electronic device 12.

[0028] Exemplarily, the first electronic device 12 may establish a connection and communicate with the first server 10 via a wireless network. The first server 10 may establish a connection and communicate with the second server 11 via a wireless network or a wired network.

[0029] Exemplarily, the first server 10 and the second server 11 may be the same server, or different servers.

[0030] For example, Figure 1 For example only, Figure 1 Three first electronic devices 12 are shown in the figure. The embodiment of the present application does not limit the number of the first electronic devices 12. The number of the first electronic devices 12 can be determined based on actual conditions and is not limited here.

[0031] Exemplarily, the first electronic device 12 can be any electronic product that can perform human-computer interaction with a user through one or more methods such as a keyboard, touchpad, touch screen, voice interaction or handwriting device, such as a mobile phone, tablet computer, PDA, wearable device, etc.

[0032] Exemplarily, the first server 10 may be a single server, or a server cluster consisting of multiple servers, or a cloud computing service center.

[0033] Exemplarily, the second server 11 may be a single server, or a server cluster consisting of multiple servers, or a cloud computing service center.

[0034] When the holder of the first electronic device 12 publishes the first information through the first electronic device 12, the first information is sent to the first server 10, and the first server 10 can send the first information to other first electronic devices 12. Exemplarily, the first server can be an information distribution platform.

[0035] For example, if the first information is a short video, the first server 10 may be a video distribution platform; if the first information is text, the first server 10 may be a text distribution platform.

[0036] Exemplarily, the first information may be at least one of video, audio, and text.

[0037] The rumor identification method provided in the embodiment of the present application can be applied to the second server 11. After obtaining the first information from the first server 10, the second server 11 can identify whether the first information is a rumor.

[0038] Exemplarily, after the second server 11 determines that the first information is a rumor, it sends a request to the first server 10 to prohibit forwarding the first information, so that the first server 10 will not send the first information to other first electronic devices 12, so that other first electronic devices 12 cannot view the first information.

[0039] Exemplarily, the first server 11 and the second server 11 may be the same server or different servers.

[0040] For example, the final result that the second server 11 determines that the first information is a rumor is not 100% accurate, so further manual review may be required. Therefore, the following second implementation environment is disclosed.

[0041] The second implementation environment, such as Figure 2 As shown, it is an architecture diagram of another implementation method of the implementation environment provided in the embodiment of the present application, and the implementation environment includes: a first server 10, a second server 11, at least one first electronic device 12 and at least one second electronic device 13.

[0042] For the description of the first server 10, the second server 11 and the first electronic device 12, please refer to Figure 1 , I will not go into details here.

[0043] After determining that the first information is a rumor, the second server 11 may send the first information to the second electronic device 13 , so that an expert manually reviews the first information based on the second electronic device 13 .

[0044] Exemplarily, the second electronic device 13 can be any electronic product that can interact with a user through one or more methods such as a keyboard, touchpad, touch screen, voice interaction or handwriting device, such as a mobile phone, tablet computer, PDA, wearable device, etc.

[0045] Exemplarily, the second electronic device 13 may establish a connection and communicate with the second server 11 via a wireless network or a wired network.

[0046] Exemplarily, the second electronic device 13 and the second server 11 may be the same device, or different devices.

[0047] Exemplarily, the final result of whether the first information is a rumor determined by the second server 11 may be different from the real result of whether the first information is a rumor in reality. For example, the final result determined by the second server 11 indicates that the first information is a rumor, but the expert verifies that the first information is not a rumor, that is, the real result indicates that the first information is not a rumor; or, the final result determined by the second server 11 indicates that the first information is not a rumor, but after the first information is widely disseminated, a large number of users report that the first information is a rumor, that is, the real result indicates that the first information is a rumor; or, the final result determined by the second server 11 indicates that the first information is not a rumor, but after the first information is widely disseminated, a large number of users report that the first information is a rumor, and further experts determine that the first information is a rumor after identification, that is, the real result indicates that the first information is a rumor.

[0048] In summary, the final result of whether the first information is a rumor determined by the second server 11 may be different from the actual result of whether the first information is a rumor in practice. The rumor identification method provided in the embodiment of the present application can be corrected based on this first information.

[0049] Exemplarily, the first information (i.e., the first information that the actual result is different from the final result) may be received by the second server or the first server or the second electronic device, and the rumor identification method contained in the second server may be modified.

[0050] Exemplarily, the rumor identification method included in the second server may be modified by other electronic devices other than the second server, the first server and the second electronic device.

[0051] It should be noted that artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that machines have the functions of perception, reasoning and decision-making.

[0052] AI technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. Its basic technologies generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics, etc. AI software technology mainly includes computer vision technology, speech processing technology, natural language processing (Nature Language Processing, NLP) technology, and machine learning / deep learning.

[0053] Among them, NLP technology mainly studies various theories and methods that can achieve effective communication between people and computers using natural language. It is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field will involve natural language, that is, the language used by people in daily life, so it is closely related to the study of linguistics. NLP technology usually includes machine translation. As the name suggests, machine translation technology refers to the study of an intelligent machine that can translate languages ​​in a similar way to human intelligence. Among them, the machine translation system is usually composed of an encoder and a decoder. In addition to machine translation, NLP technology also includes technologies such as robot question answering, text processing, semantic understanding, and knowledge graphs.

[0054] Computer Vision (CV) is a science that studies how to make machines "see". To put it more specifically, it refers to the use of cameras and computers to replace human eyes to identify targets, trace and measure targets, and further perform graphics processing to make computer processing into images that are more suitable for human observation or transmission to instruments for detection. As a scientific discipline, computer vision studies related theories and technologies, and attempts to establish an artificial intelligence system that can obtain information from images or multi-dimensional data. Computer vision technology usually includes image processing, image recognition, image semantic understanding, image retrieval, OCR (Optical Character Recognition), video processing, video semantic understanding, video content / behavior recognition, three-dimensional object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous positioning and mapping, and other technologies, as well as common biometric recognition technologies such as face recognition and fingerprint recognition.

[0055] The key technologies of speech processing technology include automatic speech recognition technology (ASR), text to speech technology (TTS) and voiceprint recognition technology. Enabling computers to listen, see, speak and feel is the future development direction of human-computer interaction, among which speech has become one of the most promising human-computer interaction methods in the future.

[0056] Machine Learning (ML) is a multi-disciplinary subject that involves probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning.

[0057] The embodiments of the present application involve multiple technologies in the field of artificial intelligence, which are specifically described through the following embodiments.

[0058] Please see attached Figure 3 , which is a flow chart of a rumor identification method provided in an embodiment of the present application. The method includes the following steps S301 to S304 during implementation.

[0059] Step S301: Obtain the title and text content corresponding to the first information to be identified.

[0060] The text content includes content representing the first information obtained directly or indirectly. Specifically, the content representing the first information obtained directly is text content attached to the first information itself, such as titles and tags; the content representing the first information obtained indirectly is text content converted from the first information, such as text content obtained from audio and video through audio-to-text conversion, or subtitles, bullet comments and other text content obtained from videos through image recognition and other technologies.

[0061] The following is an explanation of the title and text content with examples.

[0062] The first type of media form of the first information is video.

[0063] As an increasingly popular media communication carrier, videos, especially short videos, have attracted more and more attention. The development of video as a media form has led to a large amount of unreliable information being spread quickly and widely among the crowd in the form of multimedia rumors such as video rumors.

[0064] If the first information is a video, the title is the title of the video; illustratively, the video image contained in the video may include at least one of a person, an object, an animal, a scene, and a subtitle. Optionally, the text content refers to at least one of text information representing a person image contained in the video image, text information representing a person behavior contained in the video image, text information representing a scene contained in the video image, and subtitle information contained in the video image.

[0065] The above-mentioned acquisition of the text content corresponding to the video involves important technical branches of computer vision technology in the field of AI - image semantic understanding (ISU) and video semantic understanding (VSU).

[0066] Exemplarily, the text content corresponding to the video can be obtained through ISU technology. Exemplarily, the text information contained in the video image can be obtained based on the OCR technology in the ISU technology.

[0067] Exemplarily, the information representing the behavior of the person contained in the video image may be obtained based on the VSU technology.

[0068] Exemplarily, if the video includes audio, optionally, the audio can be converted into text through voice processing technology in AI technology, and the above text content includes the text converted from the audio.

[0069] For example, audio is converted into text through ASR, a speech processing technology in AI technology.

[0070] It is understandable that if the audio includes at least one of background music and dubbing for the video, if the dubbing is in dialect, or if the audio includes background music, then the accuracy of the text converted from the audio may not be high. In view of this, the text (such as subtitle information) obtained from the content of the video image contained in the video based on computer vision technology can make up for the above defects.

[0071] It is understandable that the media form of video is relatively complex. If the media form of the first information is video, the first information can be analyzed from a multimodal perspective. For example, the multimodal perspective includes but is not limited to at least one of the title, video image content, and audio.

[0072] The second media form of the first information is audio.

[0073] The audio may be converted into text by using speech processing technology in AI technology, and the above text content includes text converted from the audio. For example, the audio may be converted into text by using ASR in speech processing technology in AI technology.

[0074] The third media form of the first information is text.

[0075] The text content is the text. If the text has a title, the title of the text is the title mentioned in step S301; if the text does not have a title, then a keyword that can represent the text can be extracted from the text and the keyword is determined as the title of the text.

[0076] Exemplarily, keywords that can represent the text can be extracted from the text based on the semantic understanding technology in the NLP technology in the field of artificial intelligence technology.

[0077] For example, keywords that can represent the text are extracted from the text through the one-hot technology of the bag-of-words model or the TF-TDF (term frequency-inverse document frequency) technology in NLP technology.

[0078] It is understandable that no matter what media form the first information belongs to, it may correspond to introduction content. For example, a video corresponds to introduction content, and the introduction content corresponding to the first information is used to briefly describe the content contained in the first information. That is, the introduction content is a brief version of the text content. Exemplarily, if the first information corresponds to introduction content, the text content corresponding to the first information may include the introduction content.

[0079] Exemplarily, since the title of the first information is also a general description of the first information, if the first information corresponds to introduction content, the title corresponding to the first information may include the introduction content.

[0080] Exemplarily, neither the title of the first information nor the text content of the first information includes introduction content.

[0081] Exemplarily, the media form of the first information may be at least one of video, audio and text.

[0082] Step S302: Based on the relationship between the title and at least one preset rumor title type, a first result is obtained to indicate whether the first information is a rumor.

[0083] Exemplarily, the number of preset rumor title types is greater than or equal to 1. The number of preset rumor title types may be determined based on actual conditions, and the embodiment of the present application does not limit this.

[0084] Exemplarily, if the first information corresponds to a title and an introduction content, step S302 may include: obtaining the relationship between the title and at least one preset rumor title type to obtain a third sub-result representing whether the first information belongs to a rumor; obtaining the relationship between the introduction content and at least one preset rumor introduction type to obtain a fourth sub-result representing whether the first information belongs to a rumor, wherein the first result includes the third sub-result and the fourth sub-result.

[0085] Exemplarily, if the first information does not correspond to a title but corresponds to introduction content, then the introduction content can be used as the title.

[0086] Exemplarily, if the first information corresponds to a title and an introduction content, the title and the introduction content can be taken as a whole, for example, the title and the introduction content are called title introduction information, and step S302 may include: obtaining the relationship between the title introduction information and at least one preset rumor title type, and obtaining a first result characterizing whether the first information is a rumor.

[0087] Exemplarily, the media form of the first information is different, and the preset at least one rumor title type may be different. Exemplarily, the embodiment of the present application can classify the rumor title types in each media form based on a large number of rumors in each media form.

[0088] The following description will be made by taking video as an example of the media form.

[0089] Rumor videos involve all aspects of human social life and cover a wide variety of topics. Compared with non-rumor videos, rumor videos have a more obvious tendency. According to the titles of rumor videos, they can be roughly divided into the following types of rumor titles.

[0090] Fraud type, which is the type of rumor video titles that take advantage of people's sympathy or pursuit of profit to deceive people into paying a lot of attention or forwarding; common sense type, which is the type of rumor video titles that distort common sense about nature, history, and life to attract people's attention; social entertainment and life type, which is the type of video rumor titles about gossip about people from all walks of life or false news in the entertainment industry.

[0091] The above four types of rumor title types mentioned in the embodiments of the present application are only examples, and the "at least one preset rumor title type" is not limited to the above four types. For example, "at least one preset rumor title type" may also include: clickbait type, event type to be verified, false type formed by combining history and reality, and superstitious type.

[0092] Among them, the headline party type refers to all kinds of headlines that are difficult to distinguish between true and false, alarmist, and even deceptive and intimidating. At present, in order to attract the attention of netizens, the bottom line of headline production is constantly being refreshed, and all kinds of headlines that are difficult to distinguish between true and false, alarmist, and even deceptive and intimidating are emerging in an endless stream. For example, the headlines "The fake Spring Festival Gala raised millions of yuan and may invite Guo Degang to perform crosstalk" and "Liu Xiang may participate in the East Asian Games" are headline party types.

[0093] The event type that needs to be verified refers to the type of title that the truth of the event still needs to be investigated, studied, and discussed. For example, the title "The Real Snake Transformed into a Dragon" belongs to the event type that needs to be verified.

[0094] The false type formed by combining history and reality refers to the type of title that combines historical information with current information. For example, the title "Hawking once predicted that Qin Shihuang just couldn't get out of his own tomb!" and the title "A thousand-year-old brocade was unearthed from an ancient tomb, with eight words written on it, predicting what will happen in China in 2040" belong to the false type formed by combining history and reality.

[0095] The supernatural type refers to the type of titles that are about weirdness, bravery, rebellion, ghosts and gods. For example, titles such as "The daily life of hell messengers and ghosts", "Supernatural events in XX area", and "Are there really evil spirits?" belong to the supernatural type.

[0096] Exemplarily, the media forms of the first information are different, but at least one preset rumor title type may be the same.

[0097] Exemplarily, if the title of the first information belongs to any rumor title type of at least one preset rumor title type, then the first result is that the first information belongs to a rumor, or the first result is that the probability that the first information belongs to a rumor is greater than or equal to threshold A.

[0098] Exemplarily, if the title of the first information does not belong to any preset rumor title type, then the first result is that the first information is not a rumor, or the first result is that the probability that the first information is a rumor is less than threshold A.

[0099] Step S303: Determine a second result based on the text content, indicating whether the first information is a rumor.

[0100] For example, in the execution Figure 3 In the process of each step shown, there is no execution order between step S302 and step S303. For example, step S302 may be executed first and then step S303, or step S303 may be executed first and then step S302, or step S302 and step S303 may be executed simultaneously.

[0101] It can be known from step S302 and step S303 that the first information is analyzed from at least two dimensions, namely, the title and the text content, to obtain the results corresponding to the two dimensions, which respectively characterize whether the first information is a rumor.

[0102] Exemplarily, if the first information corresponds to an introduction content, the text content corresponding to the first information may include the introduction content. Assuming that the content representing the first information contained in the text content is the first text content, step S303 may include: determining a fifth sub-result representing whether the first information is a rumor based on the introduction content; determining a sixth sub-result representing whether the first information is a rumor based on the first text content, wherein the second result includes the fifth sub-result and the sixth sub-result.

[0103] Exemplarily, if the first information corresponds to an introduction content, the text content corresponding to the first information may include the introduction content. The first text content and the introduction content may be taken as a whole, and step S303 may include: determining a second result indicating whether the first information is a rumor based on the introduction content and the first text content.

[0104] Exemplarily, if the first information does not correspond to any introduction content, the text content only includes the first text content.

[0105] Step S304: Based on the first result and the second result, a final result is obtained, where the final result includes information indicating whether the first information is a rumor.

[0106] Exemplarily, if both the first result and the second result represent that the first information is not a rumor, then the final result represents that the first information is not a rumor; exemplarily, if at least one of the first result and the second result represents that the first information is a rumor, the final result is a result representing that the first information is a rumor.

[0107] Exemplarily, if the first result is that the probability that the first information is a rumor is a first probability, and the second result is that the probability that the first information is a rumor is a second probability, weights can be set for the first result and the second result respectively. For example, the first result corresponds to weight 1, and the second result corresponds to weight 2. Then, information A (assuming that the final result includes information characterizing whether the first information is a rumor is information A) = weight 1*first probability + weight 2*second probability.

[0108] Exemplarily, if information A is greater than or equal to threshold B, information A indicates that the first information is a rumor; if information A is less than threshold B, information A indicates that the first information is not a rumor.

[0109] Exemplarily, the final result may include the first result and the second result, as well as information A; if information A indicates that the first information is a rumor, the final result and the first information are sent to the second electronic device so that an expert can manually review the first information in combination with the final result.

[0110] In the rumor identification method provided in the embodiment of the present application, for the first information to be identified, the first information is analyzed from at least two dimensions, namely, the title and the text content, to obtain the results corresponding to the two dimensions, which characterize whether the first information is a rumor. That is, judging whether the first information is a rumor from multiple dimensions, the final result obtained based on the results corresponding to the two dimensions, which characterize whether the first information is a rumor, is more accurate.

[0111] In an optional implementation, if the media form of the first information is video, in addition to analyzing the first information from two dimensions of title and text content, the first information can also be analyzed from three dimensions: characteristics of the video playback screen, audio contained in the video, and content contained in the video image in the video.

[0112] By analyzing a large number of rumor videos, it is found that rumor videos with the characteristics of the playback screen belonging to the preset carousel type account for a large proportion. Among them, the screen playback characteristics of the video belonging to the carousel type are that one screen stays for a certain length of time before switching to another screen, similar to the playback of a slide show. In the embodiment of the present application, the video belonging to the carousel type is called a carousel video.

[0113] It is understandable that a carousel video is not necessarily a rumor video, and a non-carousel video may also be a rumor video. It is just that if a video is a carousel video, it is more likely to be a rumor. Therefore, if a video is a carousel video, it cannot be directly determined whether the video is a rumor video or a non-rumor video. Therefore, the embodiment of the present application refers to "determining whether the video belongs to the preset carousel type based on the picture change characteristics of the video" as an "auxiliary black judgment strategy."

[0114] The applicant uses the feature that "carousel videos account for a large proportion of rumor videos" to use whether a video is a carousel video as auxiliary information to determine whether the first information is a rumor.

[0115] In an optional embodiment, the method for determining whether a video belongs to a carousel video includes: determining whether the video belongs to a carousel video based on a picture change feature of the video.

[0116] The picture change feature at least includes the dwell time corresponding to at least one picture contained in the video, and the dwell time corresponding to one picture refers to the display time of the picture before switching from the picture to another picture.

[0117] Exemplarily, if the dwell time included in the picture change characteristics of the video is greater than or equal to the preset duration, it is determined that the video is a carousel video; if the dwell time included in the picture change characteristics of the video is less than the preset duration, it is determined that the video is not a carousel video.

[0118] Exemplarily, in actual applications, the preset duration corresponding to a picture can be determined in the following way, so that the user can perceive that the picture "pauses" for a while before switching to the next picture. The duration of the picture "pause" is the preset duration, or the preset duration is any value greater than the duration of the picture "pause".

[0119] Exemplarily, the picture transformation features of the video can be obtained based on the ISU technology in the computer vision technology in the field of artificial intelligence technology.

[0120] There are multiple ways to “determine whether the video belongs to a carousel video based on the picture change characteristics of the video”. The embodiments of the present application provide but are not limited to the following two.

[0121] The first implementation method involves machine learning technology in the AI ​​field, which is explained in detail below.

[0122] The first method for determining whether a video is a carousel video includes the following steps A11 to A12.

[0123] Step A11: inputting the multiple frames of video images contained in the video, or the video, into a pre-constructed second classifier, where the second classifier is used to detect whether the video is a carousel video.

[0124] The second classifier may obtain picture change features from multiple frames of video images or videos to determine whether the video is a carousel video based on the picture change features.

[0125] There are many methods for constructing the second classifier. The embodiments of the present application provide but are not limited to the following methods.

[0126] The method for constructing the second classifier includes steps A121 and A122.

[0127] Step A121: Acquire multiple first sample videos.

[0128] The plurality of first sample videos include a plurality of positive first sample videos and negative first sample videos. The positive first sample videos are carousel videos, and the negative first sample videos are non-carousel videos.

[0129] Step A122: Use the multiple first sample videos as input of the first machine model, use the labeled video types corresponding to the multiple first sample videos as output, and train the first machine model to obtain a second classifier.

[0130] Among them, the annotated video type corresponding to the positive first sample video is a carousel video, and the annotated video type corresponding to the negative first sample video is a non-carousel video.

[0131] The process of training the first machine model involves at least one of the techniques in machine learning, such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and method-based learning.

[0132] Exemplarily, the first machine model may be any one of a neural network model, a logistic regression model, a linear regression model, a support vector machine (SVM), an Adaboost, an XGboost, and a Transformer-Encoder model.

[0133] Exemplarily, the neural network model can be any one of a recurrent neural network-based model, a convolutional neural network-based model, and a Transformer-encoder-based classification model.

[0134] Exemplarily, the first machine model may be a deep hybrid model of a recurrent neural network-based model, a convolutional neural network-based model, and a Transformer-encoder-based classification model.

[0135] Exemplarily, the first machine model can be any one of an attention-based deep model, a memory network-based deep model, and a short text classification model based on deep learning.

[0136] The short text classification model based on deep learning is a recurrent neural network (RNN) or a convolutional neural network (CNN) or a variant based on the recurrent neural network or the convolutional neural network.

[0137] Exemplarily, some simple domain adaptation transformation can be performed on the pre-trained model to obtain the first machine model.

[0138] Exemplarily, “simple domain adaptation” includes but is not limited to re-pre-training an already pre-trained model using large-scale unsupervised domain corpus, and / or compressing an already pre-trained model by model distillation.

[0139] Step A121 to step A122 are supervised learning. Exemplarily, semi-supervised learning can also be performed on the first machine model. Semi-supervised learning is a learning method that combines supervised learning with unsupervised learning. Semi-supervised learning uses a large amount of unlabeled data and labeled data to perform pattern recognition.

[0140] Step A12: Obtain a second prediction result output by the second classifier, where the second prediction result is that the video is a carousel video, or that the video is not a carousel video.

[0141] The second implementation method involves image recognition technology, an important technical branch of computer vision technology in the field of AI, which is explained in detail below.

[0142] The second method for determining whether a video belongs to a carousel video includes the following steps A21 to A22.

[0143] Step A21: Acquire multiple frames of video images from the video.

[0144] Step A22: Divide the multiple frames of video images into at least one group of image sets, where one group of image sets includes one or more frames of video images with completely identical image contents.

[0145] Exemplarily, it may be possible to detect whether the image contents contained in multiple frames of video images are the same based on image recognition technology.

[0146] Step A23: obtaining the picture change feature based on the time periods respectively corresponding to at least one group of image sets.

[0147] The time period corresponding to a group of image sets refers to the difference between the latest time and the earliest time of one or more video frames contained in the group of image sets in the video.

[0148] For example, a group of image sets includes video image 1, video image 2 and video image 3, wherein video image 1 is located at 1 minute 50 seconds in the video, video image 2 is located at 1 minute 55 seconds in the video, and video image 3 is located at 2 minutes 0 seconds in the video. Then, the time period corresponding to the group of image sets is 2 minutes - 1 minute 50 seconds = 10 seconds.

[0149] Step A24: Based on the picture change characteristics of the video, determine whether the video is a carousel video.

[0150] Exemplarily, step S304 is: obtaining the final result based on the first result, the second result, and whether the video is a carousel video.

[0151] The first information is analyzed and explained below in conjunction with the audio dimension contained in the video.

[0152] By analyzing a large number of rumor videos, we found that a large proportion of the audio contained in the rumor videos is sound produced by machines (not human sounds).

[0153] In the embodiment of the present application, if the audio in the video is the sound produced by a machine, the video is called a machine-dubbed video.

[0154] It is understandable that if the video is a machine-dubbed video, the video may or may not be a rumor; however, if the video is a rumor, the probability that the video is a machine-dubbed video is higher. That is, if it is determined that the video is a machine-dubbed video, it cannot be determined whether the video is a rumor. Therefore, judging whether the video is a machine-dubbed video is also called "auxiliary black judgment strategy."

[0155] The applicant uses the feature that "a large proportion of the audio in the video is made by machines (not humans)" to determine whether the video is a machine-dubbed video as auxiliary information to determine whether the first information is a rumor.

[0156] There are many ways to determine whether a video is a machine-dubbed video. The embodiment of the present application provides but is not limited to a method: based on the attribute information of the sound, determine whether the video is a machine-dubbed video.

[0157] Exemplarily, the attribute information of the audio includes at least one of pitch, timbre, voiceprint, and waveform.

[0158] For example, since the sound produced by a machine is mechanical and has no emotional fluctuations, this feature can be combined to determine whether the video is a machine-dubbed video based on the attribute information of the audio.

[0159] Exemplarily, the attribute information of the audio may be obtained based on speech recognition technology and voiceprint recognition technology in speech processing technology in the AI ​​field.

[0160] Exemplarily, step S304 is: obtaining the final result based on the first result, the second result, and whether the video is a machine-dubbed video.

[0161] Exemplarily, if the video is a machine-dubbed video, then the probability that the final result represents that the first information is a rumor will increase; if the video is not a machine-dubbed video, then the probability that the final result represents that the first information is a rumor will decrease.

[0162] Exemplarily, step S304 is: obtaining the final result based on the first result, the second result, whether the video is a machine-dubbed video, and whether the video is a carousel video.

[0163] The first information is analyzed and explained below in combination with the dimension of the content contained in the video image in the video.

[0164] It is understandable that the types of non-rumor videos are determined by accumulating a large number of videos that are not rumors, as well as a large number of videos that are mistakenly judged as rumor videos but are actually not rumor videos.

[0165] Exemplarily, non-rumor video types include, but are not limited to: any one or more of: animation video types, game video types, ball video types, fighting video types, vulgar video types, and square dance video types.

[0166] The animation video type refers to the type of video whose content is an existing animation plot, or the type of video whose content is an animation production video. For example, if the video image contains characters from Fox Spirit Matchmaker, and the behaviors of the characters in the video image are also the behaviors of the characters in the plot related to Fox Spirit Matchmaker, then it can be determined that the video belongs to an animation video.

[0167] The game video type refers to the type of video to which the content contained in the video image is a game clip, or the type of video to which the content contained in the video image is a game live broadcast.

[0168] The ball game video type refers to the type of video to which the content contained in the video image is a ball game, for example, the content contained in the video image is a video of playing badminton, or the content contained in the video image is a video of playing table tennis.

[0169] The fighting video type refers to the type of video to which the content contained in the video images includes fighting scenes throughout or the video images containing fighting scenes account for a preset ratio (for example, 40%) of the total number of video images contained in the entire video.

[0170] Vulgar video types refer to video images that contain pornographic and vulgar content.

[0171] The square dance video type refers to the type of video that contains square dance scenes.

[0172] Exemplarily, if the video belongs to one or more types of non-rumor video types, it means that the video is a non-rumor video. If the video does not belong to any of the non-rumor videos, it means that the video may be a rumor video. Therefore, in the embodiment of the present application, "determining the relationship between the video and the preset non-rumor video types" is referred to as "direct judgment strategy."

[0173] For example, a video may belong to one or more non-rumor video types at the same time, for example, a video belongs to both an animation video and a fighting video.

[0174] There are many ways to determine the relationship between a video and the preset non-rumor video types. The embodiments of the present application provide but are not limited to the following two.

[0175] The first implementation method involves machine learning technology in the AI ​​field, which is explained in detail below.

[0176] The first method of determining the relationship between a video and each preset non-rumor video type includes the following steps C11 to C13.

[0177] Step C11: Acquire multiple frames of video images from the video.

[0178] Step C12: inputting the multiple frames of video images or the video into pre-constructed third classifiers corresponding to each non-rumor video type respectively.

[0179] The construction process of the pre-constructed third classifier corresponding to each non-rumor video type is similar, and the construction process of the third classifier corresponding to the animation video type is explained below.

[0180] Exemplarily, if the video input is directly input into the pre-built third classifiers corresponding to each non-rumor video type, step C11 may be omitted.

[0181] The process of constructing the third classifier corresponding to the animation video type includes the following steps C121 to C122.

[0182] Step C121: Obtain multiple second sample videos.

[0183] The plurality of second sample videos include a plurality of positive sample videos and negative sample videos. The positive second sample videos are cartoon videos, and the negative second sample videos are non-cartoon videos, wherein the non-cartoon videos can be non-rumor videos or rumor videos.

[0184] Step C122: taking the plurality of second sample videos as inputs of the second machine model, taking the labeled video types respectively corresponding to the plurality of second sample videos as outputs, and training the machine model to obtain a third classifier.

[0185] Among them, the annotated video type corresponding to the positive second sample video is an animation video, and the annotated video type corresponding to the negative second sample video is a non-animation video.

[0186] The description of the second machine model refers to the description of the first machine model, which will not be repeated here.

[0187] Step C121 to step C122 are supervised learning. Exemplarily, semi-supervised learning can also be performed on the second machine model. Semi-supervised learning is a learning method that combines supervised learning with unsupervised learning. Semi-supervised learning uses a large amount of unlabeled data and labeled data at the same time to perform pattern recognition.

[0188] Step C13: Determine the relationship between the video and each preset non-rumor video type based on the third prediction results respectively output by the third classifiers corresponding to each non-rumor video type.

[0189] A third prediction result output by a third classifier corresponding to a non-rumor video type is that the video belongs to the non-rumor video type, or that the video does not belong to the non-rumor video type.

[0190] The second implementation method involves machine learning technology in the AI ​​field, which is explained in detail below.

[0191] The second method of determining the relationship between a video and each preset non-rumor video type includes the following steps C21 to C22.

[0192] Step C21: input multiple frames of video images in the video, or the video, into a pre-constructed fourth classifier.

[0193] The fourth classifier is used to detect the non-rumor video type to which the video belongs based on multiple frames of video images.

[0194] Exemplarily, the fourth classifier construction process includes the following steps C121 to C122.

[0195] Step C221: Obtain multiple third sample videos.

[0196] The plurality of third sample videos include a plurality of positive sample videos and a plurality of negative sample videos. The plurality of positive third sample videos include a plurality of videos belonging to various non-rumor video types, for example, a plurality of videos belonging to the animation video type, a plurality of videos belonging to the game video type, a plurality of videos belonging to the ball video type, a plurality of videos belonging to the fighting video type, a plurality of videos belonging to the vulgar video type, and a plurality of videos belonging to the square dance video type; the negative second sample video is a video that does not belong to any non-rumor video type.

[0197] Step C222: Use multiple third sample videos or multiple frames of video images in multiple third sample videos as input of the third machine model, use the labeled video types corresponding to the multiple third sample videos as output, and train the machine model to obtain a fourth classifier.

[0198] Among them, the annotated video type corresponding to a positive third sample video is the non-rumor video type to which it belongs, and the annotated video type corresponding to a negative third sample video is that it does not belong to any non-rumor video type.

[0199] The description of the third machine model refers to the description of the first machine model, which will not be repeated here.

[0200] Step C22: Based on the fourth prediction result output by the fourth classifier, determine the relationship between the video and each preset non-rumor video type.

[0201] The fourth prediction result is that the video belongs to a target non-rumor video type in at least one non-rumor video type, or the video does not belong to any non-rumor video type.

[0202] Exemplarily, the fourth classifier includes pre-built third classifiers corresponding to each non-rumor video type.

[0203] It is understandable that the first information may correspond to introduction content. If the title does not include the introduction content and the text content does not include the introduction content, illustratively, the first information can also be analyzed and explained from the dimension of the introduction content.

[0204] In an optional implementation, the rumor identification method provided in the embodiment of the present application also includes: determining the relationship between the introduction content of the video and at least one preset rumor introduction type.

[0205] Exemplarily, the process of "determining the relationship between the introduction content of the video and at least one preset rumor introduction type" is similar to the process of "obtaining the relationship between the title and at least one preset rumor title type", which will be explained in a unified manner later.

[0206] In summary, the first information can be described from at least two of the six dimensions: title, text content, introduction content (if the title does not include introduction content and the text content does not include introduction content), characteristics of the video playback screen (if the media form of the first information is video), audio contained in the video (if the media form of the first information is video), and content contained in the video image in the video (if the media form of the first information is video). The accuracy of the information A contained in the final result is improved.

[0207] Exemplarily, step S304 may include: determining the final result characterizing whether the first information is a rumor based on whether the video belongs to a carousel video, the target prediction results output by the target classifiers corresponding to each of the non-rumor video types, whether the video is a machine-dubbed video, and at least one of the relationships between the introduction content of the video and at least one preset rumor introduction type, as well as the first result and the second result.

[0208] In an embodiment of the present application, when the media form of the first information is video, the first information can be controlled in all directions from six dimensions. The six dimensions constitute the technical system architecture of the rumor identification method, that is, through a multi-level combination strategy of multiple dimensions (for example, direct white judgment strategy, auxiliary black judgment strategy, etc.), it is possible to identify whether the video is a rumor.

[0209] In order for those skilled in the art to have a more systematic understanding of the technical system architecture of the above rumor identification method, Figure 4 To explain, Figure 4 , which is a schematic diagram of the technical system architecture of the rumor identification method provided in an embodiment of the present application.

[0210] like Figure 4 As shown, the rumor identification method can analyze the first information from six dimensions, thereby controlling the first information in all directions and obtaining a more accurate final result.

[0211] like Figure 4 As shown, the six dimensions are: dimension 41 (title), dimension 42 (introduction information), dimension 43 (text content), dimension 44 (audio), dimension 45 (picture change features) and dimension 46 (content contained in the video image in the video).

[0212] For example, Figure 4 This is just an example. In actual application, one or more dimensions can be determined from the six dimensions to analyze the first information based on the content corresponding to the first information. For example, if the media form of the first information is text, you can choose to analyze the first information from the two perspectives of dimension 41 and dimension 43; if the first information is a video, but the first information does not correspond to the introduction content, then the first information can be analyzed from the five perspectives of dimension 41, dimension 43, dimension 44, dimension 45 and dimension 46; if the first information is a video and the first information corresponds to the introduction content, then the first information can be analyzed from the above six dimensions respectively.

[0213] Dimension 44, dimension 45 and dimension 46 have been described in detail above. Dimension 41, dimension 42 and dimension 43 will be described in detail below.

[0214] First, dimension 41 is introduced.

[0215] In an optional embodiment, there are multiple ways to implement step S302, and the embodiments of the present application provide but are not limited to the following three.

[0216] The first implementation method involves semantic understanding technology and text preprocessing technology in NLP technology in the AI ​​field, which is explained in detail below.

[0217] The first implementation of step S302 includes the following steps D11 to D12.

[0218] Step D11: If the title does not conform to the keyword pattern corresponding to the at least one rumor title type, a first result is obtained indicating that the first information is not a rumor.

[0219] Step D12: If the title matches the keyword pattern corresponding to the target rumor title type in the at least one rumor title type, a first result is determined that characterizes the first information as a rumor.

[0220] Exemplarily, the keyword pattern corresponding to a rumor title type may be a keyword set, and the keyword set may include multiple keywords corresponding to multiple titles belonging to the rumor title type.

[0221] Exemplarily, for each rumor title type, the more titles that belong to this rumor title type are collected, the more comprehensive the keyword pattern corresponding to this rumor title type is obtained.

[0222] Exemplarily, for each rumor title type, a keyword pattern corresponding to the rumor title type can be obtained from a large number of titles based on semantic understanding technology and text preprocessing technology.

[0223] Exemplarily, the keyword pattern corresponding to a rumor title type may be rule information composed of at least one keyword and used to describe a title that conforms to the rumor title type.

[0224] For example, a keyword pattern can be expressed using a regular expression.

[0225] The second implementation method involves machine learning technology in the AI ​​field, which is explained in detail below.

[0226] The second implementation of step S302 includes the following steps D21 to D23.

[0227] Step D21: If the title does not match the keyword pattern corresponding to the at least one rumor title type, a first result is obtained that indicates that the first information is not a rumor.

[0228] For example, through statistics of a large number of rumors, relevant keyword patterns can be summarized for each type of rumor title.

[0229] For example, if the title is classified into a certain rumor title type, the first information is not necessarily a true rumor, but if the first information is a true rumor, then it will definitely be classified into a certain rumor title type among all rumor title types.

[0230] In summary, the keyword pattern corresponding to each rumor title type is a necessary but not sufficient condition for the first information.

[0231] For example, in the embodiment of the present application, when setting the keyword mode corresponding to each rumor title type, the recall rate is given priority. The recall rate is described below with an example.

[0232] If there are 1000 first information, among which 100 are rumors and 900 are not rumors, and the titles of the 1000 first information are classified, if 500 first information are classified into the corresponding rumor title type, among these 500 first information, 90 are rumors and 410 are not rumors, then the recall rate = 90 / 100 = 0.9.

[0233] Exemplarily, when setting the keyword patterns corresponding to each rumor title type, the embodiment of the present application tries to make the recall rate close to 100%, or even equal to 100%, to ensure that no rumors are missed.

[0234] For example, the accuracy is improved as much as possible while making the recall rate close to 100%.

[0235] Still using the above example to illustrate the accuracy, accuracy = 90 / 500 = 0.18.

[0236] For example, the recall rate and the precision rate may be combined to determine the keyword patterns corresponding to each rumor title type.

[0237] Step D22: If the title matches the keyword pattern corresponding to the target rumor title type in the at least one rumor title type, the title is input into the first classifier corresponding to the target rumor title type.

[0238] The first classifier is used to detect whether the input title belongs to the target rumor title type.

[0239] Exemplarily, if a title is classified into a target rumor title type, the first information is not necessarily a true rumor, and therefore, it is necessary to further determine whether the title classified into the target rumor title type actually belongs to the target rumor title type.

[0240] Exemplarily, different rumor title types correspond to different classifiers, and the classifier corresponding to a rumor title type is used to detect whether the input title belongs to the rumor title type.

[0241] Exemplarily, since the feature emphases of titles of different rumor title types are not the same, a first classifier is designed for each rumor title type.

[0242] The first classifier corresponding to different rumor title types has different focuses on obtaining the features of the corresponding titles.

[0243] Exemplarily, the construction process of the first classifier corresponding to each rumor title type is similar. The construction process of the first classifier corresponding to rumor title type A is taken as an example for explanation below. The process of constructing the first classifier includes the following steps D221 to D222.

[0244] Step D221: Get multiple sample titles.

[0245] The plurality of sample titles include a plurality of positive sample titles and negative sample titles. The positive sample titles are titles belonging to rumor title type A, and the negative sample titles are titles not belonging to rumor title type A.

[0246] Step D222: Use the multiple sample titles as input of the fourth machine model, use the annotation types corresponding to the multiple sample titles as output, and train the fourth machine model to obtain the first classifier.

[0247] Among them, the annotation type corresponding to the positive sample title is rumor title type A, and the annotation type corresponding to the negative sample title is non-rumor title type A.

[0248] The description of the fourth machine model refers to the description of the first machine model, which will not be repeated here.

[0249] Step D221 to step D222 are supervised learning. Exemplarily, semi-supervised learning can also be performed on the fourth machine model. Semi-supervised learning is a learning method that combines supervised learning with unsupervised learning. Semi-supervised learning uses a large amount of unlabeled data and labeled data at the same time to perform pattern recognition.

[0250] In summary, the first classifier is used to further determine whether the title really belongs to the target rumor title type, which makes up for the low accuracy while ensuring the recall rate.

[0251] Step D23: Based on the first prediction result output by the first classifier, the first result is obtained, the first prediction result is that the title belongs to the target rumor title type, or the first prediction result is that the title does not belong to the target rumor title type.

[0252] Exemplarily, if the first prediction result is that the title belongs to the target rumor title type, then the first result represents that the first information belongs to a rumor; if the first prediction result is that the title does not belong to the target rumor title type, then the first result represents that the first information does not belong to a rumor.

[0253] The third implementation method involves machine learning technology in the AI ​​field, which is explained in detail below.

[0254] The third implementation method of step S302 includes the following steps D31 to D32.

[0255] Step D31: Input the title into the fifth classifier.

[0256] The fifth classifier is used to detect which rumor title type the input title belongs to among multiple rumor title types.

[0257] Exemplarily, the fifth classifier is equivalent to the first classifier including each rumor title type respectively corresponding to the first classifier.

[0258] Step D32: Obtain the first result based on the fifth prediction result output by the fifth classifier.

[0259] The fifth prediction result is that the title belongs to a target rumor title type among multiple rumor title types, or the title does not belong to any rumor title type among multiple rumor title types.

[0260] Exemplarily, the process of training the fifth classifier includes steps D321 and D322.

[0261] Step D321: Obtain multiple sample titles.

[0262] The plurality of sample titles include a plurality of positive sample titles and negative sample titles. Titles belonging to each rumor title type are called positive sample titles, and titles not belonging to any rumor title type in each rumor title type are called negative sample titles.

[0263] Step D322: Use multiple sample titles as input of the fifth machine model, use the annotation types corresponding to the multiple sample titles as output, and train the fifth machine model to obtain a fifth classifier.

[0264] Among them, the annotation type corresponding to the positive sample title is the rumor title type to which the positive sample title belongs, and the annotation type corresponding to the negative sample title is that it does not belong to any rumor title type.

[0265] The description of the fifth machine model refers to the description of the first machine model, which will not be repeated here.

[0266] Step D321 to step D322 are supervised learning. Exemplarily, the fifth machine model can also be subjected to semi-supervised learning. Semi-supervised learning is a learning method that combines supervised learning with unsupervised learning. Semi-supervised learning uses a large amount of unlabeled data and labeled data at the same time to perform pattern recognition.

[0267] In an optional implementation, the embodiment of the present application gives priority to the recall rate when setting the keyword pattern corresponding to each rumor title type, resulting in a lower accuracy rate. In addition to using the above-mentioned steps D22 to D23 to improve the accuracy rate, the accuracy rate can also be improved through a "rejection strategy" and / or a "title fact retrieval strategy".

[0268] The following first explains the "rejection strategy".

[0269] In the above three implementation methods of step S302, even if step D12 determines that the title belongs to the target rumor title type, or the first prediction result in step D23 is that the title belongs to the target rumor title type, or the fifth prediction result in step D32 is that the title belongs to the target rumor title type, the first information may not be a rumor, and the following explanation is given using the target rumor title type as a supernatural type.

[0270] It is understandable that the first information may be at least part of the content in a movie, TV series, novel, song MV, or variety show, or the first information may be a reproduction of a similar plot in a movie, TV series, novel, song MV, or variety show. Although the title of the first information is of the supernatural type, the first information is not a rumor.

[0271] Based on this, a knowledge base is obtained based on one or more of the materials of TV series, movies, animations, novels and variety shows.

[0272] The material of a TV series includes but is not limited to at least one of the TV series title, TV series alias, names of each character in the TV series, TV series type, TV series release area, and TV series release time. The material of a movie includes but is not limited to at least one of the movie title, movie alias, names of each character in the movie, movie type, movie release area, and movie release time. The material of an animation includes but is not limited to at least one of the animation title, animation alias, names of each character in the animation, animation type, animation release area, and animation release time. The material of a novel includes but is not limited to at least one of the novel title, novel alias, names of each character in the novel, novel type, and novel publication time; the material of a variety show includes but is not limited to at least one of the variety show title, variety show alias, characters in the variety show, variety show release area, and variety show release time.

[0273] Exemplarily, the material of a TV series, the material of a movie, the material of an animation, the material of a novel, and the material of a variety show can each correspond to a plot type, that is, the embodiment of the present application includes at least 5 plot types.

[0274] For example, for TV drama materials, different types of TV dramas correspond to different plot types; the same is true for movies, animations, novels, and variety shows. That is, the number of plot types in the embodiment of the present application = the number of TV drama types + the number of movie types + the number of animation types + the number of novel types + the number of variety show types.

[0275] In an optional implementation, if step D12 determines that the title belongs to the target rumor title type, or the first prediction result in step D23 is that the title belongs to the target rumor title type, or the fifth prediction result in step D32 is that the title belongs to the target rumor title type, it is also necessary to determine the relationship between the title and at least one preset plot type.

[0276] In the embodiment of the present application, “determining the relationship between the title and at least one preset plot type” is referred to as a rejection strategy.

[0277] Exemplarily, in the above three implementations of step S302, step D12, step D23, and step D32 specifically include the following steps E11 to E13.

[0278] Step E11: If it is determined that the title belongs to the target rumor title type, determine the relationship between the title and at least one preset plot type.

[0279] Step E12: If the title belongs to the target plot type among the at least one plot type, obtain the first result indicating that the first information is not a rumor.

[0280] Step E13: If the title does not belong to the at least one plot type, obtain the first result indicating that the first information is a rumor.

[0281] In the step E11 of the embodiment of the present application, there are multiple ways to implement the step of "determining the relationship between the title and at least one preset plot type", and the embodiment of the present application provides but is not limited to the following three.

[0282] The first implementation method involves semantic understanding technology and text preprocessing technology in NLP technology in the AI ​​field, which is explained in detail below.

[0283] The first method for determining the relationship between the title and at least one preset plot type includes the following steps F11 to F12.

[0284] Step F11: Detect whether the title conforms to a preset keyword pattern corresponding to at least one plot type.

[0285] Exemplarily, the keyword pattern corresponding to a plot type may be a keyword set, and the keyword set may include a plurality of keywords corresponding to a plurality of titles belonging to the plot type.

[0286] Exemplarily, for each plot type, a keyword pattern corresponding to the plot type can be obtained from a large number of titles based on semantic understanding technology and text preprocessing technology.

[0287] Exemplarily, for each plot type, the more titles belonging to the plot type are collected, the more comprehensive the keyword pattern corresponding to the plot type is obtained.

[0288] Exemplarily, a keyword pattern corresponding to a plot type may be rule information composed of at least one keyword and used to describe a title that conforms to the plot type.

[0289] For example, a keyword pattern can be expressed using a regular expression.

[0290] Step F12: If the title matches a keyword pattern corresponding to a target plot type in at least one plot type, it is determined that the title belongs to the target plot type.

[0291] Assume that "at least one plot type" includes: movie plot type, TV drama plot type, animation plot type, novel plot type, and variety show type. For example, a title can belong to one or more plot types at the same time, for example, a title belongs to both the novel plot type and the movie plot type. For example, a movie is adapted from a novel.

[0292] Step F13: If the title does not match the keyword pattern corresponding to any plot type, it is determined that the title does not belong to any plot type.

[0293] The second implementation method involves machine learning technology in the AI ​​field, which is explained in detail below.

[0294] The second method for determining the relationship between the title and at least one preset plot type includes the following steps F21 to F22.

[0295] Step F21: input the title into a sixth classifier corresponding to at least one preset plot type.

[0296] The sixth classifier corresponding to a plot type is used to detect whether the input title belongs to the plot type. Different plot types correspond to different sixth classifiers.

[0297] Exemplarily, the manner of constructing the sixth classifier corresponding to each plot type is the same, and the construction process of the sixth classifier corresponding to one plot type is described below.

[0298] The construction process of the sixth classifier corresponding to a plot type A includes steps F211 and F212.

[0299] Step F211: Acquire a plurality of sample titles.

[0300] The plurality of sample titles include a plurality of positive sample titles and negative sample titles. The positive sample titles are titles belonging to plot type A, and the negative sample titles are titles not belonging to plot type A.

[0301] Step F212: Use the multiple sample titles as input of the sixth machine model, use the annotation types corresponding to the multiple sample titles as output, and train the sixth machine model to obtain a sixth classifier.

[0302] Among them, the annotation type corresponding to the positive sample title is plot type A, and the annotation type corresponding to the negative sample title is non-plot type A.

[0303] The description of the sixth machine model refers to the description of the first machine model, which will not be repeated here.

[0304] Steps F211 to F212 are supervised learning. Exemplarily, semi-supervised learning can also be performed on the sixth machine model. Semi-supervised learning is a learning method that combines supervised learning with unsupervised learning. Semi-supervised learning uses a large amount of unlabeled data and labeled data to perform pattern recognition.

[0305] Step F22: Obtain a sixth prediction result output by a sixth classifier corresponding to at least one plot type.

[0306] The sixth prediction result output by the sixth classifier corresponding to a plot type is that the title belongs to the plot type, or that the title does not belong to the plot type.

[0307] The third implementation method involves machine learning technology in the AI ​​field, which is explained in detail below.

[0308] A third method for determining the relationship between the title and at least one preset plot type includes the following steps F31 to F32.

[0309] Step F31: Input the title to the seventh classifier.

[0310] The seventh classifier is used to detect which plot type the input title belongs to among multiple plot types.

[0311] Step F32: Obtain the seventh prediction result output by the seventh classifier.

[0312] The seventh prediction result is that the title belongs to a target plot type among multiple plot types, or the title does not belong to any plot type among the multiple plot types.

[0313] Exemplarily, the process of training the seventh classifier includes steps F311 to F312.

[0314] Step F311: Acquire a plurality of sample titles.

[0315] The plurality of sample titles include a plurality of positive sample titles and negative sample titles. The titles belonging to each plot type are called positive sample titles, and the titles not belonging to any plot type in each plot type are called negative sample titles.

[0316] Step F321: Use multiple sample titles as input of the seventh machine model, use the annotation types corresponding to the multiple sample titles as output, and train the seventh machine model to obtain a seventh classifier.

[0317] Among them, the annotation type corresponding to the positive sample title is the plot type to which the positive sample title belongs, and the annotation type corresponding to the negative sample title is that it does not belong to any plot type.

[0318] The description of the seventh machine model refers to the description of the first machine model, which will not be repeated here.

[0319] The "title fact retrieval strategy" is explained below.

[0320] Exemplarily, if the title matches the keyword pattern corresponding to the target rumor title type in the at least one rumor title type, the "title fact retrieval strategy" can also be executed to further improve the accuracy of determining whether the first information is a rumor.

[0321] Before explaining the "title fact retrieval strategy", the rumor library and / or rumor-refuting library involved is explained first.

[0322] Exemplarily, the rumor library includes multiple rumor documents. Optionally, each rumor document included in the rumor library is a document that is confirmed by humans (such as experts) to be a true rumor.

[0323] Exemplarily, the rumor-refuting library includes a plurality of non-rumor documents. Optionally, the non-rumor documents may include rumor-refuting documents. In the rumor-refuting documents, the user explains that the rumor is false.

[0324] Exemplarily, the rumor library may include rumor documents corresponding to each first keyword combination; a first keyword combination includes one or more keywords that can characterize the corresponding rumor document.

[0325] Exemplarily, the first keyword combination that can represent the rumor document can be extracted from the rumor document based on the semantic understanding technology in the NLP technology in the field of artificial intelligence.

[0326] For example, the first keyword combination that can represent the rumor document is extracted from the rumor document through the one-hot method of the bag-of-words model or the TF-TDF (term frequency-inverse document frequency) method in NLP technology.

[0327] The above-mentioned “rumor documents corresponding to each first keyword combination” is a reverse index. Exemplarily, the reverse index can be an ElasticSearch index.

[0328] Exemplarily, the rumor-debunking library may include non-rumor documents corresponding to each second keyword combination; a second keyword combination includes one or more keywords that can characterize the corresponding non-rumor document.

[0329] The above-mentioned “non-rumor documents corresponding to each second keyword combination” is a reverse index. Exemplarily, the reverse index can be an ElasticSearch index.

[0330] In an optional implementation, the method for implementing the "title fact retrieval strategy" includes the following steps G11.

[0331] Step G11: If the title matches the keyword pattern corresponding to the target rumor title type in the at least one rumor title type, obtain the maximum first correlation between the title and the pre-stored rumor documents; and / or obtain the maximum second correlation between the title and the non-rumor documents.

[0332] Exemplarily, the relevance between the title and the rumor document (or non-rumor document) may be calculated using the BM25 algorithm; or, the relevance between the title and the rumor document (or non-rumor document) may be calculated using cosine similarity.

[0333] Correspondingly, step S304 includes: obtaining the final result based on at least one of the first correlation and the second correlation, the first result, and the second result.

[0334] Secondly Figure 4 Dimension 42 is shown for introduction.

[0335] like Figure 4As shown, dimension 42 is the introduction content. For the description of the introduction content, please refer to the description of dimension 41, i.e. the title. All the words "title" in dimension 41 will be modified to "introduction content". For example, "rumor title type" will be modified to "rumor introduction type", "sample title" will be modified to "sample introduction content", and "title fact retrieval strategy" will be modified to "introduction content fact retrieval strategy".

[0336] For example, the rumor title type may be the same as the rumor content type, or may be different, depending on the actual situation.

[0337] Again Figure 4 The dimension 43 shown is introduced.

[0338] It is understandable that the implementation of step S303 is different depending on the media form of the first information. For example, if the first information is text, then the text content is the text. If the first information is audio, then the text content is the text converted from the audio; if the first information is video, then the text content includes the text information contained in the video image (hereinafter referred to as the first text content), and optionally, may also include the text converted from the audio (hereinafter referred to as the second text content).

[0339] In an optional implementation, there are multiple implementations of step S303, and the embodiments of the present application provide but are not limited to the following two.

[0340] The first implementation method involves the text preprocessing and semantic understanding technology branches in NLP technology in the AI ​​field.

[0341] The first method for implementing step S303 includes the following steps H11 to H13.

[0342] The first implementation involves the rumor library and / or rumor refutation library involved in the "title fact retrieval strategy" or "introduction content fact retrieval strategy". The first implementation is referred to as the "content fact retrieval strategy" in the present application embodiment.

[0343] Step H11: Obtain the largest third correlation among the correlations between the first text content and each pre-stored rumor document, obtain the largest fourth correlation among the correlations between the first text content and each pre-stored non-rumor document, obtain the largest fifth correlation among the correlations between the second text content and each of the rumor documents, and obtain at least one of the largest sixth correlations among the correlations between the second text content and each of the non-rumor documents.

[0344] Exemplarily, if the "content fact retrieval strategy" corresponds to a rumor library, optionally, step H11 includes: obtaining the third largest correlation among the correlations between the first text content and each pre-stored rumor document; and obtaining at least one of the fifth largest correlations among the correlations between the second text content and each of the rumor documents.

[0345] Exemplarily, if the "content fact retrieval strategy" corresponds to a rumor debunking library, optionally, step H11 includes: obtaining the fourth largest correlation among the correlations between the first text content and each pre-stored non-rumor document, and obtaining at least one of the sixth largest correlations between the correlations between the second text content and each of the non-rumor documents.

[0346] Exemplarily, if the "content fact retrieval strategy" corresponds to a rumor library and a rumor refuting library, step H11 may include at least one of the above four steps.

[0347] Exemplarily, the correlation between the first text content and the rumor document (or non-rumor document) can be calculated using the BM25 algorithm; or, the correlation between the first text content and the rumor document (or non-rumor document) can be calculated using cosine similarity.

[0348] The calculation method of the relevance of the second text content rumor document (or non-rumor document) is similar, which will not be repeated here.

[0349] Step H12: Based on at least one of the third relevance and the fourth relevance, obtain the first sub-result representing whether the first information is a rumor.

[0350] Step H13: Based on at least one of the fifth relevance and the sixth relevance, obtain the second sub-result representing whether the first information is a rumor.

[0351] The second result includes at least one of the first sub-result and the second sub-result.

[0352] Exemplarily, if the first sub-result is obtained based on the third correlation, optionally, if the third correlation is greater than a threshold C, the first sub-result characterizes the first information as a rumor; if the third correlation is less than or equal to the threshold C, the first sub-result characterizes the first information as not a rumor.

[0353] If the first sub-result is obtained based on the fourth correlation, optionally, if the fourth correlation is greater than a threshold value D, the first sub-result represents that the first information is not a rumor; if the fourth correlation is less than or equal to the threshold value D, the first sub-result represents that the first information is a rumor.

[0354] If the first sub-result is obtained based on the third correlation and the fourth correlation, optionally, if the third correlation is greater than threshold C, or the fourth correlation is less than or equal to threshold D, the first sub-result characterizes that the first information is a rumor; if the third correlation is less than or equal to threshold C, and the fourth correlation is greater than threshold D, the first sub-result characterizes that the first information is not a rumor.

[0355] If the first sub-result is obtained based on the third correlation and the fourth correlation, optionally, weights can be assigned to the third correlation and the fourth correlation, for example, a weight of 3 is assigned to the third correlation and a weight of 4 is assigned to the fourth correlation. Then, if the third correlation*weight3-fourth correlation*weight4 is greater than the threshold E, the first sub-result characterizes that the first information is a rumor; if the third correlation*weight3-fourth correlation*weight4 is less than or equal to the threshold E, the first sub-result characterizes that the first information is not a rumor.

[0356] The fifth and sixth correlations are similar and will not be elaborated here.

[0357] The second implementation involves machine learning in the field of AI.

[0358] The second method for implementing step S303 includes the following steps H21 to H23.

[0359] Step H21: Execute at least one of the steps of inputting the first text content into the eighth classifier and inputting the second text content into the ninth classifier.

[0360] Among them, the eighth classifier is used to detect the third probability that the first input text content is a rumor; the ninth classifier is used to detect the fourth probability that the second input text content is not a rumor.

[0361] Exemplarily, the construction process of the eighth classifier includes the following steps H211 to H212.

[0362] Step H211: Obtain multiple sample documents.

[0363] The multiple sample documents include multiple positive sample documents and negative sample documents. The positive sample documents are rumor documents, and the negative sample documents are non-rumor documents.

[0364] Step H212: Use multiple sample documents as input of the eighth machine model, use the annotation types corresponding to the multiple sample documents as output, and train the eighth machine model to obtain an eighth classifier.

[0365] Among them, the label type corresponding to the positive sample document is rumor document, and the label type corresponding to the negative sample document is non-rumor document.

[0366] Exemplarily, the construction process of the ninth classifier is similar to the construction process of the eighth classifier, except that the positive sample document in step H211 is a non-rumor document, and the negative sample document is a rumor document.

[0367] For the description of the eighth and ninth classifiers, please refer to the description of the first classifier, which will not be repeated here.

[0368] Step H22: Obtain a first result based on at least one of the eighth prediction result output by the eighth classifier and the ninth prediction result output by the ninth classifier.

[0369] Exemplarily, the eighth prediction result is the third probability that the first information is a rumor, and the ninth prediction result is the fourth probability that the first information is a rumor.

[0370] Exemplarily, the third probability is equivalent to the third correlation, and the fourth probability is equivalent to the fourth correlation. For the description of step H22, please refer to the description of step H11, which will not be repeated here.

[0371] Combined with the detailed description of dimensions 41, 42, and 43, the following is a description of Figure 4 The technical architecture of the rumor identification method shown in the figure is further supplemented to obtain Figure 5 A schematic diagram of another technical system architecture of the rumor identification method shown.

[0372] For example, the rumor title type may be the same as the rumor introduction type, or may be different, depending on the actual situation. Figure 5 The following example is used to illustrate the case where the rumor title type and rumor introduction type are the same.

[0373] against Figure 5 For instructions, see Figure 4 , I will not go into details here.

[0374] It is understandable that the implementation method of obtaining the text content of the first information in step S301 is different depending on the media form of the first information. For example, if the first information is text, then the text is the text content. If the first information is audio, then the text is the text converted from the audio. If the first information is video, the implementation method of obtaining the first text content of the first information in step S301 includes the following steps J11 to J13.

[0375] Step J11: Acquire multiple frames of video images from the video.

[0376] Exemplarily, a frame of video image can be extracted from the video at every preset time interval to obtain multiple frames of video image; Exemplarily, multiple frames of video image can be randomly extracted from the video. Exemplarily, each key frame contained in the video can be used as a video image.

[0377] Step J12: performing optical character recognition on the multiple frames of the video image to obtain text information contained in at least one frame of the video image.

[0378] The above involves OCR technology in computer vision technology in the AI ​​field.

[0379] Exemplarily, the text information contained in a frame of video image may include: at least one of subtitle information corresponding to the screen displayed by the video, barrage, and the name of an application displaying the video.

[0380] The following describes steps J11 to J13 using a specific example.

[0381] like Figure 6 , which is a schematic diagram of multiple frames of video images in a video provided in an embodiment of the present application.

[0382] Figure 6 4 frames of video images are shown, namely video image 61, video image 62, video image 63 and video image 64, and, Figure 6 The 4 video frames shown are arranged in chronological order from early to late. Figure 6 This is only an example, and the number of video images obtained in step J11 is not limited.

[0383] Exemplarily, all of the multiple frames of video images contain text information, or only some of the multiple frames of video images contain text information.

[0384] like Figure 6 As shown, video image 61, video image 63, and video image 64 include text information, but video image 62 does not include text information.

[0385] Step J13: Obtain the first text content based on the text information contained in at least one frame of the video image.

[0386] Exemplarily, step J13 involves the text preprocessing and semantic understanding technology branches in NLP technology in the AI ​​field.

[0387] like Figure 6 As shown, the text information in video image 61 includes: "Tencent Video" and "Two tigers, two tigers, run fast, run fast,"; the text information in video image 63 includes: "Tencent Video," "One has no eyes, one has no tail," and "Haha! So funny"; the text information in video image 64 includes: "Tencent Video," "How strange! How strange!" and "User 1: Haha! How funny."

[0388] Among them, "Tencent Video" in video image 61, video image 63 and video image 64 is the name of the application that displays the video; "Haha! So funny" in video image 63 and video image 64 is the barrage, and the barrage displayed by video image 63 and video image 64 is the same barrage. Since video image 64 is a video image located behind video image 63, the barrage has moved from the right side to the left side; "Two tigers, two tigers, run fast, run fast," in video image 61, "One has no eyes, and one has no tail," in video image 62, and "How strange! How strange!" in video image 63 are the subtitle information corresponding to the pictures displayed in the video.

[0389] Exemplarily, different video images may include different bullet comments or the same bullet comments.

[0390] The first text content may include: "Tencent Video", "Two tigers, two tigers, run fast, run fast," "Tencent Video", "One has no eyes, one has no tail," "Haha! So funny," "Tencent Video", "So strange! So strange!", "Haha! So funny".

[0391] There are multiple ways to implement step J12 in the embodiment of the present application, and the embodiment of the present application provides but is not limited to the following two.

[0392] The first implementation of step J12 includes: splicing the text information contained in at least one frame of the video image in the order of the time when the video image is located in the video from early to late to obtain the first text content.

[0393] Still Figure 6 For example, assume that video image 61, video image 62, video image 63 and video image 64 are in the order of the time from earliest to latest in the video: video image 61, video image 62, video image 63 and video image 64.

[0394] The contents of video image 61, video image 62, video image 63 and video image 64 are stitched together to obtain: Tencent video Two tigers, two tigers, run fast, run fast, Tencent video One tiger has no eyes, haha! So funny One tiger has no tail, Tencent video So strange! So strange! Haha! So funny.

[0395] It is understandable that the same content may exist in at least two temporally consecutive video image frames, for example, the barrage in video image 63 and video image 64 , or “Tencent Video” in video image 63 and video image 64 .

[0396] Since the same content is repeated multiple times, the number of corresponding keywords included in the first text content will be affected, resulting in inaccurate calculation of the relevance of the first text content with the rumor document or the non-rumor document.

[0397] Still Figure 6 For example, "Haha! How funny" appears twice in the first text content, but in actual application, "Haha! How funny" actually only appears once. Since the number of "Haha! How funny" has increased, the calculated weight value of "Haha! How funny" may be larger, resulting in inaccurate calculation of the relevance of the first text content with the rumor document or the non-rumor document in the subsequent calculation.

[0398] Based on this, the second implementation method is introduced.

[0399] The second implementation method of step J12 includes the following steps J121 to J122.

[0400] Step J121: Obtain target text information corresponding to at least one set of video image sets.

[0401] A group of the video image sets includes at least two temporally adjacent frames of the video image, and the target text information corresponding to a group of the video image sets is text information contained in at least two temporally adjacent frames of the video image set.

[0402] Exemplarily, the number of video images involved in "at least two frames of the video images" is greater than or equal to 2, and the specific number may be determined based on actual conditions.

[0403] Still Figure 6 For example, assuming that the number of video images involved in "at least two frames of the video images" is equal to 2, Figure 6 There are three corresponding video image sets, namely video image set A1 (including video image 61 and video image 62), video image set A2 (including video image 62 and video image 63), and video image set A3 (including video image 63 and video image 64).

[0404] Since the video images 61 and 62 do not include the same text information, the video image set A1 does not correspond to the target text information; similarly, the video image set A2 does not correspond to the target text information.

[0405] Since the video image 63 and the video image 64 include the same text information “Tencent Video” and “Haha! So funny”, the target text information corresponding to the video image set A3 includes “Tencent Video” and “Haha! So funny”.

[0406] There are many ways to implement the acquisition of target text information corresponding to a set of video images in step J121. The embodiments of the present application provide but are not limited to the following two.

[0407] The first method for obtaining target text information corresponding to a set of video images includes the following steps J1211 to J1213.

[0408] The first implementation method involves the OCR technology branch in the image semantic understanding technology in computer vision in the AI ​​field, which is explained in detail below.

[0409] Step J1211: Identify text information carried by at least one location area respectively contained in at least two temporally adjacent frames of video images contained in a video image set.

[0410] Step J1212: If the target text information respectively contained in at least two temporally adjacent frames of the video images is the same, the target text information corresponding to the group of video images is obtained.

[0411] Exemplarily, in the embodiment of the present application, the information carried by a location area is referred to as text information, and when comparing the text information respectively contained in at least two frames of the video images, all the information carried by a location area is compared as a whole.

[0412] like Figure 6 When comparing the text information of the video images 63 and 64, "one has no eyes, one has no tail," will be used as a text information to be compared with "Haha! So funny," "So weird! So weird!", and "Tencent Video" in the video image 64. "One has no eyes" will not be taken out separately to be compared with "Haha! So funny," "So weird! So weird!", and "Tencent Video" in the video image 64.

[0413] The second implementation method involves the OCR technology branch in the image semantic understanding technology in computer vision in the AI ​​field, which is explained in detail below.

[0414] The second method for obtaining target text information corresponding to a set of video images includes the following steps J1213 to J1215.

[0415] Step J1213: Determine the size of the position area carrying text information respectively contained in at least two temporally adjacent frames of the video image set.

[0416] Exemplarily, the size of the location area may include the shape of the location area and a parameter that can represent the size of the location area. For example, if the location area is circular, then the parameter that can represent the size of the location area is the diameter or radius of the circle; if the location area is rectangular, then the parameter that can represent the size of the location area is the length of the location area and the width of the location area.

[0417] Exemplarily, the size of the location area carrying text information may be identified based on the text area detection technology in the OCR technology.

[0418] Step J1214: If the sizes of the target location areas in the location areas respectively included in at least two temporally adjacent frames of the video images are the same, detect whether the text information in the target location areas respectively included in at least two temporally adjacent frames of the video images are the same.

[0419] Exemplarily, the text carried by the target location area may be acquired based on the text recognition technology in the OCR technology.

[0420] For example, the same position area carrying text information in different video images may be located at different positions in the video images, such as bullet screens, but the size of the same position area carrying text information in different video images is the same. Figure 6 The location area of ​​the bullet comments included in the video image 63 and the video image 64 shown is the above-mentioned target location area.

[0421] It is understandable that the size of the location area carrying the same text information in different video images must be the same, but if different video images contain target location areas of the same size, the text information carried by the target location areas in different video images may not be the same, so further judgment is needed.

[0422] Step J1215: If the text information in the target position area respectively included in at least two temporally adjacent frames of the video image is the same, the text information in the target position area respectively included in at least two temporally adjacent frames of the video image is determined as the target text information.

[0423] Step J22: performing a deduplication operation on target text information corresponding to at least one set of the video image sets from the text information contained in at least one frame of the video image to obtain the first text content.

[0424] Exemplarily, a deduplication operation is performed on the target text information respectively corresponding to at least one group of the video image sets, that is, one target text information is retained and other target text information is deleted.

[0425] Still Figure 6Taking an example to illustrate, the first text content obtained includes: "Tencent Video", "Two tigers, two tigers, run fast, run fast," "Tencent Video", "One has no eyes, one has no tail," "Haha! How funny," "How strange! How strange!".

[0426] Exemplarily, performing a deduplication operation on the target text information corresponding to at least one group of the video image sets includes: if the first target text information corresponding to greater than or equal to a preset number of video image sets in the target text information corresponding to at least one group of the video image sets are the same, deleting all the first target text information; if the second target text information corresponding to less than a preset number of video image sets in the target text information corresponding to at least one group of the video image sets are the same, retaining one second target text information.

[0427] Exemplarily, the preset number may be determined based on the number of the multi-frame video images obtained in step J11.

[0428] Still Figure 6 As shown, if the preset number is 3, assuming that the video image 62 also includes a video image, then "Tencent Video" can be deleted. The first text content obtained includes: "Two tigers, two tigers, run fast, run fast," "One has no eyes, one has no tail," "Haha! So funny," "So weird! So weird!".

[0429] Exemplarily, there are multiple ways to implement step J13, and the embodiments of the present application provide but are not limited to the following two.

[0430] The first implementation method of J13 includes: splicing the text information contained in at least one frame of the video image in the order of the time when the video image is located in the video from early to late to obtain the first text content.

[0431] by Figure 6 Taking the first text content in the sentence as follows: "Two tigers, two tigers, running fast, running fast," "One has no eyes, one has no tail," "Haha! How funny," "How strange! How strange!" as an example, the first text content obtained by splicing is: Two tigers, two tigers, running fast, running fast, one has no eyes, one has no tail, "Haha! How funny and strange! How strange!

[0432] For example, if the text information contained in the video image contains punctuation marks, then the sentences contained in the obtained first text content have been divided by the punctuation marks. If the text information contained in the video image does not contain punctuation marks, then the sentences contained in the obtained first text content are not divided, based on which the second implementation method is introduced.

[0433] The second implementation method involves the semantic understanding technology branch of NLP technology in the AI ​​field, which is explained in detail below.

[0434] The second implementation method of J13 includes the following steps J131 to J133.

[0435] Step J131: splicing the text information contained in at least one frame of the video image in the order of the time when the video image is located in the video from early to late to obtain candidate text content.

[0436] Step J132: Determine at least one frame of target video image that does not contain text information from the multiple frames of video images.

[0437] like Figure 6 Video image 62 in.

[0438] Step J133: Based on the time when the at least one frame of target video image is located in the video, the candidate text content is divided into sentences to obtain the first text content.

[0439] Exemplarily, since the target video image does not contain text information, it means that there is a pause at the moment when the target video image is located in the video, so the candidate text content can be divided into sentences based on this feature.

[0440] Exemplarily, based on the semantic understanding technology in NLP and the time when the at least one frame of the target video image is located in the video, the candidate text content can be divided into sentences to obtain the first text content.

[0441] In summary, the above introduces a multi-dimensional strategy combination. The strategies of each dimension (for example, auxiliary black judgment strategy, direct white judgment strategy, rejection strategy, title fact retrieval strategy, introduction content fact retrieval strategy, content fact retrieval strategy) are combined to complement each other and jointly determine whether the first information is a rumor. It is also the main structural part of the video rumor identification architecture in the embodiment of the present application.

[0442] It is understandable that each first information corresponds to a publisher, that is, the user account corresponding to the user is the publisher of the first information. Modeling the behavior rumor tendency portrait corresponding to the user account is also of great significance for the overall architecture of rumor discrimination. That is, it is also possible to identify whether the first information is a rumor from the user account level that publishes the first information.

[0443] The rumor identification method provided in the embodiment of the present application also includes the following step K11.

[0444] Step K11: Obtain the rumor publishing tendency of the user account that published the first information.

[0445] The rumor publishing tendency is used to characterize the probability that the first information published by the user account is a rumor.

[0446] Exemplarily, step K11 includes the following steps K111 to step K112 during implementation.

[0447] Step K111: Obtain a behavioral rumor tendency portrait of the user account.

[0448] The behavioral rumor tendency portrait includes the number of historically forwarded rumors, the number of historically published rumors, the number of forwarded rumors within a preset time period from the current time, the number of rumors published within the preset time period, the registration time of the user account, the authentication type of the user account, the number of historically published information, the number of historically forwarded information, whether the user account has a signature, if so, the signature content of the user account, the number of followers of the user account, the behavioral rumor tendency portraits corresponding to the accounts followed by the user account, and one or more of the information representing whether the harm level of the rumors published by the user account is higher than the harm level of the rumors forwarded by the user account.

[0449] The following is an explanation of the contents of the behavioral rumor tendency profile.

[0450] The number of rumors forwarded in history refers to the total number of rumors posted by other users that the user account has forwarded since the time of registration to the current time. The number of rumors posted in history refers to the total number of rumors posted by the user account from the time of registration to the current time. The number of rumors forwarded within the preset time period from the current time refers to the total number of rumors posted by other users that the user account has forwarded within the preset time period.

[0451] Assuming that the current time is August 27, 2020, and the preset time period is 1 month, then the number of rumors forwarded within the preset time period from the current time refers to the total number of rumors forwarded by the user account from July 28, 2020 to August 27, 2020.

[0452] The number of rumors published within the preset time period refers to the total number of rumors published by the user account itself within the preset time period.

[0453] Exemplarily, the authentication type of the user account may be one or more of a government authentication type, an enterprise authentication type, an institutional group authentication type, a media authentication type, an individual user authentication type, a website authentication type, and an individual authentication type.

[0454] The number of historical published information refers to the total number of information (including rumors and non-rumors) published by the user account from the time the user account was registered to the current time. The number of historical forwarded information refers to the total number of information (including rumors and non-rumors) forwarded by the user account from the time the user account was registered to the current time.

[0455] The signature of a user account is the information filled in by the user about the user account or himself.

[0456] For example, the signature of a user account may be able to represent the inner thoughts of the user.

[0457] For example, if the number of followers of the user account is larger, the impact of the user account forwarding or publishing rumors will be greater.

[0458] It is understandable that if a user wants to see information posted or forwarded by a certain account, he or she may follow the account, that is, the accounts followed by the user account can to a certain extent characterize the type of information the user wants to see, so the behavioral rumor tendency portrait corresponding to the user account can be determined in combination with the behavioral rumor tendency portraits corresponding to the accounts followed by the user account.

[0459] It is understandable that different rumors have different levels of harm. For example, different rumors can be classified based on the degree of harm they cause to society. If the harm level of the rumor published by the user account is lower than the harm level of the rumor forwarded by the user account, it means that the user account may be a person who passively publishes rumors. For example, after seeing some rumors, he believed them to be true and felt panic, so he published some related rumors; if the harm level of the rumor published by the user account is higher than or equal to the harm level of the rumor forwarded by the user account, it means that the user account is a person who actively creates rumors.

[0460] The above-mentioned information characterizing whether the harm level of the rumors posted by the user account is higher than the harm level of the rumors forwarded by the user account includes: the average value of the harm levels of the rumors posted by the user account, the maximum value of the harm levels of the rumors posted by the user account, the minimum value of the harm levels of the rumors posted by the user account, the average value of the harm levels of the rumors forwarded by the user account, the maximum value of the harm levels of the rumors forwarded by the user account, and the minimum value of the harm levels of the rumors forwarded by the user account.

[0461] It is understandable that users who post a large number of rumors are often like online trolls. For example, user A posts dozens of rumors in a very concentrated period of time, most of which are manually determined to be rumors, and user A has not posted or forwarded any information since then. Optionally, a behavioral rumor tendency profile can be established for such users. Optionally, a behavioral rumor tendency profile can be established for all users.

[0462] For example, the user's basic information can be used to determine the user's basic credibility, and then combined with the behavioral information of the rumors posted or forwarded by the user to establish the user's behavioral rumor tendency portrait, to assist in determining whether the information posted or forwarded by the user should be judged as a rumor, and to increase the confidence of the final result of the above rumor identification method.

[0463] Among them, the basic information includes the number of historical forwarded rumors, the number of historical published rumors, the number of forwarded rumors within a preset time period from the current time, the number of rumors published within the preset time period, the registration time of the user account, the authentication type of the user account, the number of historical published information, the number of historical forwarded information, the signature content of the user account, the number of followers of the user account, and one or more of the behavioral rumor tendency portraits corresponding to the accounts followed by the user account.

[0464] Among them, the behavior information of rumors posted or forwarded by users includes information indicating whether the harm level of the rumors posted by the user account is higher than the harm level of the rumors forwarded by the user account.

[0465] For example, the user's behavior rumor tendency profile is constantly changing with the user's own behavior. Generally speaking, if a user posts more information that is determined to be a rumor in a shorter period of time, the rumor tendency corresponding to the user's behavior rumor tendency profile will increase faster. If a rumor has been posted or forwarded in the past, but the user has not posted a rumor recently, then the user's rumor tendency will continue to decrease over time.

[0466] Exemplarily, user accounts with a tendency to publish rumors higher than a threshold F may be added to the suspicious user database. Exemplarily, a more stringent judgment is required for the first information published or forwarded by the user accounts in the suspicious user database.

[0467] For example, if the user account belongs to a suspicious user database, as long as at least one of the results corresponding to the multiple dimensions involved in step S304 indicates that the first information is a rumor, then the final result indicates that the first information is a rumor.

[0468] If the user account does not belong to the user database, at least two of the results corresponding to the multiple dimensions in step S304 both indicate that the first information is a rumor, and the final result may indicate that the first information is a rumor.

[0469] Step K112: Determine the tendency to publish rumors based on the behavioral rumor tendency portrait of the user account.

[0470] Exemplarily, step S304 includes: determining a final result indicating whether the first information is a rumor based on the rumor publishing tendency, the first result, and the second result.

[0471] It is understandable that the final result obtained in step S304 contains information characterizing whether the first information is a rumor, which may be different from the real information characterizing whether the first information is a true rumor. For example, the final result contains information characterizing that the first information is a rumor, but the real information characterizes that the first information is not a rumor; or, the final result contains information characterizing that the first information is not a rumor, but the real information characterizes that the first information is a rumor.

[0472] Since the information including the final result indicating whether the first information is a rumor is obtained based on the results corresponding to multiple dimensions, such as Figure 4 or Figure 5 As shown, the final result includes information characterizing whether the first information is a rumor, which can be determined based on at least one of the results corresponding to dimension 41, the results corresponding to dimension 42, the results corresponding to dimension 43, and the results corresponding to dimension 46, so the target dimension that is different from the true result in the results corresponding to these dimensions is determined.

[0473] If the target dimension corresponds to a classifier, the first information is used as a sample to retrain the classifier involved in the target dimension. If the target dimension corresponds to a keyword pattern, the keyword pattern involved in the target dimension is supplemented based on the first information. If the target dimension corresponds to a rumor library or a rumor-refuting library, if the first information is a rumor, the first information can be added to the rumor library; if the first information is not a rumor, the first information can be added to the rumor-refuting library.

[0474] In summary, a positive closed-loop supplement to the rumor identification method is achieved.

[0475] The process of "obtaining a true result indicating whether the first information is a rumor" is described below. There are many ways to "obtain a true result indicating whether the first information is a rumor", and the embodiments of the present application provide but are not limited to the following two.

[0476] The first implementation method: after step S304, the following steps L11 to L12 are also included.

[0477] Step L11: If the final result includes information indicating that the first information is a rumor, the final result and the first information are sent to the second electronic device.

[0478] This allows the expert to manually review the first information in combination with the final result through the second electronic device.

[0479] Step L12: Receive the true result of the manual review result fed back by the second electronic device.

[0480] The second implementation method includes the following steps L21 to L22.

[0481] Step L21: Obtain multiple user-generated content corresponding to the first information, one of the user-generated content is barrage information or comment information.

[0482] Step L22: If the number of user-generated contents reporting the first information as a rumor among the plurality of user-generated contents is greater than or equal to a preset threshold, the true result indicating that the first information is a rumor is obtained.

[0483] The second implementation method includes the following steps L31 to L33.

[0484] Step L31: Obtain multiple user-generated content corresponding to the first information, one of the user-generated content is barrage information or comment information.

[0485] Step L32: If the number of user-generated contents reporting the first information as a rumor among the plurality of user-generated contents is greater than or equal to a preset threshold, the first information is sent to a second electronic device.

[0486] This allows an expert to manually review the first information through the second electronic device.

[0487] Step L33: Receive the real result fed back by the second electronic device.

[0488] In summary, further detection of the first information that has been released and determined to be non-rumor is carried out through UGC (User Generated Content), which makes up for the situation where the final result is misjudged.

[0489] In an optional implementation, if Figure 3 In the rumor identification method shown, the first information to be identified corresponds to user-generated content. In this case, the rumor identification method can also obtain multiple user-generated contents corresponding to the first information; and obtain the number of user-generated contents in the multiple user-generated contents that represent reporting the first information as a rumor.

[0490] In the embodiment of the present application, the above-mentioned "obtaining the number of user-generated contents representing reporting that the first information is a rumor from the multiple user-generated contents" is referred to as the "user-generated content strategy".

[0491] In step S304, the final result may be determined in combination with the number, the first result, and the second result.

[0492] In an embodiment of the present application, shortly after the first information is released, that is, in the early stage of the first information, it can be determined whether the first information is a rumor through the six dimensions provided in the embodiment of the present application and at least one of the "user-generated content strategies". If the final result contains information characterizing that the first information is a rumor, since the above-mentioned artificial intelligence rumor recognition method cannot guarantee that the information characterizing that the first information is a rumor contained in the final result is 100% correct, for example, it is necessary to manually (who can be an expert or operator) review the first information in combination with the final result.

[0493] After a period of time after the first information is released, that is, in the later stage of the first information, the true result corresponding to the first information can be obtained. If the true result indicates that the first information is a rumor, the first information can be included in the rumor database. If the true result indicates that the first information is not a rumor, the first information can be included in the rumor debunking database, thereby continuously replenishing the rumor database and the rumor debunking database to form a virtuous circle.

[0494] For the first information that the real result is inconsistent with the final result, the cause can be traced back to train the corresponding classifier, or the keyword pattern can be adjusted, and the behavioral rumor tendency portrait of the user account can be updated, and its corresponding rumor publishing tendency can be updated.

[0495] In summary, step S304 in the embodiment of the present application can obtain the final result based on whether the first information belongs to a carousel video, the relationship between the first information and the preset non-rumor video types, whether the first information is a machine-dubbed video, the relationship between the introduction content corresponding to the first information and at least one preset rumor introduction type, the first relevance, the second relevance, the user account's tendency to publish rumors, at least one of the number of user-generated content corresponding to the first information that represents reporting the first information as a rumor, and the first result and the second result.

[0496] Exemplarily, the final result includes at least two parameters of the first result and the second result of the above 10 parameters and information characterizing whether the first information is a rumor.

[0497] Exemplarily, weights may be set for the above 10 parameters respectively.

[0498] Exemplarily, if the first information belongs to a carousel video, the probability that the first information is a rumor is probability 1; if the first information does not belong to a carousel video, the probability that the first information is a rumor is probability 2; if the first information does not belong to any non-rumor video type, the probability that the first information is a rumor is probability 3; if the first information belongs to a certain non-rumor video type, the probability that the first information is a rumor is probability 4; if the first information is a machine-dubbed video, the probability that the first information is a rumor is probability 5; if the first information is not a machine-dubbed video, the probability that the first information is a rumor is probability 6; if the introduction content corresponding to the first information belongs to the target rumor introduction type in at least one rumor introduction type, the probability that the first information is a rumor is probability 7; if the introduction content corresponding to the first information does not belong to any rumor introduction type, the probability that the first information is a rumor is probability 8; if the number is greater than or equal to the preset threshold, the probability that the first information is a rumor is probability 9; if the number is less than the preset threshold, the probability that the first information is a rumor is probability 10.

[0499] It can be understood that the first relevance, the second relevance, the first result, and the second result themselves can be expressed by probabilities, so there is no need to set their probabilities.

[0500] Among them, probability 2 is less than probability 1; probability 3 is less than probability 4, optionally, probability 3 can be 0; probability 5 is greater than probability 4; probability 7 is greater than probability 8, optionally, probability 8 can be 0, and probability 9 is greater than probability 10.

[0501] The above-mentioned probabilities 1 to 10 may be set according to the criticality of each dimension, and the weights corresponding to the 10 parameters are set according to the criticality of each dimension.

[0502] Exemplarily, if the sum of the product of the probability corresponding to each parameter and the corresponding weight is greater than or equal to the threshold G, the final result is that the first information is a rumor; if it is less than the threshold G, the final result is that the first information is not a rumor.

[0503] Exemplarily, at least two parameters among the above 10 parameters (including at least the first result and the second result) can be combined to comprehensively determine whether the first information is a rumor.

[0504] For example, at least two parameters (including at least the first result and the second result) of the above 10 parameters are input into the tenth classifier to obtain the tenth prediction result output by the tenth classifier. The tenth prediction result is that the first information is a rumor or the first information is not a rumor.

[0505] The construction process of the tenth classifier includes the following steps M11 to M13.

[0506] Step M11: Obtain a parameter set corresponding to multiple sample information.

[0507] A parameter set corresponding to one sample information includes at least the first result and the second result of the above 10 parameters.

[0508] The multiple sample information includes positive sample information and negative sample information, wherein the positive sample information is information that is a rumor, and the negative sample information is information that is not a rumor.

[0509] Step M12: Input the parameter sets corresponding to the multiple sample information into the tenth machine model, use the real results corresponding to the multiple sample information as the output of the tenth machine model, train the tenth machine model, and obtain the tenth classifier.

[0510] For the description of the tenth classifier, please refer to the description of the first classifier, which will not be repeated here.

[0511] With the research and advancement of artificial intelligence technology, artificial intelligence technology has been studied and applied in many fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, automatic driving, drones, robots, smart medical care, smart customer service, etc. I believe that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0512] The method is described in detail in the embodiments disclosed in the above-mentioned application. The method of the application can be implemented by various forms of devices, so the application also discloses various devices. Specific embodiments are given below for detailed description.

[0513] like Figure 7 As shown, it is a schematic diagram of a rumor identification device provided in an embodiment of the present application, and the device includes: a first acquisition module 701, a second acquisition module 702, a determination module 703 and a third acquisition module 704, wherein:

[0514] A first acquisition module 701 is used to acquire a title and text content corresponding to the first information to be identified, wherein the text content includes content representing the first information acquired directly or indirectly;

[0515] A second acquisition module 702, configured to obtain a first result indicating whether the first information is a rumor based on a relationship between the title and at least one preset rumor title type;

[0516] A determination module 703, configured to determine a second result indicating whether the first information is a rumor based on the text content;

[0517] The third acquisition module 704 is used to obtain a final result based on the first result and the second result, and the final result includes information indicating whether the first information is a rumor.

[0518] In an optional implementation, the first information is a video, and the third acquisition module includes: a first determination unit, a first acquisition unit, a second determination unit, at least one of the second acquisition units, and a third determination unit, wherein: the first determination unit is used to determine whether the video is a carousel video based on a picture change feature of the video, and the picture change feature at least includes a dwell time corresponding to at least one picture contained in the video; the first acquisition unit is used to determine whether the video is a machine-dubbed video based on attribute information of the audio contained in the video; the second determination unit is used to input the video into the third classifier corresponding to each non-rumor video type, and obtain the third classifier outputted by each non-rumor video type. Three prediction results, a third prediction result output by a third classifier corresponding to the non-rumor video type is that the video belongs to the non-rumor video type or the video does not belong to the non-rumor video type; a second acquisition unit is used to determine the relationship between the introduction content of the video and at least one preset rumor introduction type; a third determination unit is used to determine the final result characterizing whether the first information is a rumor based on whether the video belongs to a carousel video, the target prediction results output by the target classifiers corresponding to each of the non-rumor video types, whether the video is a machine-dubbed video, and at least one of the relationship between the introduction content of the video and at least one preset rumor introduction type, as well as the first result and the second result.

[0519] In an optional implementation, the second acquisition module includes: a third acquisition unit, a first input unit, and a fourth acquisition unit, wherein:

[0520] A third acquisition unit is used to acquire a first result indicating that the first information is not a rumor if the title does not conform to the keyword pattern corresponding to the at least one rumor title type; a keyword pattern corresponding to the rumor title type is rule information composed of at least one keyword and used to describe a title that conforms to the rumor title type;

[0521] A first input unit, for inputting the title into a first classifier corresponding to the target rumor title type if the title matches a keyword pattern corresponding to the target rumor title type in the at least one rumor title type, the first classifier being used to detect whether the input title belongs to the target rumor title type;

[0522] The fourth acquisition unit is used to obtain the first result based on the first prediction result output by the first classifier, wherein the first prediction result is that the title belongs to the target rumor title type, or the title does not belong to the target rumor title type.

[0523] In an optional implementation, the fourth acquisition unit includes: a first determining subunit, a first acquisition subunit, and a second acquisition subunit, wherein:

[0524] A first determining subunit is configured to determine a relationship between the title and at least one preset plot type if the first prediction result is that the title belongs to the target rumor title type;

[0525] A first acquisition subunit is configured to acquire, if the title belongs to a target plot type among the at least one plot type, the first result indicating that the first information is not a rumor;

[0526] The second acquisition subunit is used to acquire the first result representing that the first information is a rumor if the title does not belong to the at least one plot type.

[0527] In an optional implementation, the rumor identification device also includes: a fourth acquisition module, which is used to obtain the largest first correlation between the title and each pre-stored rumor document if the title matches the keyword pattern corresponding to the target rumor title type in the at least one rumor title type; and / or, obtain the largest second correlation between the title and each non-rumor document; the corresponding third acquisition module is specifically used to obtain the final result based on at least one of the first correlation and the second correlation, the first result and the second result.

[0528] In an optional implementation, the text content includes first text content, the first information is a video, and the first acquisition module includes: a fifth acquisition unit, a sixth acquisition unit, and a seventh acquisition unit, wherein:

[0529] A fifth acquisition unit, used to acquire multiple frames of video images from the video;

[0530] a sixth acquisition unit, configured to perform optical character recognition on the plurality of frames of the video image to obtain text information contained in at least one frame of the video image;

[0531] The seventh acquisition unit is used to obtain the first text content based on the text information contained in at least one frame of the video image.

[0532] In an optional implementation, the sixth acquisition unit includes: a third acquisition subunit and a fourth acquisition subunit, wherein:

[0533] A third acquisition subunit is used to acquire target text information corresponding to at least one group of video image sets, wherein a group of the video image sets includes at least two frames of the video image that are adjacent in time, and the target text information corresponding to a group of the video image sets is text information contained in at least two frames of the video image that are adjacent in time and included in the video image set;

[0534] The fourth acquisition subunit is used to perform a deduplication operation on target text information corresponding to at least one group of the video image sets from the text information contained in at least one frame of the video image, so as to obtain the first text content.

[0535] In an optional implementation, determining the target text information corresponding to a set of video images in the third acquisition subunit includes: a first acquisition submodule, a detection submodule, and a determination submodule, wherein:

[0536] A first acquisition submodule is used to acquire the size of the position area carrying text information respectively contained in at least two frames of the video image set that are temporally adjacent;

[0537] A detection submodule, configured to detect whether text information in the target location areas respectively included in at least two temporally adjacent frames of the video images is the same if the sizes of the target location areas in the location areas respectively included in at least two temporally adjacent frames of the video images are the same;

[0538] The determination submodule is used to determine the text information in the target position area respectively contained in at least two temporally adjacent frames of the video image as the target text information if the text information in the target position area respectively contained in at least two temporally adjacent frames of the video image is the same.

[0539] In an optional implementation, the seventh acquisition unit is specifically used to: splice the text information contained in at least one frame of the video image in order from early to late when the video image is located in the video to obtain the first text content.

[0540] In an optional implementation, the seventh acquisition unit includes: a sorting subunit, a second determining subunit, and a sentence dividing subunit, wherein:

[0541] A sorting subunit, used for splicing the text information contained in at least one frame of the video image in the order of the time when the video image is located in the video from early to late to obtain candidate text content;

[0542] A second determining subunit is used to determine at least one frame of target video image that does not contain text information from the multiple frames of video images;

[0543] The sentence division subunit is used to divide the candidate text content into sentences based on the time when the at least one frame of target video image is located in the video to obtain the first text content.

[0544] In an optional implementation, the text content includes second text content, and the first acquisition module further includes: a speech recognition unit, configured to perform speech recognition on the audio included in the video to obtain the second text content.

[0545] In an optional implementation, the second result includes at least one of the first sub-result and the second sub-result, and the determination module includes: at least one of an eighth acquisition unit, a ninth acquisition unit, a tenth acquisition unit, and an eleventh acquisition unit, and at least one of a twelfth acquisition unit and a thirteenth acquisition unit, wherein:

[0546] an eighth acquisition unit, for acquiring the third largest correlation among the correlations between the first text content and each pre-stored rumor document; a ninth acquisition unit, for acquiring the fourth largest correlation among the correlations between the first text content and each pre-stored non-rumor document; a tenth acquisition unit, for acquiring the fifth largest correlation among the correlations between the second text content and each rumor document; an eleventh acquisition unit, for acquiring the sixth largest correlation among the correlations between the second text content and each non-rumor document;

[0547] The twelfth acquisition unit is used to obtain the first sub-result representing whether the first information is a rumor based on at least one of the third correlation and the fourth correlation; the thirteenth acquisition unit is used to obtain the second sub-result representing whether the first information is a rumor based on at least one of the fifth correlation and the sixth correlation.

[0548] In an optional implementation, it also includes: a fifth acquisition module, used to obtain the rumor publishing tendency of the user account that published the first information, and the rumor publishing tendency is used to characterize the probability that the first information published by the user account is a rumor; the corresponding third acquisition module is specifically used to: determine the final result characterizing whether the first information is a rumor based on the rumor publishing tendency, the first result and the second result.

[0549] In an optional implementation, it also includes: a sixth acquisition module, used to obtain multiple user-generated content corresponding to the first information; a seventh acquisition module, used to obtain the number of user-generated content representing the report that the first information is a rumor among the multiple user-generated content; the corresponding third acquisition module is specifically used to: determine the final result representing whether the first information is a rumor based on the number, the first result and the second result.

[0550] like Figure 8 As shown, it is a structural diagram of an electronic device provided in an embodiment of the present application. The electronic device may be a second server or an integrated device of a second server and a second electronic device. The electronic device includes:

[0551] Memory 801, used for storing programs;

[0552] The processor 802 is used to execute the program, and the program is specifically used to:

[0553] Acquire a title and text content corresponding to the first information to be identified, wherein the text content includes content representing the first information acquired directly or indirectly;

[0554] Based on the relationship between the title and at least one preset rumor title type, obtaining a first result indicating whether the first information belongs to a rumor;

[0555] Determining, based on the text content, a second result indicating whether the first information is a rumor;

[0556] Based on the first result and the second result, a final result is obtained, and the final result includes information characterizing whether the first information is a rumor.

[0557] The processor 802 may be a central processing unit (CPU) or an application specific integrated circuit (ASIC).

[0558] The control device may further include a communication interface 803 and a communication bus 804 , wherein the memory 801 , the processor 802 and the communication interface 803 communicate with each other via the communication bus 804 .

[0559] An embodiment of the present invention further provides a readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps included in any one of the above-mentioned rumor identification method embodiments applied to an electronic device are implemented.

[0560] The embodiment of the present invention also provides a computer program product or a computer program, which includes a computer instruction stored in a computer-readable storage medium. The processor of the computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the above-mentioned rumor identification method.

[0561] It should be noted that the features described in the various embodiments in this specification can be replaced or combined with each other. For the device or system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0562] It should also be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0563] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0564] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A rumor identification method, It is characterized in that include: Acquire a title and text content corresponding to the first information to be identified, wherein the text content includes content representing the first information acquired directly or indirectly; The first information is a video; If the title does not conform to at least one preset keyword pattern corresponding to the rumor title type, a first result indicating that the first information is not a rumor is obtained; a keyword pattern corresponding to the rumor title type is rule information composed of at least one keyword and used to describe a title that conforms to the rumor title type; If the title matches the keyword pattern corresponding to the target rumor title type in the at least one rumor title type, the title is input into the first classifier corresponding to the target rumor title type to obtain a first prediction result output by the first classifier; the first prediction result is that the title belongs to the target rumor title type, or the title does not belong to the target rumor title type; If the first prediction result is that the title belongs to the target rumor title type, determining the relationship between the title and at least one preset plot type; If the title belongs to a target plot type among the at least one plot type, obtaining a first result indicating that the first information is not a rumor; If the title does not belong to the at least one plot type, obtaining a first result indicating that the first information is a rumor; Determining, based on the text content, a second result indicating whether the first information is a rumor; Determining a final result indicating whether the first information is a rumor based on the first result and the second result includes: Determine whether the video is a carousel video based on the picture change feature of the video, wherein the picture change feature at least includes the dwell time corresponding to at least one picture contained in the video; determine whether the video is a machine dubbing video based on the attribute information of the audio contained in the video; Based on whether the video is a carousel video, whether the video is a machine-dubbed video, and the first result and the second result, a final result characterizing whether the first information is a rumor is determined.

2. According to the rumor identification method of claim 1, It is characterized in that The determining, based on the first result and the second result, a final result indicating whether the first information is a rumor further includes: Input the video into the third classifier corresponding to each non-rumor video type respectively, obtain the third prediction result output by the third classifier corresponding to each non-rumor video type respectively, the third prediction result output by the third classifier corresponding to one non-rumor video type is that the video belongs to the non-rumor video type or the video does not belong to the non-rumor video type; determine the relationship between the introduction content of the video and at least one preset rumor introduction type; as well as, Based on whether the video belongs to a carousel video, whether the video is a machine-dubbed video, the target prediction results output by the target classifiers corresponding to each of the non-rumor video types, and the relationship between the introduction content of the video and at least one preset rumor introduction type, as well as the first result and the second result, determine the final result characterizing whether the first information is a rumor.

3. According to the rumor identification method of claim 1, It is characterized in that Also includes: Acquire a plurality of user-generated contents corresponding to the first information; Obtaining the number of user-generated contents reporting that the first information is a rumor from among the plurality of user-generated contents; Determining the final result representing whether the first information is a rumor based on the first result and the second result includes: determining the final result representing whether the first information is a rumor based on the number, the first result, and the second result.

4. According to the rumor identification method of claim 1, It is characterized in that Also includes: Acquire a rumor publishing tendency of the user account that published the first information, where the rumor publishing tendency is used to represent the probability that the first information published by the user account is a rumor; Determining the final result representing whether the first information is a rumor based on the first result and the second result includes: determining the final result representing whether the first information is a rumor based on the tendency to publish rumors, the first result, and the second result.

5. According to the rumor identification method of claim 1, It is characterized in that Also includes: If the title matches the keyword pattern corresponding to the target rumor title type in the at least one rumor title type, obtaining at least one of the largest first correlation between the title and each pre-stored rumor document and the largest second correlation between the title and each non-rumor document; The determining, based on the first result and the second result, of a final result representing whether the first information is a rumor includes: The final result is obtained based on at least one of the first correlation and the second correlation, the first result, and the second result.

6. The rumor identification method according to any one of claims 1 to 4, It is characterized in that The text content includes first text content, the first information is a video, and obtaining the text content corresponding to the first information to be identified includes: Acquire multiple frames of video images from the video; Performing optical character recognition on multiple frames of the video image to obtain text information contained in at least one frame of the video image; The first text content is obtained based on text information contained in at least one frame of the video image.

7. According to the rumor identification method of claim 6, It is characterized in that The performing optical character recognition on the multiple frames of the video image to obtain text information contained in at least one frame of the video image includes: Acquire target text information corresponding to at least one set of video image sets, wherein one set of the video image sets includes at least two frames of the video image that are adjacent in time, and the target text information corresponding to one set of the video image sets is text information contained in at least two frames of the video image that are adjacent in time and included in the video image set; From the text information contained in at least one frame of the video image, a deduplication operation is performed on target text information corresponding to at least one group of the video image sets to obtain the first text content.

8. According to the rumor identification method of claim 7, It is characterized in that Determining the target text information corresponding to a set of the video image sets includes: Obtaining the size of the position area carrying text information respectively contained in at least two frames of the video images that are temporally adjacent to each other and included in the video image set; If the sizes of the target location areas in the location areas respectively included in at least two temporally adjacent frames of the video images are the same, detecting whether the text information in the target location areas respectively included in at least two temporally adjacent frames of the video images are the same; If the text information in the target position area respectively included in at least two temporally adjacent frames of the video image is the same, the text information in the target position area respectively included in at least two temporally adjacent frames of the video image is determined as the target text information.

9. According to the rumor identification method of claim 1, It is characterized in that The text content includes second text content, and obtaining the text content corresponding to the first information to be identified also includes: Perform speech recognition on the audio contained in the video to obtain the second text content.

10. The rumor identification method according to claim 9, It is characterized in that The second result includes at least one of the first sub-result and the second sub-result, and the second result of determining whether the first information is a rumor based on the text content includes: Obtaining at least one of the largest third correlation among the correlations between the first text content and each pre-stored rumor document, obtaining the largest fourth correlation among the correlations between the first text content and each pre-stored non-rumor document, obtaining the largest fifth correlation among the correlations between the second text content and each rumor document, and obtaining the largest sixth correlation among the correlations between the second text content and each non-rumor document; Based on at least one of the third relevance and the fourth relevance, obtaining the first sub-result representing whether the first information is a rumor; Based on at least one of the fifth relevance and the sixth relevance, the second sub-result representing whether the first information is a rumor is obtained.

11. A rumor identification device, It is characterized in that include: A first acquisition module, used to acquire a title and text content corresponding to the first information to be identified, wherein the text content includes content representing the first information acquired directly or indirectly; The first information is a video; A third acquisition unit is used to acquire a first result indicating that the first information is not a rumor if the title does not conform to a preset keyword pattern corresponding to at least one rumor title type; a keyword pattern corresponding to the rumor title type is rule information composed of at least one keyword and used to describe a title that conforms to the rumor title type; A first input unit is used to input the title into a first classifier corresponding to the target rumor title type if the title matches a keyword pattern corresponding to the target rumor title type in the at least one rumor title type, and obtain a first prediction result output by the first classifier; the first prediction result is that the title belongs to the target rumor title type, or the title does not belong to the target rumor title type; A first determining subunit is configured to determine a relationship between the title and at least one preset plot type if the first prediction result is that the title belongs to the target rumor title type; A first acquisition subunit is configured to acquire a first result indicating that the first information is not a rumor if the title belongs to a target plot type among the at least one plot type; A second acquisition subunit is configured to acquire a first result indicating that the first information is a rumor if the title does not belong to the at least one plot type; A determination module, configured to determine, based on the text content, a second result indicating whether the first information is a rumor; A third acquisition module, configured to determine a final result indicating whether the first information is a rumor based on the first result and the second result; The third acquisition module includes: A first determination unit, configured to determine whether the video is a carousel video based on a picture change feature of the video, wherein the picture change feature at least includes a dwell time corresponding to at least one picture included in the video; A first acquisition unit, configured to determine whether the video is a machine dubbing video based on attribute information of the audio contained in the video; The third determination unit is used to determine a final result representing whether the first information is a rumor based on whether the video is a carousel video, whether the video is a machine-dubbed video, and the first result and the second result.

12. The rumor identification device according to claim 11, It is characterized in that The third acquisition module further includes: A second determination unit is used to input the video into the third classifiers corresponding to the non-rumor video types respectively, and obtain the third prediction results output by the third classifiers corresponding to the non-rumor video types respectively, wherein the third prediction results output by the third classifiers corresponding to the non-rumor video types are that the video belongs to the non-rumor video type or that the video does not belong to the non-rumor video type; A second acquisition unit is used to determine the relationship between the video introduction content and at least one preset rumor introduction type; The third determination unit is used to determine the final result characterizing whether the first information is a rumor based on whether the video belongs to a carousel video, whether the video is a machine-dubbed video, the target prediction results output by the target classifiers corresponding to each of the non-rumor video types, and the relationship between the introduction content of the video and at least one preset rumor introduction type, as well as the first result and the second result.

13. The rumor identification device according to claim 11, It is characterized in that Also includes: A sixth acquisition module, used to acquire a plurality of user-generated contents corresponding to the first information; A seventh acquisition module, configured to acquire the number of user-generated contents reporting the first information as a rumor from among the plurality of user-generated contents; The third acquisition module is specifically used to determine a final result indicating whether the first information is a rumor based on the number, the first result and the second result.

14. The rumor identification device according to claim 11, It is characterized in that Also includes: A fifth acquisition module, used to acquire a rumor publishing tendency of the user account that published the first information, wherein the rumor publishing tendency is used to characterize the probability that the first information published by the user account is a rumor; The third acquisition module is specifically used to determine a final result representing whether the first information is a rumor based on the tendency to publish rumors, the first result and the second result.

15. The rumor identification device according to claim 11, It is characterized in that Also includes: A fourth acquisition module, for acquiring at least one of the largest first correlation between the title and each pre-stored rumor document and the largest second correlation between the title and each non-rumor document, if the title matches the keyword pattern corresponding to the target rumor title type in the at least one rumor title type; The third acquisition module is specifically used for: The final result is obtained based on at least one of the first correlation and the second correlation, the first result, and the second result.

16. The rumor identification device according to any one of claims 11 to 14, It is characterized in that The text content includes first text content, the first information is a video, and the first acquisition module includes: A fifth acquisition unit, used to acquire multiple frames of video images from the video; a sixth acquisition unit, configured to perform optical character recognition on the plurality of frames of the video image to obtain text information contained in at least one frame of the video image; The seventh acquisition unit is used to obtain the first text content based on the text information contained in at least one frame of the video image.

17. The rumor identification device according to claim 16, It is characterized in that The sixth acquisition unit includes: A third acquisition subunit is used to acquire target text information corresponding to at least one group of video image sets, wherein a group of the video image sets includes at least two frames of the video image that are adjacent in time, and the target text information corresponding to a group of the video image sets is text information contained in at least two frames of the video image that are adjacent in time and included in the video image set; The fourth acquisition subunit is used to perform a deduplication operation on target text information corresponding to at least one group of the video image sets from the text information contained in at least one frame of the video image, so as to obtain the first text content.

18. The rumor identification device according to claim 17, It is characterized in that The third acquisition subunit determines that the target text information corresponding to a set of the video image sets includes: A first acquisition submodule is used to acquire the size of the position area carrying text information respectively contained in at least two frames of the video image set that are temporally adjacent; A detection submodule, configured to detect whether text information in the target location areas respectively included in at least two temporally adjacent frames of the video images is the same if the sizes of the target location areas in the location areas respectively included in at least two temporally adjacent frames of the video images are the same; The determination submodule is used to determine the text information in the target position area respectively contained in at least two temporally adjacent frames of the video image as the target text information if the text information in the target position area respectively contained in at least two temporally adjacent frames of the video image is the same.

19. The rumor identification device according to claim 11, It is characterized in that The first acquisition module also includes: A speech recognition unit is used to perform speech recognition on the audio contained in the video to obtain the second text content.

20. The rumor identification device according to claim 19, It is characterized in that The determination module includes: at least one of an eighth acquisition unit, a ninth acquisition unit, a tenth acquisition unit, and an eleventh acquisition unit, and at least one of a twelfth acquisition unit and a thirteenth acquisition unit; The eighth acquisition unit is used to acquire the third largest correlation between the correlations between the first text content and each pre-stored rumor document; The ninth acquisition unit is used to acquire the largest fourth correlation between the correlations between the first text content and the pre-stored non-rumor documents; The tenth acquisition unit is used to acquire the fifth largest correlation among the correlations between the second text content and each of the rumor documents; The eleventh obtaining unit is used to obtain the sixth largest correlation among the correlations between the second text content and each of the non-rumor documents; The twelfth obtaining unit is configured to obtain, based on at least one of the third relevance and the fourth relevance, the first sub-result representing whether the first information is a rumor; The thirteenth acquisition unit is used to obtain the second sub-result representing whether the first information is a rumor based on at least one of the fifth correlation and the sixth correlation.

21. An electronic device, It is characterized in that include: Memory, used to store programs; A processor is used to execute the program, wherein the program is specifically used to execute the steps included in the rumor identification method described in any one of claims 1 to 10.

22. A storage medium, It is characterized in that A computer program is stored thereon, and when the computer program is executed by a processor, the steps included in the rumor identification method as described in any one of claims 1 to 10 are implemented.

23. A computer program product, It is characterized in that The computer program product includes computer instructions, which are stored in a computer-readable storage medium; a processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the steps included in the rumor identification method as described in any one of claims 1 to 10 above.