Video author identification method and apparatus, electronic device, and medium

Abstract

The present disclosure relates to a video author identification method and device, electronic equipment and medium, and relates to the technical field of computer. Author related information of a to-be-identified author and content related information of a published video created by the to-be-identified author are obtained; the author related information and the content related information are input into a target identification model to encode the author related information into a text vector sequence and encode the content related information into a video vector sequence; the target identification model is trained based on sample author related information of a sample author, sample content related information of a published video created by the sample author and a sample author label, and the sample author label is used to represent whether the sample author is a knowledge author; an author type classification identification is performed on the to-be-identified author based on the text vector sequence and the video vector sequence to obtain an author type identification result of the to-be-identified author; the author type identification result is used to represent whether the to-be-identified author is a knowledge author, and the identification accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computer, and particularly relates to a video author identification method and device, electronic equipment and medium. BACKGROUND

[0002] With the development of network technology, a large number of video platforms (for example, short video platforms) emerge. Platform users as creators can realize communication and sharing with other users through uploading or publishing videos. Accordingly, platform parties can recommend and encourage different types of videos according to actual needs of users.

[0003] Among them, the proportion of knowledge type video sharing in the platform is increasing, therefore, how to accurately identify whether the creator in the platform is a knowledge type author becomes a problem to be solved. SUMMARY

[0004] The present disclosure provides a video author identification method, device, electronic equipment and medium to at least solve the problem of how to accurately identify whether the creator in the platform is a knowledge type author. The technical solutions of the present disclosure are as follows:

[0005] According to a first aspect of the embodiments of the present disclosure, a video author identification method is provided, comprising:

[0006] obtaining author related information of a to-be-identified author and content related information of a published video created by the to-be-identified author;

[0007] inputting the author related information and the content related information into a target identification model to encode the author related information into a text vector sequence and encode the content related information into a video vector sequence; the target identification model is trained based on sample author related information of a sample author, sample content related information of a published video created by the sample author and a sample author label, and the sample author label is used to represent whether the sample author is a knowledge type author;

[0008] performing author type classification identification on the to-be-identified author based on the text vector sequence and the video vector sequence to obtain an author type identification result of the to-be-identified author; the author type identification result is used to represent whether the to-be-identified author is a knowledge type author.

[0009] Optionally, the number of published videos created by the to-be-identified author is at least two, and the content related information includes a video label and video multi-modal information; the video label is used to represent whether the published video is a knowledge type video, and the video label is determined based on the video content of the published video; the method further comprises: obtaining the video upload time of the published video created by the to-be-identified author.

[0010] encoding the content-related information into a sequence of video vectors, comprising:

[0011] encoding video tags of each of the published videos to obtain a first vector of each of the published videos;

[0012] classifying the published videos into a plurality of data sets corresponding to different time intervals based on video upload times of the published videos; the plurality of data sets belong to at least two different time intervals;

[0013] obtaining a confidence of a published video in each of the data sets, and generating a second vector based on the confidence of each of the published videos;

[0014] generating a third vector for each of the published videos based on video multi-modal information of each of the published videos;

[0015] generating the sequence of video vectors based on the first vector, the second vector and the third vector of each of the published videos.

[0016] Optionally, the classifying the published videos into a plurality of data sets corresponding to different time intervals based on video upload times of the published videos comprises:

[0017] for any of the time intervals, determining a published video belonging to the time interval from each of the published videos;

[0018] determining a set duration corresponding to a data set belonging to the time interval based on a time span of the time interval and a number of data sets corresponding to the time interval; the set duration is used to represent an upper limit value of a difference between video upload times of videos in the data set;

[0019] determining a time range corresponding to each of the data sets belonging to the time interval based on the set duration and the time span of the time interval;

[0020] for any of the published videos belonging to the time interval, classifying the published video into a data set corresponding to a time range to which a video upload time of the published video belongs based on the video upload time of the published video.

[0021] Optionally, the generating the second vector based on the confidence of each of the published videos comprises:

[0022] for any of the time intervals, generating a first sub-vector for each of the data sets belonging to the time interval based on a confidence of a published video in each of the data sets belonging to the time interval;

[0023] generating a second sub-vector corresponding to the time interval based on a first sub-vector of each of the data sets, and splicing second sub-vectors corresponding to all time intervals;

[0024] generating a second vector for each of the published videos based on the spliced second sub-vectors.

[0025] Optionally, the generating the video vector sequence based on the first vector, the second vector and the third vector of each of the published videos comprises:

[0026] splicing the first vector, the second vector and the third vector of each of the published videos to obtain a video vector of each of the published videos.

[0027] For any of the published videos, processing the video vector of the published video based on a first encoding layer in the target recognition model to generate a target confidence of the published video.

[0028] splicing the video vectors of the published videos with a target confidence not less than a preset confidence threshold to generate the video vector sequence.

[0029] Optionally, the author-related information comprises attribute information of different dimensions of the to-be-recognized author; and the encoding the author-related information into a text vector sequence comprises:

[0030] splicing text included in the attribute information of different dimensions based on a second encoding layer in the target recognition model to obtain a target text.

[0031] encoding each character in the target text to obtain a character vector of each of the characters.

[0032] splicing the character vectors of each of the characters to generate the text vector sequence.

[0033] Optionally, the generating a first sub-vector for each data set belonging to the time interval based on the confidence of the published videos in the data set comprises:

[0034] For any data set belonging to the time interval, sorting the published videos in the data set according to a preset sorting index to obtain a sorting result.

[0035] sorting the confidence of the published videos in the data set based on the sorting result, and generating the first sub-vector of the data set based on the sorted confidence.

[0036] Optionally, the target recognition model is obtained by training in the following manner:

[0037] obtaining sample author related information of the sample author and sample content related information of published videos created by the sample author;

[0038] inputting the sample author related information and the sample content related information into the author type identification model to be trained, encoding the sample author related information into a sample text vector sequence, encoding the sample content related information into a sample video vector sequence, and performing author type classification identification on the sample author based on the sample text vector sequence and the sample video vector sequence to obtain a sample identification result of the sample author;

[0039] adjusting the author type identification model to be trained based on the sample identification result and a sample author label of the sample author;

[0040] in a case where a training process of the author type identification model to be trained reaches a stop condition, determining the author type identification model reaching the stop condition as the target identification model.

[0041] According to a second aspect of the embodiments of the present disclosure, a video author identification apparatus is provided, which comprises:

[0042] an information obtaining module configured to perform obtaining author related information of an author to be identified and content related information of published videos created by the author to be identified;

[0043] an information inputting module configured to perform inputting the author related information and the content related information into a target identification model to encode the author related information into a text vector sequence and encode the content related information into a video vector sequence; the target identification model is trained based on sample author related information of a sample author, sample content related information of published videos created by the sample author, and a sample author label; the sample author label is used to represent whether the sample author is a knowledge type author;

[0044] a type identification module configured to perform author type classification identification on the author to be identified based on the text vector sequence and the video vector sequence to obtain an author type identification result of the author to be identified; the author type identification result is used to represent whether the author to be identified is a knowledge type author.

[0045] Optionally, the number of published videos created by the author to be identified is at least two, the content related information comprises a video label and video multi-modal information; the video label is used to represent whether the published video is a knowledge type video, and the video label is determined based on video content of the published video; the information inputting module comprises:

[0046] a first vector generation submodule configured to perform encoding processing on a video tag of each of the published videos to obtain a first vector of each of the published videos;

[0047] a classification submodule configured to perform classification of the published videos into a plurality of data sets corresponding to different time intervals based on a video upload time of each of the published videos; the plurality of data sets belong to at least two different time intervals;

[0048] a second vector generation submodule configured to perform obtaining of a confidence of a published video in each of the data sets and generating a second vector based on the confidence of each of the published videos;

[0049] a third vector generation submodule configured to perform generating a third vector for each of the published videos based on video multi-modal information of each of the published videos;

[0050] a vector sequence generation submodule configured to perform generating the video vector sequence based on the first vector, the second vector and the third vector of each of the published videos.

[0051] Optionally, the classification submodule is specifically configured to perform:

[0052] for any of the time intervals, determining a published video belonging to the time interval from each of the published videos;

[0053] determining a set duration corresponding to a data set belonging to the time interval based on a time span corresponding to the time interval and a number of data sets corresponding to the time interval; the set duration is used to represent an upper limit value of a difference between video upload times of videos in the data set;

[0054] determining a time range corresponding to each data set belonging to the time interval based on the set duration and the time span of the time interval;

[0055] for any of the published videos belonging to the time interval, classifying the published video into a data set corresponding to a time range to which a video upload time of the published video belongs based on the video upload time of the published video.

[0056] Optionally, the second vector generation submodule is specifically configured to perform:

[0057] for any of the time intervals, generating a first sub-vector for each data set belonging to the time interval based on a confidence of a published video in each data set belonging to the time interval;

[0058] generate a second sub-vector corresponding to the time interval based on a first sub-vector of each of the data sets, and splice second sub-vectors corresponding to all time intervals;

[0059] generate a second vector for each of the published videos based on the spliced second sub-vectors.

[0060] Optionally, the vector sequence generation submodule is specifically configured to perform:

[0061] splice the first vector, the second vector and the third vector of each of the published videos to obtain a video vector of each of the published videos;

[0062] For any of the published videos, process the video vector of the published video based on a first encoding layer in the target recognition model to generate a target confidence of the published video.

[0063] splice the video vectors of the published videos with a target confidence not less than a preset confidence threshold to generate the video vector sequence.

[0064] Optionally, the author-related information includes attribute information of different dimensions of the to-be-identified author; and the information input module includes:

[0065] a text splicing submodule configured to perform splicing of text included in the attribute information of different dimensions based on a second encoding layer in the target recognition model to obtain target text;

[0066] a character encoding submodule configured to perform encoding of each character in the target text to obtain a character vector of each of the characters;

[0067] a text vector generation submodule configured to perform splicing of the character vectors of each of the characters to generate the text vector sequence.

[0068] Optionally, the second vector generation submodule is specifically configured to perform:

[0069] For any data set belonging to the time interval, sort the published videos in the data set according to a preset sorting index to obtain a sorting result;

[0070] sort the confidences of the published videos in the data set based on the sorting result, and generate a first sub-vector of the data set based on the sorted confidences.

[0071] Optionally, the target recognition model is obtained by training in the following manner:

[0072] obtaining sample author related information of the sample author and sample content related information of published videos created by the sample author;

[0073] inputting the sample author related information and the sample content related information into an author type identification model to be trained, to encode the sample author related information into a sample text vector sequence, encode the sample content related information into a sample video vector sequence, and perform author type classification identification on the sample author based on the sample text vector sequence and the sample video vector sequence, to obtain a sample identification result of the sample author;

[0074] adjusting the author type identification model to be trained based on the sample identification result and a sample author label of the sample author;

[0075] in a case where a training process of the author type identification model to be trained reaches a stop condition, determining the author type identification model reaching the stop condition as the target identification model.

[0076] According to a third aspect of the embodiments of the present disclosure, an electronic device is provided, comprising:

[0077] a processor;

[0078] a memory for storing instructions executable by the processor;

[0079] The processor is configured to execute the instructions to implement the method according to any one of the first aspect.

[0080] According to a fourth aspect of the embodiments of the present disclosure, a storage medium is provided, when instructions in the storage medium are executed by a processor of an electronic device, the electronic device performs the method according to any one of the first aspect.

[0081] According to a fifth aspect of the embodiments of the present disclosure, a computer program product is provided, the computer program product comprises readable program instructions, when the readable program instructions are executed by a processor of an electronic device, the electronic device performs the method according to any one of the first aspect.

[0082] The embodiments of the present disclosure provide at least the following beneficial effects: in the embodiments of the present disclosure, author related information of a to-be-identified author and content related information of a published video created by the to-be-identified author are acquired; the author related information and the content related information are input into a target identification model to encode the author related information into a text vector sequence and encode the content related information into a video vector sequence; the target identification model is trained based on sample author related information of a sample author, sample content related information of a published video created by the sample author and a sample author label, the sample author label being used to represent whether the sample author is a knowledge type author; an author type classification identification is performed on the to-be-identified author based on the text vector sequence and the video vector sequence to obtain an author type identification result of the to-be-identified author; the author type identification result is used to represent whether the to-be-identified author is a knowledge type author. In this way, since the target identification model is trained based on the sample author related information of the sample author, the sample content related information of the published video created by the sample author and the sample author label, and the sample author label can represent whether the sample author is a knowledge type author, the target identification model can identify and classify the to-be-identified author according to multi-aspect and multi-dimensional data by encoding and identifying the author related information and the content related information of the to-be-identified author, thereby improving the accuracy of video author identification to some extent and enabling the to-be-identified author to be accurately identified as a knowledge type author.

[0083] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and are not limiting to the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0084] The accompanying drawings incorporated in and forming a part of the specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the disclosure without imposing on the disclosure any undue limitations.

[0085] Figure 1 is a flowchart of a video author identification method according to an exemplary embodiment;

[0086] Figure 2 is an identification flowchart of a target identification model according to an exemplary embodiment of the present disclosure;

[0087] Figure 3 is a schematic diagram of a target identification model according to an exemplary embodiment of the present disclosure;

[0088] Figure 4 is a block diagram of a video author identification device according to an exemplary embodiment;

[0089] Figure 5 is a block diagram of an apparatus for video author identification according to an example embodiment;

[0090] Figure 6 is a block diagram of another apparatus for video author identification according to an example embodiment. DETAILED DESCRIPTION

[0091] In order to make the ordinary person skilled in the art better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be described clearly and completely below in conjunction with the drawings.

[0092] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The implementation described in the following example embodiments does not represent all implementations consistent with the present disclosure. Rather, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0093] Figure 1 is a flowchart of a video author identification method according to an example embodiment, as shown in Figure 1 the method can include the following steps:

[0094] Step 101, obtaining author-related information of an author to be identified and content-related information of a published video created by the author to be identified.

[0095] Among them, the above-identified author refers to the publisher of the video, and the method can be applied to any video platform, so that the above-identified author can be selected from all publishers on the platform, specifically, it can be any publisher in all publishers of the video platform who have not been identified as an author type. Further, the above-mentioned author-related information refers to the attribute information related to the author to be identified, which can include author name, author profile, number of fans, and number of published videos, and different dimensions of information.

[0096] Correspondingly, the published videos can be all the videos published by the to-be-identified author, or a specified number of videos selected from all the videos published by the to-be-identified author. The content-related information can be information related to the content of the video, and can include the content of the video itself, a video category, and the like. The video category can be used to represent the classification of the published video. For example, when a published video is a video clip, the corresponding video category can be entertainment or film and television; when a published video is a daily life record, the video category can be life or daily; when a published video is a science popularization content or a skill sharing in a certain field, the video category can be knowledge. Specifically, the knowledge video refers to a video that enables the viewer to obtain knowledge or skills, and is useful to the viewer. The video is usually in the form of science popularization, and can be knowledge popularization in different fields.

[0097] Specifically, the content-related information can be obtained from the identification of the video and the video frames contained therein. Specifically, the content of the video itself can be obtained by extracting a preset number of key frames (video frames containing characters or text) from the video frames contained in the video. Further, the video category can be obtained by a preset video category recognition model, and according to the audio data in the video and the content of the key frames or all video frames. Alternatively, the video category of the published video can be obtained based on the identification.

[0098] Specifically, the author-related information and the content-related information can be obtained in advance.

[0099] Step 102: inputting the author-related information and the content-related information into a target recognition model to encode the author-related information into a text vector sequence and the content-related information into a video vector sequence; the target recognition model is trained based on sample author-related information of a sample author, sample content-related information of published videos created by the sample author, and a sample author label used to represent whether the sample author is a knowledge author.

[0100] Step 103: performing author type classification and identification on the to-be-identified author based on the text vector sequence and the video vector sequence to obtain an author type identification result of the to-be-identified author; the author type identification result is used to represent whether the to-be-identified author is a knowledge author.

[0101] The target recognition model can be pre-trained.

[0102] The text vector sequence can be obtained by encoding the author-related information. The author-related information is usually in the form of text or characters, and thus the encoding algorithm can be a preset character encoding algorithm, for example, American Standard Code for Information Interchange (ASCII) encoding, Unicode encoding, or the like. When the author-related information includes information of different dimensions, the information of different dimensions can be separated, for example, by using a separator to obtain a separated string or text string, and then the corresponding text vector sequence can be obtained by encoding the string or text string.

[0103] The video vector sequence can be obtained by encoding the content-related information. The video vector sequence can be obtained by linear mapping.

[0104] Further, after obtaining the text vector sequence and the video vector sequence, the two can be spliced to obtain a spliced vector. Since the text vector sequence and the video vector sequence respectively include relevant information of different dimensions, the target recognition model can perform multi-dimensional classification and recognition on the to-be-recognized author by using the spliced vector to obtain an author type recognition result.

[0105] Specifically, the target recognition model can include an encoding layer and a classification layer. The encoding of the author-related information and the content-related information can be performed by the encoding layer. After splicing the text vector sequence and the video vector sequence output by the encoding layer to obtain a spliced vector, the spliced vector can be input into the classification layer, and the output of the classification layer can be determined as the author type recognition result, which can represent the type of the to-be-recognized author. Specifically, the author type recognition result can represent whether the to-be-recognized author is a knowledge author. The result can be 1 / 0 or Y / N. Specifically, when the recognition result is 1 or Y, the to-be-recognized author can be considered as a knowledge author, and when the recognition result is 0 or N, the to-be-recognized author can be considered as a non-knowledge author. Of course, the recognition result can also be in the form of text, for example, knowledge and non-knowledge. The present disclosure does not limit the expression form of the recognition result.

[0106] Further, any publisher or author of a video platform can be identified by the target identification model, obtaining the knowledge author of the platform, so that the video published by the knowledge author can be further processed, for example, the video of the knowledge author can be increased in push traffic or push intensity, etc.

[0107] To sum up, the video author identification method provided by the embodiments of the present disclosure, by obtaining author-related information of the to-be-identified author and content-related information of the published video created by the to-be-identified author; inputting the author-related information and the content-related information into a target identification model to encode the author-related information into a text vector sequence and the content-related information into a video vector sequence; the target identification model is trained based on sample author-related information of a sample author, sample content-related information of a published video created by the sample author, and a sample author label, the sample author label is used to represent whether the sample author is a knowledge author; based on the text vector sequence and the video vector sequence, the author type of the to-be-identified author is classified and identified to obtain the author type identification result of the to-be-identified author; the author type identification result is used to represent whether the to-be-identified author is a knowledge author. In this way, since the target identification model is trained based on the sample author-related information of the sample author, the sample content-related information of the published video created by the sample author, and the sample author label, and the sample author label can represent whether the sample author is a knowledge author, the target identification model can identify and classify the to-be-identified author according to multi-aspect and multi-dimensional data, which can improve the accuracy of video author identification to some extent, so that whether the to-be-identified author is a knowledge author can be accurately identified.

[0108] In an optional embodiment, the number of published videos of the to-be-identified author is at least two, and the content-related information includes video tags and video multi-modal information; the video tags are used to represent whether the published video is a knowledge video, and the video tags are determined based on the video content of the published video; the embodiments of the present disclosure can further include the following steps:

[0109] Step 201, obtaining the video upload time of the published video created by the to-be-identified author.

[0110] Among them, the video upload time refers to the publishing time of the video, that is, the time when the video is uploaded to the video platform. For example, when the video publishing time is January 1, 2022 20:00, the upload time is January 1, 2022 20:00. The upload time of each video can be retrieved from the video library of the video platform according to the identifier of each video.

[0111] The operation of encoding the content-related information into a sequence of video vectors, the embodiments of the present disclosure can specifically include the following steps:

[0112] Step 202, encode the video tags of each published video to obtain a first vector of each published video.

[0113] The video tag is used to represent the category of the video, that is, whether it is a knowledge video. Specifically, the video tag can be obtained by detecting and analyzing all video frames contained in the video through a preset video recognition model, and can be represented by 1 and 0, where 1 represents knowledge and 0 represents non-knowledge.

[0114] In the embodiments of the present disclosure, the encoding process can be linear mapping of the video tag. Specifically, linear mapping can be performed through the encoding layer or linear layer in the model to encode the video tag into a d-dimensional vector v1, that is, the first vector. Wherein, the above-mentioned d can be set according to actual needs (for example, 20, 10 or 30, etc.).

[0115] Step 203, based on the upload time of each published video, classify the published video into a plurality of data sets corresponding to different time intervals; the plurality of data sets belong to at least two different time intervals.

[0116] The time range corresponding to different data sets can be determined according to the time span of the time interval to which each data set belongs.

[0117] Specifically, the above-mentioned data set can be understood as grouping different videos, and each video can be allocated to a time interval with different time spans according to the different upload time of the published video. Further, one or more data sets can be set for any time interval, so that the videos in the interval are further divided into data sets according to the time range of each set. Specifically, the number of time intervals L and the number of sets L i (i∈[0,L-1]) can be set by themselves.

[0118] Exemplarily, two time intervals (L0 and L1) can be set, and the time spans of the two are one year and half a year respectively. When the span of the publishing time of a video from the current time is not greater than one year, the video is classified into L0. Correspondingly, when the span of the publishing time of the video from the current time is not greater than half a year, the video is classified into L1. It can be understood that under this division mode, the video classified into L1 will also be classified into L0. Further, assuming that two sets (M0 and M1) are set for the above L1, the time ranges of the two can be the first three months and the last three months respectively. The video in L1 whose publishing time from the current time is not greater than three months can be classified into M0, and the video in L1 whose publishing time from the current time is greater than three months can be classified into M1. Alternatively, taking the above L0 and L1 as an example, the video whose publishing time from the current time is greater than half a year and less than one year can be classified into L0, so as to avoid the repetition of video data in different time intervals.

[0119] Step 204, obtaining the confidence of the published video in each data set, and generating a second vector based on the confidence of each published video.

[0120] The above confidence refers to the confidence level, which can be understood as the probability that the published video is a knowledge video. Specifically, the confidence of the published video can be obtained through a preset confidence evaluation model or a confidence algorithm, which is not limited by the embodiments of the present disclosure. Specifically, the confidence of the published video in each data set can be linearly processed through a linear layer to map it into a d-dimensional vector v2, that is, a second vector, where the above d can be the same as the d of the first vector, and of course, it can also be different.

[0121] Step 205, generating a third vector for each published video based on the video multi-modal information of each published video.

[0122] The above video multi-modal information can be obtained through a preset multi-modal algorithm (for example, VIT) or multi-modal calculation model. The multi-modal information refers to the description information of the same published video in different angles and different types of media (for example, text, image, sound, video, etc.), which is more comprehensive and comprehensive than single-modal information, so that the multi-modal information can more accurately describe the video content of the published video.

[0123] Specifically, the multi-modal information of the published video can be obtained by performing multi-modal calculation on each video frame of the published video. Further, a d-dimensional vector, that is, a third vector v3, can be obtained after linear mapping of the multi-modal information. The above d can be the same as the d of the first vector and the second vector, and of course, it can also be different.

[0124] In step 206, the video vector sequence is generated based on the first vector, the second vector and the third vector of each published video.

[0125] Specifically, the video vector Vi of each published video can be obtained by adding the first vector, the second vector and the third vector, i.e. Vi = v1 + v2 + v3, where i represents different published videos. The video vector sequence can be obtained by splicing the video vectors of different published videos. It can be understood that when the knowledge videos and non-knowledge videos of the same author are mixed, the number and frequency of the knowledge videos published by the author can be determined through the video tags and the upload time. Since the video vector sequence includes the first vector corresponding to the video tags, the second vector corresponding to the upload time and the third vector corresponding to the multi-modal information of the video, the subsequent processing and calculation of the video vector sequence can comprehensively evaluate the author by considering the number and frequency of the knowledge videos published by the author, thereby avoiding the recognition error caused by the mixing of the knowledge videos and non-knowledge videos of the same author. Meanwhile, the third vector generated according to the multi-modal information can make the video vector more accurately represent the video content of the published video, thereby further improving the accuracy of the subsequent processing of the video vector.

[0126] In the embodiments of the present disclosure, the number of published videos created by the to-be-identified author is at least two, and the content-related information includes video tags and video multi-modal information; the video tags are used to represent whether the published videos are knowledge videos, and the video tags are determined based on the video content of the published videos; the video upload time of the published videos created by the to-be-identified author is obtained; the video tags of each of the published videos are encoded to obtain a first vector of each of the published videos; based on the video upload time of each of the published videos, the published videos are classified into a plurality of data sets corresponding to different time intervals; the plurality of data sets belong to at least two different time intervals; the confidence of the published videos in each of the data sets is obtained, and a second vector is generated based on the confidence of each of the published videos; a third vector is generated for each of the published videos based on the video multi-modal information of each of the published videos; and the video vector sequence is generated based on the first vector, the second vector and the third vector of each of the published videos. In this way, by dividing the videos into different data sets according to the upload time of the published videos, parallel processing of each set can be facilitated, thereby improving the efficiency of generating the second vector and the video vector sequence, and further improving the data processing efficiency of the target identification model. At the same time, the number and frequency of knowledge videos published by the author can be determined through the video tags and the upload time. Since the video vector sequence includes the first vector corresponding to the video tags, the second vector corresponding to the upload time, and the third vector corresponding to the video multi-modal information, the number and frequency of knowledge videos published by the author can be evaluated when the video vector sequence is processed and calculated subsequently, thereby avoiding recognition errors caused by the mixing of knowledge videos and non-knowledge videos of the same author. In addition, the third vector generated based on the multi-modal information can make the video vector more accurately represent the video content of the published videos, thereby further improving the accuracy of video author identification and enabling accurate identification of whether the to-be-identified author is a knowledge author.

[0127] In an optional embodiment, the operation of classifying the published videos into a plurality of data sets corresponding to different time intervals based on the video upload time of each of the published videos can include the following steps:

[0128] Step 301: For any of the time intervals, determine the published videos belonging to the time interval from each of the published videos.

[0129] Step 302: Determine the set duration corresponding to the data set belonging to the time interval based on the time span corresponding to the time interval and the number of data sets corresponding to the time interval; the set duration is used to represent the upper limit value of the difference between the video upload times of the videos in the data set.

[0130] Step 303, determining the time range corresponding to each data set belonging to the time interval based on the time span of the time interval and the set duration of the data set.

[0131] Step 304, for any published video belonging to the time interval, classifying the published video into the data set corresponding to the time range to which the video upload time of the published video belongs based on the video upload time of the published video.

[0132] Specifically, the number of time intervals, the time span of each time interval, and the number of data sets contained in each time interval can be set in advance. Then, for any time interval, the set duration of each data set in the time interval, which can also be referred to as the scale of the data set, can be obtained by dividing the time span of the time interval by the number of data sets in the time interval. Alternatively, the scale of each data set in the time interval can also be obtained by dividing the time difference between the earliest upload time in the time interval and the latest upload time in the time interval by the number of data sets in the time interval. The earliest upload time in the time interval refers to the upload time of the published video with the earliest video upload time in the time interval, and the latest upload time in the time interval refers to the upload time of the published video with the latest video upload time in the time interval. The upload time can be in units of days / hours.

[0133] The collection duration can represent an upper limit of a difference between the earliest upload time and the latest upload time of the videos in the data collection, and the difference between the upload time of any video in the data collection and the upload time of other videos in the data collection is less than the upper limit. The time range of each data collection can be obtained according to the time span of the time interval and the collection duration. For example, the time span of the time interval L1 in step 203 is half a year, and two data collections (M0 and M1) are included. The scale of each data collection in L1 can be obtained as half a year / 2, that is, three months. That is, the difference between the upload time of any video in M0 and M1 and the upload time of other videos in the data collection is less than three months. Therefore, the time range of each data collection can be the first three months and the last three months in half a year, respectively. If the current time is December 31, 2021, the time range corresponding to M0 is from September 30, 2021 to December 31, 2021. Correspondingly, the time range corresponding to M1 is from June 30, 2021 to September 30, 2021. Further, for any time interval, the videos with upload times belonging to the time span of the time interval can be obtained according to the time span of the time interval, and the videos are divided into the time interval. Further, in any time interval, according to the time range of any data collection included in the time interval, the videos with upload times belonging to the time range of the data collection are divided into the data collection from the published videos belonging to the time interval.

[0134] For example, the number of time intervals is 3, the time span of the first time interval is 1 year, the number of data collections is 256, the time span of the second time interval is half a year, the number of data collections is 128, and the time span of the third time interval is one month, and the number of data collections is 64. The scale of each data collection in the first time interval can be obtained by rounding up to 35 hours, the scale of each data collection in the second time interval can be obtained by rounding up to 35 hours, and the scale of each data collection in the third time interval can be obtained by rounding up to 12 hours. Therefore, when the upload time of a published video is not more than 1 year from the current time, the video is divided into the first time interval. Further, when the upload time of the video is more than 364 days from the current time and less than 365 days, the video can be divided into the first data collection in the first time interval.

[0135] In the embodiments of the present disclosure, for any time interval, the published videos belonging to the time interval are determined from each of the published videos; based on the time span corresponding to the time interval and the number of data sets corresponding to the time interval, the set duration corresponding to the data set belonging to the time interval is determined; the set duration is used to represent the upper limit value of the difference between the video upload times of the videos in the data set; based on the set duration and the time span of the time interval, the time range corresponding to each data set belonging to the time interval is determined; for any published video belonging to the time interval, based on the video upload time of the published video, the published video is classified into the data set corresponding to the time range to which the video upload time of the published video belongs. In this way, by dividing the published videos into different data sets contained in different time intervals according to the upload time, the video data of different data sets can be conveniently processed in parallel, improving the processing efficiency, and at the same time, by dividing different intervals and sets according to the upload time, the recognition error caused by different author creation frequencies can be avoided.

[0136] In an optional embodiment, the operation of generating the second vector based on the confidence of each published video can specifically include the following steps:

[0137] Step 401: For any time interval, based on the confidence of the published videos in each data set belonging to the time interval, a first sub-vector corresponding to each data set belonging to the time interval is generated.

[0138] Specifically, for any data set in any time interval, the confidence of the video with the knowledge class corresponding to the video tag in the data set can be taken as the confidence corresponding to the data set, so that the confidence corresponding to each data set can be taken as the first sub-vector corresponding to each data set.

[0139] Step 402: The second sub-vector corresponding to the time interval is generated based on the first sub-vector of each data set, and the second sub-vectors corresponding to all time intervals are spliced.

[0140] Specifically, for any time interval, the data sets in the time interval can be identified or numbered (index), for example, the first time interval can be {0, 1, 2, … 255}, and the confidence of the video with the knowledge class corresponding to the video tag is taken as the value corresponding to each index, to obtain the second sub-vector of the time interval (for example, the second sub-vector of the first time interval can be [v0, v1, …, v 255 ])).

[0141] Further, after obtaining the second sub-vectors corresponding to all time intervals, the second sub-vectors can be spliced according to a preset splicing order (for example, time span from far to near or from near to far) to obtain a spliced second sub-vector.

[0142] Step 403, generating a second vector for each of the published videos based on the spliced second sub-vectors.

[0143] Specifically, after obtaining the spliced second sub-vectors, the spliced second sub-vectors can be mapped to a d-dimensional vector v2, i.e., a second vector, through a linear layer in the target recognition model. It can be understood that, since the second vector is obtained by splicing the second sub-vectors of all intervals, and all time intervals contain all published videos, the second vectors of each published video of the same to-be-recognized author obtained through the above steps are all the same.

[0144] In the embodiments of the present disclosure, for any time interval, a first sub-vector is generated for each data set belonging to the time interval based on the confidence of the published videos in the data set; a second sub-vector corresponding to the time interval is generated based on the first sub-vectors of each data set, and all second sub-vectors corresponding to time intervals are spliced; and a second vector is generated for each of the published videos based on the spliced second sub-vectors. In this way, by obtaining second sub-vectors of different time intervals and splicing all second sub-vectors to generate a second vector, the second vector can cover videos published by an author in different time periods, and the second vector can be used to comprehensively evaluate the type of the to-be-recognized author in different time periods, thereby improving the accuracy of video author recognition.

[0145] In an optional embodiment, the operation of generating the video vector sequence based on the first vector, the second vector, and the third vector of each of the published videos can specifically include the following steps:

[0146] Step 501, obtaining a video vector of each of the published videos based on the first vector, the second vector, and the third vector of each of the published videos.

[0147] Specifically, the video vector V i may be obtained by adding the first vector, the second vector, and the third vector. i , where i represents different published videos.

[0148] Step 502, for any of the published videos, processing the video vector of the published video based on the first encoding layer in the target recognition model to generate a target confidence of the published video.

[0149] Specifically, the first encoding layer can be a sub-model in an actual scene, which can include a transformation layer (Transformer), a linear layer and a normalization function layer (sigmoid), and the target confidence of the published video can be obtained according to the video vector through the sub-model, that is, the probability that the published video is a knowledge video.

[0150] In step 503, the video vectors of the published videos with a target confidence not less than a preset confidence threshold are spliced to generate the video vector sequence.

[0151] The preset confidence threshold can be set according to actual needs, for example, it can be set to 0.3, and of course it can also be set to 0.25 or 0.5, and the present disclosure does not limit this.

[0152] Specifically, from all the published videos, the video vectors V i of K published videos with a target confidence not less than a preset confidence threshold can be selected. i After splicing the K V K , the video vector sequence V = {V1, V2, …, V

[0153] Alternatively, an initial vector sequence can also be generated for all published videos. For example, the initial vector sequence can be represented as V' = {V1', V2', …, V N ' when the number of published videos is N. Then, the V i ' corresponding to the target confidence less than the preset confidence threshold in the initial vector sequence is deleted, and then the filtered video vector sequence V = {V1, V2, …, V K} is obtained. It can be understood that K is less than N.

[0154] In the present disclosure, the video vector of each published video is obtained based on the first vector, the second vector and the third vector of each published video. For any published video, the video vector of the published video is processed based on the first encoding layer in the target recognition model to generate the target confidence of the published video. The video vectors of the published videos with a target confidence not less than a preset confidence threshold are spliced to generate the video vector sequence. In this way, all video vectors are filtered according to the preset confidence threshold, which can reduce unnecessary data, avoid data processing loss, and make the final video vector sequence more referential in the model optimization process.

[0155] In an optional embodiment, the author-related information includes attribute information of different dimensions of the to-be-identified author; and the operation of encoding the author-related information into a text vector sequence can specifically include the following steps:

[0156] In step 601, the text included in the attribute information of different dimensions is spliced based on a second encoding layer in the target identification model, to obtain a target text.

[0157] The attribute information of different dimensions refers to author information of different angles and different types, which can include an author name, an author introduction, a number of fans, a number of published videos, and the like. Specifically, the target text can be obtained by splicing the attribute information of different dimensions by using a preset separator. For example, if the preset separator is " / ", the target text S = author name / author introduction / ... / number of fans / number of published videos.

[0158] The second encoding layer can also correspond to a sub-model in an actual scenario.

[0159] In step 602, each character in the target text is encoded to obtain a character vector of each character.

[0160] It can be understood that the target text S obtained by step 601 is usually in the form of a string, and therefore each character included in the string can be encoded to obtain a character vector corresponding to each character. i Specifically, the character can be encoded to obtain a character vector by using a pre-training model (Bidirectional Encoder Representations from Transformer, BERT) or a deep / bidirectional recurrent neural network (Deep / Bidirectional RNN), a long short-term memory network (Long short-term memory, LSTM), a convolutional neural network (Convolutional Neural Network, CNN), or the like.

[0161] In step 603, the character vectors of each character are spliced to generate the text vector sequence.

[0162] Specifically, after obtaining the character vectors w i of all characters, the character vectors can be spliced to obtain a text vector sequence W = {w1, w2,..., w p}, where p is the number of characters included in the target text.

[0163] In the embodiments of the present disclosure, the author-related information includes attribute information of different dimensions of the to-be-identified author; the text included in the attribute information of different dimensions is spliced based on a second encoding layer in the target identification model to obtain a target text; each character in the target text is encoded to obtain a character vector of each character; and the character vectors of each character are spliced to generate the text vector sequence. In this way, the text vector sequence is generated based on the author-related information including different dimension information, so that the text vector sequence can represent author information of different dimensions, and thus the author type identification result generated based on the text vector sequence can be more accurate, further improving the identification accuracy of video authors.

[0164] Figure 2 is a schematic diagram of an identification process of a target identification model according to an example embodiment of the present disclosure, as shown in Figure 2 The identification process can include obtaining author basic information, obtaining all videos and video knowledge tags of an author, model automatic sampling, and model feature encoding and classification.

[0165] The author basic information is author-related information, which can include the author's name, biography, number of fans, and the like. All videos and video knowledge tags of the author refer to all video content published by the author and the video tags (whether knowledge or not) corresponding to each video. It can be understood that Figure 2 The knowledge and non-knowledge blocks in the above-mentioned knowledge, non-knowledge correspond to different videos. The model automatic sampling refers to encoding the obtained all videos and video knowledge tags of the author to generate a first vector, a second vector, and a third vector, adding them together, and splicing the video vectors of different videos to obtain a video vector sequence and enter the next process of model feature encoding and classification. At the same time, the model feature encoding and classification is used for encoding the obtained author basic information to generate a text vector sequence, which is concatenated with the above-mentioned video vector sequence, for example, spliced to obtain a spliced vector, so that the author can be identified and classified according to the spliced vector, and the classification result is Y (knowledge author) or N (non-knowledge author).

[0166] It should be noted that in addition to invited knowledge authors and institutions in the video platform, there are often a large number of users who spontaneously organize and share knowledge, including sharing some life aspects such as small tips and the like. These authors usually have extremely dispersed distribution of publishing frequency, and the content, form, and quality of the published content are highly diverse, making it difficult to identify knowledge-sharing authors from them. The target identification model of the embodiments of the present disclosure can Figure 2The flow shown can automatically identify knowledge type authors in a large number of authors, facilitate better subsequent incentives for the creators. Meanwhile, whether or not to be a knowledge type author can be used as a signal to help recommendation and search technology better understand user needs, for example, the proportion of knowledge type authors can be obtained by identifying the user's viewing history data, and then it can be determined whether to recommend knowledge type authors to the user or whether to select knowledge type authors as the first search result when the user searches, thereby improving the user demand satisfaction.

[0167] In an optional embodiment, based on the confidence of the published videos in each data set belonging to the time interval, the operation of generating a first sub-vector for each data set belonging to the time interval, the embodiments of the present disclosure can specifically include the following steps:

[0168] Step 701, for any data set belonging to the time interval, the published videos in the data set are sorted according to a preset sorting index, and a sorting result is obtained.

[0169] The above-mentioned preset sorting index can be understood as an index most relevant to author type identification, which can be set according to actual needs. For example, the videos in the data set can be sorted from high to low (or from low to high) according to the effective number of plays of each video, or the total number of plays, or the cumulative number of plays.

[0170] Step 702, based on the sorting result, the confidence of the published videos in the data set is sorted, and a first sub-vector of the data set is generated based on the sorted confidence.

[0171] Specifically, after obtaining the above-mentioned sorting result, the video arrangement order of each published video according to the sorting result can be obtained, so as to sort the confidence of each video according to the video arrangement order, and obtain the first sub-vector of the data set. For example, there are three published videos (A, B, C) in a certain data set, and the corresponding confidence is 0.8, 0.4, and 0.9 respectively. When the sorting result obtained according to step 701 is BCA, the generated set vector is [0.4, 0.9, 0.8]. Further, in order to facilitate subsequent processing and calculation, the confidence can also be normalized, for example, the confidence greater than 0.5 (normalization threshold) can be determined as 1, and the confidence not greater than 0.5 can be determined as 0, so that the set vector can be obtained as [0, 1, 1]. The above-mentioned normalization threshold can be set according to actual needs, and the embodiments of the present disclosure do not limit this.

[0172] In the embodiments of the present disclosure, for any data set belonging to the time interval, the published videos in the data set are sorted according to a preset sorting index to obtain a sorting result; the confidence of the published videos in the data set is sorted based on the sorting result, and a first sub-vector of the data set is generated based on the sorted confidence. In this way, by setting different preset sorting indexes, the first sub-vector generated for the data set can meet different identification requirements, thereby improving the diversity of author type identification.

[0173] Figure 3 is a schematic diagram of a target identification model according to an exemplary embodiment of the present disclosure, as shown in Figure 3 , wherein the BERT in the figure is a classification layer, which is used to output a classification result as a knowledge author or a non-knowledge author according to the input information. Alternatively, the classification layer can also use other deep learning classification networks (for example, TextCNN, etc.).

[0174] , wherein the left part of the figure represents the attribute information of the obtained author, and the title in the figure can be understood as the name of the author, and the description can be understood as the introduction of the author. Figure 3 , wherein the left part of the figure represents the attribute information of the obtained author, and the title in the figure can be understood as the name of the author, and the description can be understood as the introduction of the author.

[0175] , wherein the right part of the figure represents the encoding processing of all videos (video 1-video N) published by the author “Dr. Zhang San”, and the encoding process can include obtaining the multi-modal video representation, the knowledge label representation and the upload time representation of each video, and inputting the sum of each representation to the conversion layer (Transformer) to generate a video vector, wherein the multi-modal video representation refers to the third vector generated according to the video multi-modal information, the knowledge label representation refers to the first vector generated according to the video label, and correspondingly, the upload time representation refers to the second vector generated according to the upload time of each video. The specific generation method can refer to the description of the foregoing embodiments, which will not be described here.

[0176] Further, the BERT layer can also perform corresponding encoding processing on the basic attribute information of the author to generate a text vector sequence, and identify the final classification result according to the text vector sequence and the video vector sequence obtained through the first, second and third vectors. Figure 3 The target identification model shown in the embodiments of the present disclosure can encode the user historical videos according to various representation features (upload time, knowledge label, multi-modal video information), thereby improving the identification effect of the knowledge author of the video platform.

[0177] In an optional embodiment, the target identification model is obtained by the following method:

[0178] Step 801, obtaining sample author related information of the sample author and sample content related information of a published video created by the sample author.

[0179] The sample author can be preselected from any video platform. Specifically, a preset number of publishers can be randomly selected from all publishers included in the video platform as sample authors. Further, the sample author related information refers to basic information related to the sample author, which can include the name, biography, number of fans, and number of published videos of the sample author, and the like. Correspondingly, the sample content related information can be obtained from the video content published by the sample author.

[0180] Specifically, the sample author related information and the sample content related information can be pre-obtained. Step 802, inputting the sample author related information and the sample content related information into a to-be-trained author type identification model to encode the sample author related information into a sample text vector sequence, encode the sample content related information into a sample video vector sequence, and perform author type classification and identification on the sample author based on the sample text vector sequence and the sample video vector sequence to obtain a sample identification result of the sample author.

[0181] The to-be-trained author type identification model can be pre-constructed or pre-selected. Specifically, a previously built author type identification model can be randomly selected as the to-be-trained model.

[0182] The sample text vector sequence is obtained by encoding the sample author related information. Specifically, the encoding algorithm pre-set in the model can be used. It can be understood that the sample author related information is usually in the form of text or characters, so the encoding algorithm can use a pre-set character encoding algorithm, for example, ASCII encoding, Unicode encoding, or the like, which is not limited in the present disclosure. Further, when the sample author related information includes information of different dimensions, the information of different dimensions can be separated, for example, a separator such as [sep] can be used to separate the separated string or text string, and then the corresponding sample text vector sequence can be obtained by encoding the string or text string. The sample video vector sequence is obtained by encoding the sample content related information.

[0183] Further, after obtaining the sample text vector sequence and the sample video vector sequence, the two can be spliced to obtain a spliced vector, so that the model can perform multi-dimensional classification and identification on the sample author to obtain a sample identification result.

[0184] Specifically, the author type identification model to be trained can include an encoding layer and a classification layer, so that the operation of encoding the sample author related information and the sample content related information can be performed by the encoding layer, and after the sample text vector sequence and the sample video vector sequence output by the encoding layer are spliced to obtain a spliced vector, the spliced vector can be taken as the input of the classification layer, and the output of the classification layer is determined as the sample identification result.

[0185] Step 803, based on the sample identification result and the sample author label of the sample author, the author type identification model to be trained is adjusted.

[0186] The sample author label is used to represent whether the sample author is a knowledge type author. Specifically, the sample author label can be related to the author related information and the content related information, and can include a knowledge type and other types. When the sample author label is a knowledge type, the corresponding sample author is considered to be a knowledge type author. Correspondingly, when the sample author label is a type other than the knowledge type, the corresponding sample author is considered to be a non-knowledge type author. The sample author label can be represented by 1, 0, Y or N. Specifically, when the sample author label is 1 or Y, the corresponding sample author is considered to be a knowledge type author. Correspondingly, when the sample author label is 0 or N, the corresponding sample author is considered to be a non-knowledge type author.

[0187] The sample author label can be obtained by manual marking. For example, the input information of the staff for different sample authors can be received to obtain the sample author label corresponding to each sample author. The input information can be used to represent whether the sample author is a knowledge type author.

[0188] Step 804, in the case that the training process of the author type identification model to be trained reaches a stop condition, the author type identification model reaching the stop condition is determined as the target identification model.

[0189] It can be understood that the sample identification result output by the author type identification model to be trained can be obtained by the above step 802. The result can represent the type of the sample author, i.e., whether it is a knowledge type author. Therefore, the model output sample identification result and the sample author label in step 803 can be used to measure whether the model can perform author type identification, and the model to be trained can be optimized.

[0190] Specifically, in the embodiments of the present disclosure, the loss value of the to-be-trained model can be obtained through the sample recognition result, the sample author label and the preset loss function (for example, cross entropy), and then the model parameters can be optimized through an optimizer, for example, a stochastic gradient optimization, an adaptive gradient optimization (Adagrad) or an adaptive moment estimation (Adam). Further, the stop condition can be that the loss value is minimum, the loss value is reduced to a loss value threshold, or the training round number reaches a preset round number threshold, so that when the stop condition is met, the optimization is stopped, and the author type recognition model obtained at present, that is, the optimized author type recognition model, is taken as the target recognition model.

[0191] In the embodiments of the present disclosure, the sample author related information of the sample author and the sample content related information of the published video created by the sample author are obtained; the sample author related information and the sample content related information are input into a to-be-trained author type recognition model to encode the sample author related information into a sample text vector sequence and the sample content related information into a sample video vector sequence, and the sample author is classified and recognized in terms of author type based on the sample text vector sequence and the sample video vector sequence to obtain a sample recognition result of the sample author; the to-be-trained author type recognition model is adjusted based on the sample recognition result and a sample author label of the sample author; and in the case that the training process of the to-be-trained author type recognition model reaches a stop condition, the author type recognition model that reaches the stop condition is determined as the target recognition model. In this way, the target recognition model generated in the above manner can automatically recognize a large number of video authors in a video platform and recognize knowledge class authors, thereby facilitating the platform to perform corresponding operations on the knowledge class authors. Meanwhile, the target recognition model can recognize and classify authors according to multi-aspect and multi-dimensional data by encoding and recognizing author related information and content related information, thereby improving the accuracy of author classification and recognition and enabling the to-be-recognized author to be accurately recognized as a knowledge class author.

[0192] Figure 4 is a block diagram of a video author recognition device according to an exemplary embodiment, as shown in Figure 4 The device 90 can include:

[0193] The information obtaining module 901 is configured to perform obtaining author related information of a to-be-recognized author and content related information of a published video created by the to-be-recognized author;

[0194] The information input module 902 is configured to input the author-related information and the content-related information into a target identification model to encode the author-related information into a text vector sequence and encode the content-related information into a video vector sequence. The target identification model is trained based on sample author-related information of a sample author, sample content-related information of published videos created by the sample author, and a sample author label, where the sample author label is used to represent whether the sample author is a knowledge author.

[0195] The type identification module 903 is configured to perform author type classification and identification on the to-be-identified author based on the text vector sequence and the video vector sequence to obtain an author type identification result of the to-be-identified author, where the author type identification result is used to represent whether the to-be-identified author is a knowledge author.

[0196] In an optional embodiment, the number of published videos created by the to-be-identified author is at least two, and the content-related information includes video tags and video multi-modal information. The video tags are used to represent whether the published videos are knowledge videos, and the video tags are determined based on video content of the published videos. The information input module 902 includes:

[0197] The first vector generation submodule is configured to perform encoding processing on the video tags of each of the published videos to obtain a first vector of each of the published videos.

[0198] The classification submodule is configured to perform classification of the published videos into a plurality of data sets corresponding to different time intervals based on video upload times of the published videos. The plurality of data sets belong to at least two different time intervals.

[0199] The second vector generation submodule is configured to perform obtaining of a confidence of a published video in each of the data sets and generation of a second vector based on the confidence of each of the published videos.

[0200] The third vector generation submodule is configured to perform generation of a third vector for each of the published videos based on video multi-modal information of each of the published videos.

[0201] The vector sequence generation submodule is configured to perform generation of the video vector sequence based on the first vector, the second vector, and the third vector of each of the published videos.

[0202] In an optional embodiment, the classification submodule is specifically configured to perform:

[0203] For any of the time intervals, determine published videos belonging to the time interval from the respective published videos;

[0204] Based on the time span corresponding to the time interval and the number of data sets corresponding to the time interval, determine the set duration corresponding to the data set belonging to the time interval; the set duration is used to represent the upper limit value of the difference between the video upload times of the videos in the data set;

[0205] Based on the set duration and the time span of the time interval, determine the time range corresponding to each data set belonging to the time interval;

[0206] For any published video belonging to the time interval, based on the video upload time of the published video, classify the published video to the data set corresponding to the time range to which the video upload time of the published video belongs.

[0207] In an optional embodiment, the second vector generation submodule is specifically configured to perform:

[0208] For any of the time intervals, generate a first sub-vector for each data set belonging to the time interval based on the confidence of the published video in each data set belonging to the time interval;

[0209] Generate a second sub-vector corresponding to the time interval based on the first sub-vector of each data set, and splice the second sub-vectors corresponding to all time intervals;

[0210] Generate a second vector for each of the published videos based on the spliced second sub-vectors.

[0211] In an optional embodiment, the vector sequence generation submodule is specifically configured to perform:

[0212] Splice the first vector, the second vector and the third vector of each of the published videos to obtain the video vector of each of the published videos;

[0213] For any of the published videos, based on the first encoding layer in the target recognition model, process the video vector of the published video to generate a target confidence of the published video;

[0214] Splice the video vectors of the published videos with a target confidence not less than a preset confidence threshold to generate the video vector sequence.

[0215] In an optional embodiment, the author-related information includes attribute information of different dimensions of the to-be-identified author; the information input module 902 includes:

[0216] a text splicing submodule configured to perform splicing of text included in the attribute information of different dimensions based on a second encoding layer in the target recognition model to obtain target text;

[0217] a character encoding submodule configured to perform encoding of each character in the target text to obtain a character vector of each character;

[0218] a text vector generation submodule configured to perform splicing of the character vectors of each character to generate the text vector sequence.

[0219] In an optional embodiment, the second vector generation submodule is specifically configured to perform:

[0220] for any data set belonging to the time interval, sorting the published videos in the data set according to a preset sorting index to obtain a sorting result;

[0221] sorting the confidence of the published videos in the data set based on the sorting result, and generating a first sub-vector of the data set based on the sorted confidence.

[0222] In an optional embodiment, the target recognition model is obtained by the following method:

[0223] obtaining sample author related information of the sample author and sample content related information of the published videos created by the sample author;

[0224] inputting the sample author related information and the sample content related information into a to-be-trained author type recognition model to encode the sample author related information into a sample text vector sequence, encode the sample content related information into a sample video vector sequence, and perform author type classification and recognition of the sample author based on the sample text vector sequence and the sample video vector sequence to obtain a sample recognition result of the sample author;

[0225] adjusting the to-be-trained author type recognition model based on the sample recognition result and a sample author label of the sample author;

[0226] in a case where a training process of the to-be-trained author type recognition model reaches a stop condition, determining the author type recognition model reaching the stop condition as the target recognition model.

[0227] To sum up, the video author identification device provided by the embodiments of the present disclosure, by obtaining author related information of a to-be-identified author and content related information of a published video created by the to-be-identified author; inputting the author related information and the content related information into a target identification model to encode the author related information into a text vector sequence and encode the content related information into a video vector sequence; the target identification model is trained based on sample author related information of a sample author, sample content related information of a published video created by the sample author, and a sample author label, the sample author label is used to represent whether the sample author is a knowledge type author; based on the text vector sequence and the video vector sequence, the to-be-identified author is classified and identified to obtain an author type identification result of the to-be-identified author; the author type identification result is used to represent whether the to-be-identified author is a knowledge type author. In this way, since the target identification model is trained based on the sample author related information of the sample author, the sample content related information of the published video created by the sample author, and the sample author label, and the sample author label can represent whether the sample author is a knowledge type author, the target identification model can identify and classify the to-be-identified author according to multi-aspect and multi-dimensional data, and the accuracy of video author identification can be improved to some extent, so that whether the to-be-identified author is a knowledge type author can be accurately identified.

[0228] According to an embodiment of the present disclosure, an electronic device is provided, comprising: a processor, a memory for storing processor-executable instructions, wherein the processor is configured to execute the steps of the video author identification method in any one of the above embodiments.

[0229] According to an embodiment of the present disclosure, a computer-readable storage medium is also provided, when the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device can execute the steps of the video author identification method in any one of the above embodiments.

[0230] According to an embodiment of the present disclosure, a computer program product is also provided, the computer program product comprises readable program instructions, when the readable program instructions are executed by the processor of the electronic device, the electronic device can execute the steps of the video author identification method in any one of the above embodiments.

[0231] Figure 5is a block diagram of an apparatus for video author identification according to an example embodiment. The apparatus 1000 can include a processing component 1002, a memory 1004, a power supply component 1006, a multimedia component 1008, an audio component 1010, an input / output (I / O) interface 1012, a sensor component 1014, a communication component 1016, and a processor 1020. The processing component 1002 can include one or more processors 1020 to execute instructions to complete all or a part of steps of the above method for video author identification. In an example embodiment, a storage medium can be provided including instructions, e.g., the memory 1004 including instructions, which can be executed by the processor 1020 of the apparatus 1000 to complete the above method. Alternatively, the storage medium can be a non-transitory computer-readable storage medium, e.g., the non-transitory computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disc, and an optical data storage device, etc.

[0232] Figure 6 is a block diagram of another apparatus for video author identification according to an example embodiment. The apparatus 1100 can include a processing component 1122, a memory 1132, an input / output interface 1158, a network interface 1150, and a power supply component 1126. The apparatus 1100 can be provided as a server. The applications stored in the memory 1132 can include one or more than one module each corresponding to a set of instructions. In addition, the processing component 1122 is configured to execute the instructions to perform the above method for video author identification.

[0233] The user information (including but not limited to the user's device information, the user's personal information, etc.) and related data involved in the present disclosure are information authorized by the user or authorized by each party.

[0234] Other embodiments of the present disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the aspects of the present disclosure disclosed herein. It is intended that the present disclosure cover any and all variations of the present disclosure including combinations or permutations of individual features of the present disclosure. The true scope of the present disclosure is indicated by the appended claims, along with the full scope of equivalents to which such claims are entitled. The specification and examples given are intended as illustrative only and not in limitation of the present disclosure, which is to be governed by the claims appended hereto.

[0235] It should be understood that the present disclosure is not limited to the precise structures herein described and illustrated, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is indicated by the appended claims, rather than by the description.

Claims

1. A method for identifying video authors, characterized in that, The method includes: The process involves obtaining author-related information of the author to be identified, content-related information of the published videos created by the author to be identified, and the upload time of the published videos created by the author to be identified; the number of published videos created by the author to be identified is at least two, and the content-related information includes video tags and video multimodal information; the video tags are used to characterize whether the published videos are knowledge-based videos, and the video tags are determined based on the video content of the published videos; The author-related information and content-related information are input into the target recognition model to encode the author-related information into a text vector sequence and the content-related information into a video vector sequence. The target recognition model is trained based on the sample author-related information of the sample author, the sample content-related information of the published videos created by the sample author, and the sample author tags. The sample author tags are used to characterize whether the sample author is a knowledge-based author. The author type is classified and identified based on the text vector sequence and the video vector sequence to obtain the author type identification result of the author to be identified; the author type identification result is used to characterize whether the author to be identified is a knowledge author; Encoding the content-related information into a video vector sequence includes: The video tags of each of the published videos are encoded to obtain the first vector of each of the published videos; Based on the video upload time of each of the published videos, the published videos are classified into multiple data sets corresponding to different time intervals, and the confidence level of the published videos in each of the data sets is obtained. A second vector is generated based on the confidence level of each of the published videos; the multiple data sets belong to at least two different time intervals. A third vector is generated for each of the published videos based on the video multimodal information of each of the published videos; The video vector sequence is generated based on the first vector, second vector, and third vector of each of the published videos.

2. The method according to claim 1, characterized in that, The step of classifying the published videos into multiple data sets corresponding to different time intervals based on their upload times includes: For any given time interval, determine the published videos belonging to that time interval from among the published videos; Based on the time span corresponding to the time interval and the number of data sets corresponding to the time interval, the set duration corresponding to the data set belonging to the time interval is determined; the set duration is used to characterize the upper limit of the difference between the video upload times of the videos in the data set. Based on the duration of the data set and the time span of the time interval, determine the time range corresponding to each data set belonging to the time interval; For any published video belonging to the time interval, based on the video upload time of the published video, the published video is classified into the data set corresponding to the time range to which the video upload time of the published video belongs.

3. The method according to claim 2, characterized in that, The generation of the second vector based on the confidence scores of each of the published videos includes: For any of the time intervals, a first sub-vector is generated for each data set belonging to the time interval based on the confidence of the published videos in each data set belonging to the time interval. Generate a second sub-vector corresponding to the time interval based on the first sub-vector of each of the data sets, and concatenate the second sub-vectors corresponding to all time intervals; Based on the concatenated second sub-vector, a second vector is generated for each of the published videos.

4. The method according to claim 1, characterized in that, The step of generating the video vector sequence based on the first vector, second vector, and third vector of each of the published videos includes: Based on the first vector, second vector, and third vector of each of the published videos, the video vector of each of the published videos is obtained; For any of the published videos, the video vector of the published video is processed based on the first encoding layer in the target recognition model to generate the target confidence score of the published video. The video vector sequence is generated by concatenating the video vectors of published videos whose target confidence level is not less than a preset confidence threshold.

5. The method according to any one of claims 1-4, characterized in that, The author-related information includes attribute information of the author to be identified in different dimensions; encoding the author-related information into a text vector sequence includes: Based on the second encoding layer in the target recognition model, the text included in the attribute information of different dimensions is concatenated to obtain the target text; Each character in the target text is encoded to obtain a character vector for each character; The character vectors of each character are concatenated to generate the text vector sequence.

6. The method according to claim 3, characterized in that, The step of generating a first sub-vector for each data set belonging to the time interval based on the confidence score of the published videos in each data set belonging to the time interval includes: For any data set belonging to the time interval, the published videos in the data set are sorted according to a preset sorting index to obtain the sorting result; The confidence scores of the published videos in the dataset are sorted based on the sorting results, and a first sub-vector of the dataset is generated based on the sorted confidence scores.

7. The method according to any one of claims 1-4 and 6, wherein the target recognition model is trained in the following manner: Obtain the sample author information of the sample author and the sample content information of the published videos created by the sample author; The relevant information of the sample author and the relevant information of the sample content are input into the author type recognition model to be trained, so as to encode the relevant information of the sample author into a sample text vector sequence and the relevant information of the sample content into a sample video vector sequence. Based on the sample text vector sequence and the sample video vector sequence, the author type of the sample author is classified and identified to obtain the sample recognition result of the sample author. Based on the sample identification results and the sample author labels of the sample authors, the author type identification model to be trained is adjusted. If the training process of the author type recognition model to be trained reaches the stopping condition, the author type recognition model that has reached the stopping condition is determined as the target recognition model.

8. A video author identification device, characterized in that, The device includes: The information acquisition module is configured to acquire author-related information of the author to be identified, content-related information of published videos created by the author to be identified, and the video upload time of the published videos created by the author to be identified; the number of published videos created by the author to be identified is at least two, and the content-related information includes video tags and video multimodal information; the video tags are used to characterize whether the published videos are knowledge-based videos, and the video tags are determined based on the video content of the published videos; The information input module is configured to input the author-related information and content-related information into the target recognition model, so as to encode the author-related information into a text vector sequence and the content-related information into a video vector sequence; the target recognition model is trained based on the sample author-related information of the sample author, the sample content-related information of the published video created by the sample author, and the sample author tags, wherein the sample author tags are used to characterize whether the sample author is a knowledge-based author; The type recognition module is configured to perform author type classification and recognition on the author to be identified based on the text vector sequence and the video vector sequence, and obtain the author type recognition result of the author to be identified; the author type recognition result is used to characterize whether the author to be identified is a knowledge author; The information input module includes: The first vector generation submodule is configured to perform encoding processing on the video tags of each of the published videos to obtain the first vector of each of the published videos; The classification submodule is configured to classify the published videos into multiple data sets corresponding to different time intervals based on the video upload time of each of the published videos; the multiple data sets belong to at least two different time intervals; The second vector generation submodule is configured to obtain the confidence scores of published videos in each of the datasets and generate a second vector based on the confidence scores of each of the published videos. The third vector generation submodule is configured to generate a third vector for each of the published videos based on the video multimodal information of each of the published videos; The vector sequence generation submodule is configured to generate the video vector sequence based on the first vector, second vector, and third vector of each of the published videos.

9. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor of the electronic device, the electronic device performs the method as described in any one of claims 1 to 7.

Images