Detection method, device, equipment, medium and product

Abstract

The invention discloses a detection method and device, equipment, a medium and a product, and the method comprises the steps: carrying out the retrieval of a resource pool according to at least one feature of a target video after the target video is obtained, and obtaining a plurality of candidate videos; screening out a plurality of rough videos from the plurality of candidate videos according to a similarity calculation result between each candidate video and the target video; then, according to the similarity score between each roughed video and the target video, screening out a plurality of carefully selected videos from the plurality of roughed videos; and finally, for any selected video, determining a judgment result between the target video and the selected video by using the generative model, so that the judgment result is used for indicating whether the target video and the selected video describe the same thing or not, thereby reasoning whether the two videos are similar or not by means of the generative capability of the generative model. Therefore, similarity can be found as comprehensively as possible, and whether similarity exists or not can be judged as accurately as possible, so that the repeated detection accuracy and the recommendation effect are improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a detection method, device, equipment, medium, and product. Background Art

[0002] For some video recommendation fields, such as the news recommendation field composed of videos, in order to avoid recommending videos with repeated information, after obtaining a new video waiting to be recommended, it is necessary to detect whether the new video is repeated in content with the videos recommended in the past. If so, the new video will not be recommended; if not, the new video will be recommended.

[0003] However, for some videos, when these videos describe the same thing in different words, there are certain differences between the contents described in these videos, making it difficult to identify the similarities of these videos during repeated detection, thus affecting the recommendation effect. Summary of the invention

[0004] In order to solve the above technical problems, the present application provides a detection method, device, equipment, medium, and product, which are conducive to improving the recommendation effect.

[0005] In order to achieve the above objectives, the technical solutions provided by this application are as follows:

[0006] The present application provides a detection method, which includes: after obtaining a target video, searching a resource pool according to at least one feature of the target video to obtain multiple candidate videos, wherein the resource pool is used to record videos that have been recommended in a historical time period; based on the similarity calculation results between each of the candidate videos and the target video, screening out multiple rough-selected videos from the multiple candidate videos; the similarity calculation results include at least one index value, and different index values ​​are used to characterize similarities under different dimensions; based on the similarity scores between each of the rough-selected videos and the target video, screening out multiple selected videos from the multiple rough-selected videos; the similarity scores are obtained using a target model, and the target model is trained using a sample video, a positive example video corresponding to the sample video, and a negative example video corresponding to the sample video, wherein the sample video and the positive example video both describe the same thing, and the sample video and the negative example video respectively describe different things; for any of the selected videos, using a generative model to determine a judgment result between the target video and the selected video, and the judgment result is used to indicate whether the target video and the selected video describe the same thing.

[0007] In one possible implementation, the at least one feature includes: a video vector of the target video, a text vector of a title of the target video, and keywords extracted from the title; the process of determining the multiple candidate videos includes: obtaining retrieval results of the video vector, retrieval results of the text vector, and retrieval results of the keywords from the resource pool; and determining the multiple candidate videos based on the retrieval results of the video vector, the retrieval results of the text vector, and the retrieval results of the keywords.

[0008] In one possible implementation, for any of the candidate videos, the similarity calculation result between the candidate video and the target video includes at least one of a first index value, a second index value, and a third index value; the first index value is determined based on the similarity between the video vector of the candidate video and the video vector of the target video; the second index value is determined based on the similarity between the title of the candidate video and the title of the target video; the third index value is determined based on the similarity between the text vector of the title of the candidate video and the text vector of the title of the target video.

[0009] In one possible implementation, the similarity calculation result includes N index values, where N is a positive integer; the process of determining the multiple roughly selected videos includes: screening the multiple candidate videos according to the nth index value between each of the candidate videos and the target video to obtain a screening result corresponding to the nth index value, where n is a positive integer, n≤N; and determining the multiple roughly selected videos according to the screening results corresponding to the N index values.

[0010] In one possible implementation, for any of the roughly selected videos, the process of determining the similarity score between the roughly selected video and the target video includes: inputting the roughly selected video, the title of the roughly selected video, the target video, and the title of the target video into the target model to obtain output data of the target model, wherein the output data includes the similarity score between the roughly selected video and the target video.

[0011] In one possible implementation, the target model includes a text encoding module, a video encoding module, a fully connected layer and a scoring module; the text encoding module is used to obtain the encoding result of the title of the roughly selected video and the encoding result of the title of the target video; the video encoding module is used to obtain the encoding result of the roughly selected video and the encoding result of the target video; the fully connected layer is used to process the splicing result between the encoding result of the title of the roughly selected video and the encoding result of the roughly selected video, and to process the splicing result between the encoding result of the title of the target video and the encoding result of the target video; the scoring module is used to process the output data of the fully connected layer to obtain a similarity score between the roughly selected video and the target video.

[0012] In a possible implementation manner, the method further includes: determining whether to recommend the target video based on the judgment result.

[0013] The present application provides a detection device, comprising: a retrieval unit, which is used to retrieve a resource pool according to at least one feature of the target video after acquiring a target video, and obtain multiple candidate videos, wherein the resource pool is used to record videos that have been recommended in a historical time period; a rough selection unit, which is used to filter out multiple rough selected videos from the multiple candidate videos according to the similarity calculation result between each of the candidate videos and the target video; the similarity calculation result includes at least one index value, and different index values ​​are used to characterize the similarity under different dimensions; a selection unit, which is used to filter out multiple selected videos from the multiple rough selected videos according to the similarity score between each of the rough selected videos and the target video; the similarity score is obtained using a target model, and the target model is trained using a sample video, a positive example video corresponding to the sample video, and a negative example video corresponding to the sample video, wherein the sample video and the positive example video both describe the same thing, and the sample video and the negative example video respectively describe different things; a judgment unit, which is used to determine, for any of the selected videos, a judgment result between the target video and the selected video using a generative model, and the judgment result is used to indicate whether the target video and the selected video describe the same thing.

[0014] The present application provides an electronic device, which includes: a processor and a memory; the memory is used to store instructions or computer programs; the processor is used to execute the instructions or computer programs in the memory, so that the electronic device executes the detection method provided by the present application.

[0015] The present application provides a computer-readable medium, in which instructions or computer programs are stored. When the instructions or computer programs are executed on a device, the device executes the detection method provided by the present application.

[0016] The present application provides a computer program product, which includes a computer program carried on a non-transitory computer-readable medium, and the computer program contains program codes for executing the detection method provided by the present application.

[0017] Compared with the related art, this application has at least the following advantages:

[0018] In the detection scheme provided by the present application, for a target video (such as a new video waiting to be recommended), after obtaining the target video, firstly, based on at least one feature of the target video, a resource pool is searched to obtain multiple candidate videos, so that these candidate videos can represent videos that have been recommended in a historical time period and may be similar to the target video; then, based on the similarity calculation results (such as multiple index values) between each candidate video and the target video, multiple roughly selected videos are screened out from the multiple candidate videos, so that these roughly selected videos can represent videos that are obtained by a rough sorting method and are relatively similar to the target video in at least one dimension; then, based on the similarity scores between each roughly selected video and the target video, multiple fine selected videos are screened out from the multiple roughly selected videos. Select videos so that these selected videos can be obtained through a precise sorting method and are more likely to be similar to the target video; finally, for any selected video, use the generative model to determine the judgment result between the target video and the selected video, so that the judgment result is used to indicate whether the target video and the selected video describe the same thing, so that the generative ability of the generative model can be used to infer whether the two videos are similar, thereby effectively avoiding defects when manually setting a similarity threshold to determine whether the two videos are similar, such as poor flexibility or inaccurate judgment results, and then it is possible to find similarities as comprehensively as possible and judge whether they are similar as accurately as possible, so as to improve the accuracy of duplicate detection, which is conducive to improving the recommendation effect.

[0019] In addition, because the target model is trained using sample videos, positive example videos corresponding to the sample videos, and negative example videos corresponding to the sample videos, the model can more accurately identify different videos used to describe the same thing, so that the similarity score obtained based on the model can more accurately describe the similarity between the two videos, and then the video selected based on the similarity score is more likely to be similar to the target video, which is conducive to better improving the accuracy of duplicate detection, thereby helping to better improve the recommendation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 A flow chart of a detection method provided in an embodiment of the present application;

[0022] Figure 2 A schematic diagram of recall and rough sorting provided in an embodiment of the present application;

[0023] Figure 3 A schematic diagram of the structure of a target model provided in an embodiment of the present application;

[0024] Figure 4 A schematic diagram of a model training process provided in an embodiment of the present application;

[0025] Figure 5 A schematic diagram of the structure of a detection device provided in an embodiment of the present application;

[0026] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0027] Through research, it is found that some duplicate detection schemes are implemented based on text matching or keyword analysis. To facilitate understanding, the following is an explanation with some examples.

[0028] Example 1: A duplicate detection solution based on keyword analysis includes: first extracting keywords from each content (such as an article); and then comparing these keywords to determine whether different contents are similar.

[0029] Example 2, a duplicate detection scheme based on Term Frequency-Inverse Document Frequency (TF-IDF) includes: first, for any content (such as an article), using TF-IDF, measure the importance of each word in the content; then, based on the importance, determine whether different contents are similar.

[0030] Example 3: A duplicate detection solution based on string matching includes: using a similarity evaluation algorithm (such as edit distance, Jaccard similarity, etc.) to directly compare whether different contents are similar.

[0031] Example 4, a duplicate detection scheme based on fingerprint analysis includes: first, for any content (such as an article), the content is hashed to obtain the hash value of the content as the fingerprint corresponding to the content; then, by comparing the fingerprints, it is determined whether different contents are similar.

[0032] Example 5, a duplicate detection scheme based on a vector space model includes: first, for any content (such as an article), the content is converted into a vector representation by means of a certain vector conversion method (such as Bag of Words model, TF-IDF, text encoding, etc.); then, by comparing the vector representations, it is determined whether different contents are similar.

[0033] Example 6: A duplicate detection scheme based on clustering and classification includes: firstly grouping all contents by using a clustering algorithm or a classification algorithm; and then determining whether different contents are similar by determining whether they belong to the same group.

[0034] The study also found that the duplicate detection schemes described in the above examples have the following defects: because these schemes rely on text matching or keyword analysis, they cannot accurately capture the deep semantics of the text, making them susceptible to vocabulary changes, making it difficult for these schemes to identify the content that describes the same thing in different wordings (such as different word order, synonyms, and paraphrasing, etc.) as duplicates, thus affecting the duplicate detection performance. For ease of understanding, the following is an explanation with some examples.

[0035] Example 1: For the content "mercury is poisonous" and the content "mercury is poisonous", mercury and mercury are synonyms, so that mercury and mercury both describe the same thing, but because the number of words and shapes of these two words are very different, it is difficult for the above duplication detection scheme to identify that the two contents are similar.

[0036] Example 2: For the content "Very good weather, today" and the content "The weather is very good today", although these two contents describe the same thing, the word order presented in these two contents is completely different, making it difficult for the above-mentioned duplication detection scheme to identify that these two contents are similar.

[0037] Example 3: For the content "This dress is very beautiful" and the content "This dress is not ugly at all", although these two contents describe the same thing, the latter uses a paradoxical way of stating the thing, which makes the two contents appear quite different, making it difficult for the above-mentioned duplication detection scheme to identify that the two contents are similar.

[0038] Based on the above research, in order to improve the performance of duplicate detection as much as possible, the present application provides a detection method, which includes: for a target video (such as a new video waiting to be recommended), after obtaining the target video, first search the resource pool according to at least one feature of the target video to obtain multiple candidate videos, so that these candidate videos can represent videos that have been recommended in a historical time period and may be similar to the target video; then, based on the similarity calculation results (such as multiple index values) between each candidate video and the target video, select multiple roughly selected videos from the multiple candidate videos, so that these roughly selected videos can represent videos that are obtained by a rough sorting method and are relatively similar to the target video in at least one dimension; then, based on the similarity scores between each roughly selected video and the target video, A plurality of selected videos are screened out from a plurality of roughly selected videos, so that these selected videos can be obtained through fine sorting and are more likely to be similar to the target video; finally, for any selected video, a generative model is used to determine a judgment result between the target video and the selected video, so that the judgment result is used to indicate whether the target video and the selected video describe the same thing, so that the generative ability of the generative model can be used to infer whether the two videos are similar, thereby effectively avoiding defects when manually setting a similarity threshold to determine whether the two videos are similar, such as poor flexibility or inaccurate judgment results, and thus achieving the goal of finding similarities as comprehensively as possible and judging whether they are similar as accurately as possible, so as to improve the accuracy of duplicate detection, which is conducive to improving the recommendation effect.

[0039] In addition, because the target model is trained using sample videos, positive example videos corresponding to the sample videos, and negative example videos corresponding to the sample videos, the model can more accurately identify different videos used to describe the same thing, so that the similarity score obtained based on the model can more accurately describe the similarity between the two videos, and then the video selected based on the similarity score is more likely to be similar to the target video, which is conducive to better improving the accuracy of duplicate detection, thereby helping to better improve the recommendation effect.

[0040] In addition, the present application does not limit the execution subject of the detection method. For example, the method can be applied to a terminal device or a server. For another example, the method can also be implemented with the help of a data interaction process between a terminal device and a server. Among them, the terminal device can be a smart phone, a computer, a personal digital assistant (PDA), a tablet computer, etc. The server can be a stand-alone server, a cluster server, or a cloud server.

[0041] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0042] In order to better understand the technical solution provided by the present application, the detection method provided by the present application is described below with reference to some drawings. Figure 1 As shown, the detection method provided in the embodiments of the present application includes S1-S4 below.

[0043] S1: After acquiring a target video, a resource pool is searched based on at least one feature of the target video to obtain a plurality of candidate videos. The resource pool is used to record videos that have been recommended in a historical time period.

[0044] Among them, the target video refers to the video selected in the current recommendation and meets the recommendation conditions, such as a video describing a certain event (such as a news event), so that further repeated detection of the video is needed in the future to determine whether the event described by the video has been recommended before. If it has been recommended, the video will be discarded. If it has not been recommended, a recommendation list is generated based on the video, such as a hot news list.

[0045] At least one feature of the target video is used to describe some characteristics of the video, such as video title content, title semantics, video content, some tag information (such as video publisher, identification category), etc., so that videos that may be similar to the video can be recalled based on each feature.

[0046] In addition, the present application does not limit the implementation methods of the at least one feature mentioned above. For example, in order to maximize the comprehensiveness of recall, at least one feature of the target video may include a video vector of the target video, a text vector of the title of the target video, and keywords extracted from the title. The text vector is obtained by extracting text features from the title using a pre-trained text embedding model, such as Bidirectional Encoder Representations from Transformers (Bert), so that the text vector is used to represent the semantic information carried by the title. The video vector is obtained by extracting video features from the target video using a video embedding model (such as an optical flow network or video Bert), so that the video vector can better represent the things described by the target video. The keyword is a keyword extracted from the word segmentation result of the title, so that the keyword can represent the core content described by the title.

[0047] For another example, in order to improve the comprehensiveness of recall, the at least one feature of the target video may also include at least one tag of the video. The tag refers to information annotated for the video and used to describe the characteristics of the video; and the application does not limit the at least one tag, for example, it may include the publisher of the video, the type of the video, etc.

[0048] For example, in order to better improve the recall effect, at least one feature of the above-mentioned target video may also include an identity document (ID) of the news composed of the video and / or the ID of the video, so that it can be determined later with the help of the ID whether the video or the news composed of the video has been recommended.

[0049] The resource pool is used to record videos that have been recommended in the historical time period, such as Figure 2 It should be noted that the present application does not limit the implementation method of the historical time period. For example, the historical time period can be set according to the actual application scenario to ensure that the things described in the target video to be recommended have not been recommended in the historical time period.

[0050] Multiple candidate videos refer to videos that are recalled from a resource pool and may be similar to a target video. For example, when at least one feature of a target video includes M features, the multiple candidate videos may include a video recalled from the resource pool according to the first feature, a video recalled from the resource pool according to the second feature, ..., and a video recalled from the resource pool according to the Mth feature. Wherein M is a positive integer. For ease of understanding, a possible implementation of S1 is used as an example for explanation below.

[0051] As an example, when at least one feature of the target video includes a video vector of the target video, a text vector of the title of the target video, and keywords extracted from the title, the above S1 may specifically include: first obtaining the retrieval results of the video vector, the retrieval results of the text vector, and the retrieval results of the keyword from the resource pool; and then determining multiple candidate videos based on the retrieval results of the video vector, the retrieval results of the text vector, and the retrieval results of the keyword.

[0052] For the video vector of the target video, the retrieval result of the video vector refers to the result obtained by searching the resource pool with the video vector as the retrieval keyword, so that the retrieval result represents the videos recalled from the resource pool according to the video vector, thereby making the retrieval result able to represent the videos that may be similar to the target video in the dimension of video vector and have been recommended in the historical time period.

[0053] For the text vector of the target video title, the retrieval result of the text vector refers to the result obtained by searching the resource pool using the text vector as the retrieval keyword, so that the retrieval result represents the videos recalled from the resource pool according to the text vector, thereby enabling the retrieval result to represent videos that may be similar to the target video in the dimension of the text vector of the video title and have been recommended in a historical time period.

[0054] With respect to extracting keywords from the title of a target video, the search result of the keyword refers to that obtained by searching the resource pool (such as an inverted index) with the keyword as the search keyword, so that the search result represents videos recalled from the resource pool according to the keyword, thereby enabling the search result to represent videos that have been recommended within a historical time period and whose video titles include the keyword.

[0055] In addition, the present application does not limit the implementation method of the above-mentioned step of "determining multiple candidate videos based on the retrieval results of the video vector, the retrieval results of the text vector, and the retrieval results of the keyword". For example, it can be specifically: first aggregate the retrieval results of the video vector, the retrieval results of the text vector, and the retrieval results of the keyword to obtain an aggregated result; then deduplicate the multiple videos with the same video ID in the aggregated result to obtain the multiple candidate videos.

[0056] Based on the relevant content of S1 above, it can be known that in some scenarios, for the target video waiting to be recommended, after obtaining the video, first, the title of the video is segmented to obtain the segmentation result, and keywords are extracted from the segmentation result so that the keyword can represent the core content of the title; Bert is used to convert the title into a text vector so that the text vector can represent the semantic information described by the title; the optical flow network is used to extract video features of the video to obtain a video vector so that the video vector can represent the things described by the video; then, according to the keyword, the text vector and the video vector, the resource pool is recalled to obtain multiple candidate videos, so that these candidate videos can represent the videos that have been recommended and may be similar to the target video, so that these candidate videos can cover all videos that are suspected to be similar to the target video as much as possible, which is conducive to achieving the purpose of finding similarities.

[0057] S2: Filter out a plurality of roughly selected videos from a plurality of candidate videos according to a similarity calculation result between each candidate video and the target video; the similarity calculation result includes at least one index value, and different index values ​​are used to represent similarities under different dimensions.

[0058] Among them, the similarity calculation result between the i-th candidate video and the target video is used to describe the similarity between the i-th candidate video and the target video, i is a positive integer, i≤I, I is a positive integer, and I represents the number of videos in the above-mentioned multiple candidate videos.

[0059] In addition, for the similarity calculation result between the i-th candidate video and the target video, the similarity calculation result can at least satisfy the following constraints: the similarity calculation result includes at least one index value, and different index values ​​are used to characterize the similarity under different dimensions (such as video content, title string, and semantic information of the title).

[0060] In addition, the present application does not limit the above similarity calculation results. For example, the similarity calculation result between the i-th candidate video and the target video may include at least one of the first index value, the second index value, and the third index value. Wherein, the first index value is determined based on the similarity between the video vector of the i-th candidate video and the video vector of the target video (such as cosine similarity), so that the first index value can represent the degree of overlap of the video content. The second index value is determined based on the similarity between the title of the i-th candidate video and the title of the target video (such as edit distance), so that the second index value can represent the degree of string matching of different titles. The third index value is determined based on the similarity between the text vector of the title of the i-th candidate video and the text vector of the title of the target video (such as cosine similarity), so that the third index value can represent the degree of semantic similarity of different titles.

[0061] The multiple roughly selected videos are obtained by roughly sorting the multiple candidate videos to ensure that the number of the roughly selected videos is less than the number of the candidate videos. It should be noted that the roughly sorting is used to eliminate videos that are obviously not similar to the target video to reduce the search range of similar videos as much as possible, which is conducive to better finding similarities.

[0062] In addition, the present application does not limit the determination process of multiple roughly selected videos (that is, the implementation method of S2 above), for example, it can be implemented using any rough sorting method.

[0063] For another example, in order to better improve the rough sorting effect, when the above-mentioned similarity calculation result includes N index values, N is a positive integer, the above-mentioned S2 can be specifically as follows: first, according to the nth index value between each candidate video and the target video, multiple candidate videos are screened to obtain the screening result corresponding to the nth index value, so that the screening result can represent the result obtained by rough sorting according to the nth index value, so that the screening result can represent the video that exists in the multiple candidate videos and is similar to the target video in the dimension represented by the nth index value (such as video content, title string or title semantic information), n is a positive integer, n≤N; then, according to the screening results corresponding to the N index values, multiple roughly selected videos are determined, so that the multiple roughly selected videos can represent videos that are similar to the target video in at least one dimension, so that there is no video that is dissimilar to the target video in all dimensions in the multiple roughly selected videos, so that the videos that are obviously dissimilar to the target video can be eliminated by means of rough sorting, which is beneficial to improving the subsequent similarity search efficiency.

[0064] Based on the relevant content of S2 above, it can be known that in some scenarios, for the recalled multiple candidate videos, first, the cosine similarity between the video vector of the target video and the video vector of each candidate video is calculated, the edit distance between the title of the target video and the title of each candidate video is calculated, and the semantic similarity between the title of the target video and the title of each candidate video is calculated using the text vector of the video title; then, the multiple candidate videos are screened based on the cosine similarity, the multiple candidate videos are screened based on the edit distance, and the multiple candidate videos are screened based on the semantic similarity; finally, the results of these three screenings are summarized to obtain multiple roughly selected videos to ensure that there are no videos in these roughly selected videos that are obviously dissimilar to the target video, which can effectively and significantly narrow the similarity search range, thereby helping to improve the subsequent similarity search efficiency.

[0065] S3: Filtering out multiple selected videos from multiple rough selected videos based on similarity scores between each rough selected video and the target video; the similarity scores are obtained using a target model, and the target model is trained using a sample video, a positive video corresponding to the sample video, and a negative video corresponding to the sample video, the sample video and the positive video both describe the same thing, and the sample video and the negative video respectively describe different things.

[0066] Among them, the similarity score between the jth roughly selected video and the target video is used to more accurately represent the similarity between the jth roughly selected video and the target video, j is a positive integer, j≤J, J is a positive integer, and J represents the number of videos in the above-mentioned multiple roughly selected videos.

[0067] In addition, for the similarity score between the jth rough selected video and the target video, the similarity score is obtained by using a pre-trained target model with good similarity determination performance (such as Figure 3 The present application does not limit the determination process of the similarity score. For example, it may specifically be: inputting the j-th roughly selected video, the title of the j-th roughly selected video, the target video, and the title of the target video into the target model to obtain output data of the target model, so that the output data includes the similarity score between the j-th roughly selected video and the target video.

[0068] In addition, in order to better improve the accuracy of similarity assessment, the target model may include a text encoding module (such as RoBert a), a video encoding module (such as video Bert), a fully connected layer and a scoring module. Among them, the text encoding module is used to obtain the encoding result of the title of the jth rough selected video and the encoding result of the title of the target video; the video encoding module is used to obtain the encoding result of the jth rough selected video and the encoding result of the target video; the fully connected layer is used to process the splicing result between the encoding result of the title of the jth rough selected video and the encoding result of the jth rough selected video, and the splicing result between the encoding result of the title of the target video and the encoding result of the target video; the scoring module is used to process the output data of the fully connected layer (such as the processing result of the above two splicing results) to obtain the similarity score between the jth rough selected video and the target video.

[0069] It can be seen that in a possible implementation mode, the working principle of the target model may include: first, using a text encoding module to encode the title of the j-th rough selected video to obtain the title encoding result of the j-th rough selected video, using a text encoding module to encode the title of the target video to obtain the title encoding result of the target video, using a video encoding module to encode the j-th rough selected video to obtain the video encoding result of the j-th rough selected video, and using a video encoding module to encode the target video to obtain the video encoding result of the target video; then, splicing the title encoding result of the j-th rough selected video with the video encoding result of the j-th rough selected video to obtain the spliced ​​video corresponding to the j-th rough selected video. The result is obtained by splicing the title encoding result of the target video with the video encoding result of the target video to obtain the splicing result corresponding to the target video; secondly, the splicing result corresponding to the j-th rough selected video is processed by the fully connected layer to obtain the processing result corresponding to the j-th rough selected video, and the splicing result corresponding to the target video is processed by the fully connected layer to obtain the processing result corresponding to the target video; finally, the scoring module performs similarity evaluation processing on the processing result corresponding to the j-th rough selected video and the processing result corresponding to the target video to obtain the similarity score between the j-th rough selected video and the target video, so that the similarity score can more accurately represent the similarity between the j-th rough selected video and the target video.

[0070] Furthermore, in order to better improve the performance of the target model, the training process of the target model may include the following steps 11 to 14.

[0071] Step 11: Obtain a sample video, a positive video corresponding to the sample video, and a negative video corresponding to the sample video.

[0072] The sample video refers to a video that needs to be evaluated for similarity with other videos during the training process.

[0073] In addition, the positive example video corresponding to the sample video refers to the positive sample corresponding to the sample video, so that the positive example video is used to represent videos similar to the sample video, so that the things described by the positive example video are the same as the things described by the sample video, and then the positive example video can guide the target model to learn what videos similar to the sample video are like.

[0074] In addition, the negative example video corresponding to the sample video refers to the negative sample corresponding to the sample video, so that the negative example video is used to represent a video that is dissimilar to the sample video, so that the things described by the negative example video are different from the things described by the sample video, and then the negative example video can guide the target model to learn what the video that is dissimilar to the sample video is like.

[0075] Step 12: Input the sample video, the title of the sample video, the positive video and the title of the positive video into the target model to obtain the positive example score output by the target model, and input the sample video, the title of the sample video, the negative video and the title of the negative video into the target model to obtain the negative example score output by the target model.

[0076] The positive example score refers to a similarity score between a sample video and a positive example video corresponding to the sample video, determined by a target model, so that the score is used to describe the degree of similarity between the sample video and the positive example video.

[0077] The negative example score refers to a similarity score between a sample video and a negative example video corresponding to the sample video, determined by a target model, such that the score is used to describe the degree of similarity between the sample video and the negative example video.

[0078] Step 13: Determine the triplet loss (Triplet Loss) based on the positive example score and the negative example score so that the loss (such as Figure 4 The loss shown) can indicate the performance of the target model.

[0079] Step 14: Update the target model based on the triplet loss, and return to continue executing the above step 11 and its subsequent steps until it is determined that the preset stop condition is met (for example, the loss is lower than the preset loss threshold, the rate of change of the loss is lower than the preset rate of change threshold, or the number of updates of the model reaches the preset number threshold), and then end the iterative training process for the target model.

[0080] Based on the relevant contents of the above steps 11 to 14, it can be known that in some scenarios, the target model can be trained using a sample video, a positive video corresponding to the sample video, and a negative video corresponding to the sample video, so that the trained model has better performance, so that the trained model can more accurately judge the similarity between different videos, and then the model can be used to achieve fine sorting in the future. It should be noted that fine sorting is used to select videos that are similar to the target video to achieve finding similarities.

[0081] Based on the relevant content of S3 above, it can be known that in some scenarios, for multiple roughly selected videos obtained by rough sorting, the pre-built target model (such as Figure 3The refined ranking model shown in the figure) is used to determine the similarity score between each roughly selected video and the target video, so that the similarity score can more accurately represent the similarity between each roughly selected video and the target video; and then based on the similarity scores, a plurality of refined videos are screened out from the roughly selected videos, so that the plurality of refined videos can represent the videos existing in the roughly selected videos that are highly similar to the target video (such as the top 5 videos in the similarity score ranking), so that the plurality of refined videos can represent the videos that are relatively high in similarity to the target video.

[0082] S4: For any selected video, a generative model is used to determine a judgment result between a target video and the selected video, where the judgment result is used to indicate whether the target video and the selected video describe the same thing.

[0083] Among them, the generative model refers to a model with generative capabilities (such as the capabilities shown in Table 1 below), so that it can subsequently accurately determine whether two videos are similar with the help of this capability, so as to overcome the defects that exist when selecting the most similar video by manually setting the scoring threshold, such as the different thresholds in different scenarios, resulting in the video selected by the threshold method may be similar to the target video, or may not be similar. It should be noted that the present application does not limit the implementation method of the generative model. For example, it can be implemented using any large language model (such as GPT).

[0084]

[0085] Table 1. Identifying similar methods using generative models

[0086] Based on the relevant content of the above-mentioned generative model, it can be known that after obtaining multiple selected videos through fine sorting, for any selected video, the model is used to determine the judgment result between the target video and the selected video, so that the judgment result can indicate whether the target video and the selected video describe the same thing, so that the judgment result can accurately indicate whether the two videos are similar. It can be seen that the present application can make autonomous inferences on different contents with the help of the powerful semantic understanding ability of the model to determine the repetitiveness (such as similarity) of different contents, which can effectively improve the accuracy and efficiency of duplicate content detection.

[0087] Based on the relevant contents of S1 to S4 above, it can be known that the duplicate detection scheme provided by the present application includes: for a target video (such as a new video waiting to be recommended), after obtaining the target video, first search the resource pool according to at least one feature of the target video to obtain multiple candidate videos, so that these candidate videos can represent videos that have been recommended in a historical time period and may be similar to the target video; then, based on the similarity calculation results (such as multiple index values) between each candidate video and the target video, screen out multiple rough-selected videos from the multiple candidate videos, so that these rough-selected videos can represent videos that are obtained by rough sorting and are relatively similar to the target video in at least one dimension; then, based on the similarity scores between each rough-selected video and the target video, select the rough-selected videos from the multiple rough-selected videos. A plurality of selected videos are screened out from the selected videos, so that these selected videos can be obtained through fine sorting and are more likely to be similar to the target video; finally, for any selected video, a generative model is used to determine the judgment result between the target video and the selected video, so that the judgment result is used to indicate whether the target video and the selected video describe the same thing, so that the generative ability of the generative model can be used to infer whether the two videos are similar, thereby effectively avoiding defects when manually setting a similarity threshold to determine whether the two videos are similar, such as poor flexibility or inaccurate judgment results, and then it is possible to find similarities as comprehensively as possible and judge whether they are similar as accurately as possible, so as to improve the accuracy of duplicate detection, which is conducive to improving the recommendation effect.

[0088] In addition, because the target model is trained using sample videos, positive example videos corresponding to the sample videos, and negative example videos corresponding to the sample videos, the model can more accurately identify different videos used to describe the same thing, so that the similarity score obtained based on the model can more accurately describe the similarity between the two videos, and then the video selected based on the similarity score is more likely to be similar to the target video, which is conducive to better improving the accuracy of duplicate detection, thereby helping to better improve the recommendation effect.

[0089] In addition, the present application also provides a video recommendation solution, which includes not only the above S1-S4, but also the following step 21.

[0090] Step 21: Determine whether to recommend the target video based on the judgment result between the target video and each selected video.

[0091] In the present application, after obtaining the judgment results between the target video and the selected videos, if these judgment results indicate that the target video is not similar to any of the selected videos, it can be determined that the thing described by the target video has not been recommended, so it can be determined that the target video is recommended; however, if these judgment results indicate that the target video is similar to at least one selected video, it can be determined that the thing described by the target video has been recommended, so it can be determined that the target video is not recommended. This can effectively overcome the defects caused by repeatedly recommending the same thing, such as affecting user experience.

[0092] Based on the relevant content of step 21 above, it can be known that the video recommendation solution provided by the present application may include: first determining a target video with recommendation value; then determining whether the thing described by the target video has been recommended through historical video recall, coarse sorting, fine sorting and large model enhanced recognition, so that when it is determined that it has not been recommended, the target video can be directly recommended. This can effectively overcome the defects caused by repeatedly recommending the same thing, such as affecting user experience.

[0093] Based on the detection method provided in the embodiment of the present application, the embodiment of the present application also provides a detection device. Figure 5 Explain and illustrate. Figure 5 This is a schematic diagram of the structure of a detection device provided in an embodiment of the present application. It should be noted that for the technical details of the detection device provided in an embodiment of the present application, please refer to the relevant content of the detection method above.

[0094] like Figure 5 As shown, the detection device 500 provided in the embodiment of the present application includes:

[0095] A retrieval unit 501 is used to retrieve a resource pool according to at least one feature of the target video after acquiring the target video, and obtain multiple candidate videos, wherein the resource pool is used to record videos that have been recommended in a historical time period;

[0096] A rough selection unit 502 is used to select a plurality of rough selected videos from the plurality of candidate videos according to a similarity calculation result between each of the candidate videos and the target video; the similarity calculation result includes at least one index value, and different index values ​​are used to represent similarities under different dimensions;

[0097] A selection unit 503 is used to select multiple selected videos from the multiple roughly selected videos according to the similarity scores between each of the roughly selected videos and the target video; the similarity scores are obtained using a target model, and the target model is trained using a sample video, a positive example video corresponding to the sample video, and a negative example video corresponding to the sample video, the sample video and the positive example video both describe the same thing, and the sample video and the negative example video respectively describe different things;

[0098] The judgment unit 504 is used to determine, for any of the selected videos, a judgment result between the target video and the selected video using a generative model, wherein the judgment result is used to indicate whether the target video and the selected video describe the same thing.

[0099] In a possible implementation manner, the at least one feature includes: a video vector of the target video, a text vector of a title of the target video, and keywords extracted from the title;

[0100] The retrieval unit 501 is specifically used to: obtain the retrieval results of the video vector, the retrieval results of the text vector, and the retrieval results of the keyword from the resource pool; and determine the multiple candidate videos based on the retrieval results of the video vector, the retrieval results of the text vector, and the retrieval results of the keyword.

[0101] In one possible implementation, for any of the candidate videos, the similarity calculation result between the candidate video and the target video includes at least one of a first index value, a second index value, and a third index value; the first index value is determined based on the similarity between the video vector of the candidate video and the video vector of the target video; the second index value is determined based on the similarity between the title of the candidate video and the title of the target video; the third index value is determined based on the similarity between the text vector of the title of the candidate video and the text vector of the title of the target video.

[0102] In a possible implementation manner, the similarity calculation result includes N index values, where N is a positive integer;

[0103] The rough selection unit 502 is specifically used to: screen the multiple candidate videos according to the nth index value between each of the candidate videos and the target video to obtain a screening result corresponding to the nth index value, where n is a positive integer, n≤N; and determine the multiple rough selected videos according to the screening results corresponding to the N index values.

[0104] In one possible implementation, for any of the roughly selected videos, the process of determining the similarity score between the roughly selected video and the target video includes: inputting the roughly selected video, the title of the roughly selected video, the target video, and the title of the target video into the target model to obtain output data of the target model, wherein the output data includes the similarity score between the roughly selected video and the target video.

[0105] In one possible implementation, the target model includes a text encoding module, a video encoding module, a fully connected layer and a scoring module; the text encoding module is used to obtain the encoding result of the title of the roughly selected video and the encoding result of the title of the target video; the video encoding module is used to obtain the encoding result of the roughly selected video and the encoding result of the target video; the fully connected layer is used to process the splicing result between the encoding result of the title of the roughly selected video and the encoding result of the roughly selected video, and to process the splicing result between the encoding result of the title of the target video and the encoding result of the target video; the scoring module is used to process the output data of the fully connected layer to obtain a similarity score between the roughly selected video and the target video.

[0106] In a possible implementation manner, the detection device 500 further includes:

[0107] A determination unit is used to determine whether to recommend the target video according to the judgment result.

[0108] Based on the relevant contents of the above-mentioned detection device 500, it can be known that the working principle of the device 500 includes: after obtaining the target video, searching the resource pool according to at least one feature of the target video to obtain multiple candidate videos, and the resource pool is used to record the videos that have been recommended in the historical time period; according to the similarity calculation result between each of the candidate videos and the target video, screening out multiple rough selected videos from the multiple candidate videos; the similarity calculation result includes at least one index value, and different index values ​​are used to characterize the similarity under different dimensions; according to the similarity score between each of the rough selected videos and the target video, screening out multiple selected videos from the multiple rough selected videos; the similarity score is obtained using a target model, and the target model is trained using a sample video, a positive example video corresponding to the sample video, and a negative example video corresponding to the sample video, the sample video and the positive example video both describe the same thing, and the sample video and the negative example video respectively describe different things; for any of the selected videos, using a generative model to determine the judgment result between the target video and the selected video, and the judgment result is used to indicate whether the target video and the selected video describe the same thing.

[0109] In addition, an embodiment of the present application also provides an electronic device, which includes a processor and a memory: the memory is used to store instructions or computer programs; the processor is used to execute the instructions or computer programs in the memory, so that the electronic device executes any implementation of the detection method provided in the embodiment of the present application.

[0110] See also Figure 6 , which shows a schematic diagram of the structure of an electronic device 600 suitable for implementing the embodiment of the present disclosure. The terminal device in the embodiment of the present disclosure may include but is not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), vehicle terminals (such as vehicle navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 6 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0111] like Figure 6 As shown, the electronic device 600 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. In the RAM 603, various programs and data required for the operation of the electronic device 600 are also stored. The processing device 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0112] Typically, the following devices may be connected to the I / O interface 605: an input device 606 including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 607 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 608 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 609. The communication device 609 may allow the electronic device 600 to communicate with other devices wirelessly or by wire to exchange data. Although Figure 6 The electronic device 600 is shown with various devices, but it should be understood that it is not required to implement or possess all the devices shown. More or fewer devices may be implemented or possessed instead.

[0113] In particular, according to an embodiment of the present disclosure, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a non-transitory computer-readable medium, and the computer program contains program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication device 609, or installed from the storage device 608, or installed from the ROM 602. When the computer program is executed by the processing device 601, the above-mentioned functions defined in the method of the embodiment of the present disclosure are executed.

[0114] The electronic device provided by the embodiment of the present disclosure and the method provided by the above embodiment belong to the same inventive concept, and the technical details not fully described in this embodiment can be referred to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0115] The embodiment of the present application also provides a computer-readable medium, in which instructions or computer programs are stored. When the instructions or computer programs are executed on a device, the device executes any implementation of the detection method provided in the embodiment of the present application.

[0116] It should be noted that the computer-readable medium disclosed above may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that may be used by or in combination with an instruction execution system, device or device. In the present disclosure, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which a computer-readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. The computer readable signal medium may also be any computer readable medium other than a computer readable storage medium, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0117] In some embodiments, the client and the server may communicate using any currently known or future developed network protocol such as HTTP (Hyper Text Transfer Protocol), and may be interconnected with any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), an internet (e.g., the Internet), and a peer-to-peer network (e.g., an ad hoc peer-to-peer network), as well as any currently known or future developed network.

[0118] The computer-readable medium may be included in the electronic device, or may exist independently without being incorporated into the electronic device.

[0119] The computer-readable medium carries one or more programs, and when the one or more programs are executed by the electronic device, the electronic device can execute the method.

[0120] Computer program code for performing the operations of the present disclosure may be written in one or more programming languages ​​or a combination thereof, including, but not limited to, object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0121] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some implementations as replacements, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0122] The units involved in the embodiments described in the present disclosure may be implemented by software or hardware, wherein the name of a unit / module does not, in some cases, constitute a limitation on the unit itself.

[0123] The functions described above herein may be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0124] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0125] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the system or device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0126] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0127] It should also be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0128] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0129] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A detection method, characterized in that: The method comprises: After acquiring the target video, searching a resource pool according to at least one feature of the target video to obtain a plurality of candidate videos, wherein the resource pool is used to record videos that have been recommended in a historical time period; According to the similarity calculation results between each of the candidate videos and the target video, a plurality of roughly selected videos are screened out from the plurality of candidate videos; the similarity calculation results include at least one index value, and different index values ​​are used to represent similarities under different dimensions; According to the similarity scores between each of the rough selected videos and the target video, a plurality of selected videos are screened out from the plurality of rough selected videos; the similarity scores are obtained using a target model, the target model is trained using a sample video, a positive example video corresponding to the sample video, and a negative example video corresponding to the sample video, the sample video and the positive example video both describe the same thing, and the sample video and the negative example video respectively describe different things; For any of the selected videos, a generative model is used to determine a judgment result between the target video and the selected video, wherein the judgment result is used to indicate whether the target video and the selected video describe the same thing.

2. The method according to claim 1, characterized in that The at least one feature includes: a video vector of the target video, a text vector of a title of the target video, and keywords extracted from the title; The process of determining the multiple candidate videos includes: Acquire the search result of the video vector, the search result of the text vector, and the search result of the keyword from the resource pool; The plurality of candidate videos are determined according to the retrieval results of the video vector, the retrieval results of the text vector, and the retrieval results of the keywords.

3. The method according to claim 1, characterized in that For any of the candidate videos, a similarity calculation result between the candidate video and the target video includes at least one of a first index value, a second index value, and a third index value; The first index value is determined based on the similarity between the video vector of the candidate video and the video vector of the target video; The second index value is determined based on the similarity between the title of the candidate video and the title of the target video; The third index value is determined based on the similarity between the text vector of the title of the candidate video and the text vector of the title of the target video.

4. The method according to claim 1, characterized in that The similarity calculation result includes N index values, where N is a positive integer; The process of determining the plurality of roughly selected videos includes: According to the nth index value between each of the candidate videos and the target video, the plurality of candidate videos are screened to obtain a screening result corresponding to the nth index value, where n is a positive integer, n≤N; The plurality of roughly selected videos are determined according to the screening results corresponding to the N indicator values.

5. The method according to claim 1, characterized in that For any of the roughly selected videos, a process of determining a similarity score between the roughly selected video and the target video includes: The roughly selected video, the title of the roughly selected video, the target video, and the title of the target video are input into the target model to obtain output data of the target model, wherein the output data includes a similarity score between the roughly selected video and the target video.

6. The method according to claim 5, characterized in that The target model includes a text encoding module, a video encoding module, a fully connected layer and a scoring module; The text encoding module is used to obtain the encoding result of the title of the roughly selected video and the encoding result of the title of the target video; The video encoding module is used to obtain the encoding result of the roughly selected video and the encoding result of the target video; The fully connected layer is used to process the splicing result between the encoding result of the title of the roughly selected video and the encoding result of the roughly selected video, and to process the splicing result between the encoding result of the title of the target video and the encoding result of the target video; The scoring module is used to process the output data of the fully connected layer to obtain a similarity score between the roughly selected video and the target video.

7. The method according to any one of claims 1 to 6, characterized in that: The method further comprises: According to the judgment result, it is determined whether to recommend the target video.

8. A detection device, characterized in that: include: A retrieval unit, configured to retrieve a resource pool according to at least one feature of the target video after acquiring the target video, and obtain a plurality of candidate videos, wherein the resource pool is used to record videos that have been recommended in a historical time period; A rough selection unit, configured to select a plurality of rough selected videos from the plurality of candidate videos according to a similarity calculation result between each of the candidate videos and the target video; the similarity calculation result includes at least one index value, and different index values ​​are used to represent similarities under different dimensions; A selection unit, configured to select a plurality of selected videos from the plurality of roughly selected videos according to a similarity score between each of the roughly selected videos and the target video; The similarity score is obtained using a target model, wherein the target model is trained using a sample video, a positive video corresponding to the sample video, and a negative video corresponding to the sample video, wherein the sample video and the positive video both describe the same thing, and the sample video and the negative video respectively describe different things; A judgment unit is used to determine, for any of the selected videos, a judgment result between the target video and the selected video using a generative model, wherein the judgment result is used to indicate whether the target video and the selected video describe the same thing.

9. An electronic device, characterized in that: The device comprises: a processor and a memory; The memory is used to store instructions or computer programs; The processor is used to execute the instructions or computer programs in the memory so that the electronic device executes the method according to any one of claims 1 to 7.

10. A computer-readable medium, characterized in that The computer-readable medium stores instructions or computer programs, and when the instructions or computer programs are executed on a device, the device is enabled to execute the method according to any one of claims 1 to 7.

11. A computer program product, characterized in that It comprises a computer program carried on a non-transitory computer-readable medium, the computer program comprising a program code for executing the method according to any one of claims 1 to 7.

Images