A similar content determination method and apparatus

Abstract

The embodiment of the present application provides a similar content determination method, comprising: obtaining at least one target sub-content corresponding to each content category from target content according to the content category; for each content category, vectorizing at least one target sub-content corresponding to the content category to obtain at least one target sub-content vector corresponding to the content category; determining a multi-modal similarity between a candidate similar content and the target content according to at least one candidate similar sub-content vector corresponding to each content category of the candidate similar content in a candidate similar content set and at least one target sub-content vector corresponding to each content category of the target content; and determining a similar content matched with the target content from the candidate similar content set according to the multi-modal similarity between each candidate similar content in the candidate similar content set and the target content.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of search, in particular to a similar content determination method and device. BACKGROUND

[0002] The current mainstream similarity calculation technology is mostly based on single-dimensional calculation of video similarity, text similarity, picture similarity, etc. In all candidate sets, only whether similar content exists in the target dimension is investigated. If not, it is determined as non-repeated material.

[0003] In the prior art, there are also some multimedia content similarity calculation technologies, but they are also based on graphic and text dimensions. Similarity calculation is performed on graphics and texts respectively. A threshold is set for similar picture pairs. Similar picture pairs reaching a certain number are determined as similar picture content. Then, text content similarity is considered. Video content is judged for similarity separately.

[0004] In the process of implementing the present application, the applicant found at least the following problems in the prior art:

[0005] When performing content similarity calculation, separately calculating the similarity of each similar dimension and separately judging the similarity leads to a high misjudgment rate. SUMMARY

[0006] The embodiments of the present application provide a similar content determination method and device, which is also a multi-modal similar content determination method and device, solving the problem of high misjudgment rate when performing content similarity calculation by separately calculating the similarity of each similar dimension and separately judging the similarity.

[0007] To achieve the above purpose, on the one hand, the embodiments of the present application provide a similar content determination method, comprising:

[0008] Obtaining at least one target sub-content corresponding to each content category from the target content according to the content category;

[0009] For each content category, vectorizing at least one target sub-content corresponding to the content category to obtain at least one target sub-content vector corresponding to the content category;

[0010] According to at least one candidate similar sub-content vector corresponding to each content category of the candidate similar content in the candidate similar content set and at least one target sub-content vector corresponding to each content category of the target content, determining the multi-modal similarity between the candidate similar content and the target content;

[0011] According to the multi-modal similarity between each candidate similar content and the target content in the candidate similar content set, determining similar content matching the target content from the candidate similar content set;

[0012] wherein the candidate similar content set comprises at least one candidate similar content; at least one candidate similar sub-content corresponding to each content category is obtained from each candidate similar content according to the content category, and at least one candidate similar sub-content vector corresponding to each content category is obtained by vectorizing the at least one candidate similar sub-content corresponding to the content category for each content category of each candidate similar content.

[0013] In another aspect, an embodiment of the present application provides a similar content determination apparatus, comprising:

[0014] a target sub-content obtaining unit configured to obtain at least one target sub-content corresponding to each content category from target content according to the content category;

[0015] a target sub-content vectorizing unit configured to vectorize at least one target sub-content corresponding to each content category to obtain at least one target sub-content vector corresponding to the content category for each content category;

[0016] a multi-modal similarity determining unit configured to determine a multi-modal similarity between the candidate similar content and the target content according to at least one candidate similar sub-content vector corresponding to each content category of each candidate similar content in the candidate similar content set and at least one target sub-content vector corresponding to each content category of the target content;

[0017] a similar content determining unit configured to determine a similar content matching the target content from the candidate similar content set according to the multi-modal similarity between each candidate similar content in the candidate similar content set and the target content;

[0018] wherein the candidate similar content set comprises at least one candidate similar content; at least one candidate similar sub-content corresponding to each content category is obtained from each candidate similar content according to the content category, and at least one candidate similar sub-content vector corresponding to the content category is obtained by vectorizing the at least one candidate similar sub-content corresponding to the content category for each content category of each candidate similar content.

[0019] The above technical solution has the following beneficial effects: By calculating the multimodal similarity between each candidate similar content and the target content, the similar content of the target content is further determined. This achieves the determination of similar content by combining multiple similarity dimensions, avoiding the need to determine similar content by comparing each dimension individually with its preset threshold, thus significantly reducing the misjudgment rate of similar content analysis. Specifically, by calculating the similarity scores of each candidate similar content and the target content in each content category, and combining the similarity scores of each candidate similar content in each content category to obtain the overall similarity of each candidate similar content, the similar content of the target content is determined based on the overall similarity. This achieves the determination of similar content by combining multiple similarity dimensions to obtain the overall similarity, avoiding the need to determine similar content by comparing each dimension individually with its preset threshold. This application uses a comprehensive overall similarity to significantly reduce the misjudgment rate of similar content analysis. Specifically, this application obtains the overall similarity by combining text, images, and videos, thus fully reflecting the correlation between the information recorded in the text, images, and videos. Identifying similar content based on the overall similarity obtained thereby significantly reduces the misjudgment rate. Furthermore, before performing similarity judgment, candidate content is filtered to obtain candidate similar content, which significantly reduces the workload of similarity comparison and improves the efficiency of identifying similar content. Furthermore, by limiting the timeliness of candidate content and using target content from the most recent period as candidate content, similar content identification can be adaptively performed to target content that changes over time or with social trends. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of a method for determining similar content according to one embodiment of the present invention;

[0022] Figure 2 This is an architectural diagram of a similar content determination device according to one embodiment of the present invention;

[0023] Figure 3 This is another flowchart of a similar content determination method according to one embodiment of the present invention. Detailed Implementation

[0024] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work belong to the scope of protection of the present application.

[0025] Noun explanation:

[0026] The text generation model adopts the BERT model, which is based on the Transformer architecture released by Google in 2017, and trains the pre-trained deep bidirectional representation by jointly adjusting the bidirectional Transformer in all layers. It is stacked by the encoder of the Transformer. Each encoder contains two sub-layers, the first sub-layer is a multi-head attention layer, which can pay attention to other words when encoding a certain word; the second sub-layer is a forward feedback layer.

[0027] The picture vector generation model adopts the contrast learning technology, obtains a positive sample pair by enhancing the original image, and obtains a 128-dimensional vector through infoNCE loss contrast learning.

[0028] The video vector generation model decodes and frames the video to obtain an image sequence, evenly cuts the sequence to obtain a segment, and extracts a 128-dimensional vector for each segment using contrast learning.

[0029] The Faiss similarity search library is an open source library for clustering and similarity search provided by the Facebook AI team, which provides efficient similarity search and clustering for dense vectors, supports search of billions of vectors, and is the most mature approximate neighbor search library at present.

[0030] The inventor analyzes the prior art and finds the following problems of the prior art:

[0031] When the prior art judges the similarity of multi-picture content, it first judges whether the similarity of each picture group reaches a threshold value, and then judges whether the similar picture groups reach the required number. The existence of two preset values causes the multi-picture content with similar picture groups close to the threshold value to be misjudged as dissimilar;

[0032] When the prior art judges the similarity of video content, it often only considers the video itself and ignores the information amount in the video title and caption. When the title and caption of the video contain a certain amount of information, the information amount of the text consistent with the video content and the information amount of the picture-text content containing multiple key frame screenshots of the video can be almost ignored. The prior art cannot judge the similarity of the two types of content because it lacks processing of the text of the video content.

[0033] The prior art often needs to compare and calculate the target content with the candidate set content when judging whether the target content is repeated, and the size of the candidate set and the resource utilization rate are selected - if the candidate set is too large, too much calculation is wasted on non-repeated content, and the retrieval efficiency of similar content is reduced; if the candidate set is too small, the probability of not containing corresponding similar content increases. At the same time, due to the one-to-one limitation of similarity judgment, it is difficult to better cluster and unify a large number of repeated contents.

[0034] In view of the above problems found by the inventors, the inventors propose the technical solutions of the present application.

[0035] The embodiments of the present application are described below:

[0036] On the one hand, as shown in Figure 1 The present application provides a similar content determination method, comprising:

[0037] Step S11: obtaining at least one target sub-content corresponding to each content category from the target content according to the content category;

[0038] Step S12: vectorizing at least one target sub-content corresponding to each content category to obtain at least one target sub-content vector corresponding to the content category;

[0039] Step S13: determining the multi-modal similarity between the candidate similar content and the target content according to at least one candidate similar sub-content vector corresponding to each content category of the candidate similar content in the candidate similar content set and at least one target sub-content vector corresponding to each content category of the target content;

[0040] Step S14: determining the similar content matching the target content from the candidate similar content set according to the multi-modal similarity between each candidate similar content in the candidate similar content set and the target content;

[0041] The candidate similar content set includes at least one candidate similar content; at least one candidate similar sub-content corresponding to each content category is obtained from each candidate similar content according to the content category, and at least one candidate similar sub-content vector corresponding to each content category is vectorized to obtain at least one candidate similar sub-content vector corresponding to the content category.

[0042] In some embodiments, the target content currently received is to be processed as target content by the embodiments of the present application; the target content determined as similar content in the history before the current processing is historical target content, and the candidate similar content in the candidate similar content set can be pre-constructed according to the project needs, such as various media content or advertising content published by advertisers, and can also come from the historical target content; the content categories of the target content can be established according to the specific characteristics of the target content and the specific similar content identification needs. The target sub-content belonging to each content category is obtained from the target content. The target sub-content of different content categories can be vectorized using the same or different vectorization methods. The candidate similar sub-content is obtained from the candidate similar content according to the same content category as the target content; and the candidate similar sub-content contained in each content category of the candidate similar content is also vectorized using the same vectorization method as the same content category of the target content, so as to obtain the candidate similar sub-content vector corresponding to the candidate similar sub-content contained in each content category of the candidate similar content. After determining the similar content of the target content, the content displayed to the user and the frequency of the target content and the similar content displayed to the user are determined according to the determined target content and the similar content thereof. For example, when the target content has similar content, the similar content may have been displayed in a specific application, and multiple similar content has been displayed. At this time, since the total number of times of displaying the multiple similar content of the target content reaches the preset similar display number, the target content will not be displayed this time.

[0043] The embodiments of the present application have the following technical effects: by calculating the multi-modal similarity of each candidate similar content and the target content corresponding to the content category, the similar content of the target content is further determined, the similar content is determined by combining multiple similar dimensions, and the similar content is determined by sequentially comparing the multiple similar dimensions with the preset threshold of each dimension, thereby significantly reducing the misjudgment rate of similar content analysis.

[0044] Further, the content categories include: text, picture, video;

[0045] The at least one target sub-content vector contained in the content category is vectorized for each content category to obtain at least one target sub-content vector corresponding to the content category, including:

[0046] In the case where the target content includes text, all the text in the target content is vectorized into a corresponding target text vector;

[0047] In the case where the target content includes pictures, at least one picture in the target content is respectively vectorized into at least one target picture vector;

[0048] In a case that the target content comprises a video, the video in the target content is vectorized into a corresponding target video segment vector sequence;

[0049] The target sub-content vector comprises a target text vector, a target picture vector, and / or a target video segment vector sequence.

[0050] In some embodiments, the target content is classified according to text, picture and / or video; the target content is composed of one, two or any combination of more than one of text, picture and / or video; all the text in the target content is taken as a whole, and the text is vectorized into a target text vector; each picture in the target content corresponds to a corresponding target picture vector respectively; the video in the target content is segmented according to a preset video time length to obtain at least one video segment, and each video segment is vectorized to obtain a corresponding target video segment vector, and the target video segment vectors are arranged into a target video segment vector sequence according to the order of the video segments in the original video. There are many vectorization methods, including but not limited to using BERT model, picture vector generation model and video vector generation model to generate a corresponding number of vectors for the text, all the pictures and all the video segments of the target content (such as a microblog containing only text content, a microblog containing multiple pictures and text content, or a microblog containing video content and text content, etc.).

[0051] The embodiments of the present application have the following technical effects: the multi-modal similarity between the candidate similar content and the target content is obtained by combining the text, the picture and the video, thereby fully reflecting the correlation between the information recorded in the text, the picture and the video, and significantly reducing the misjudgment rate.

[0052] Further, the multi-modal similarity between the candidate similar content and the target content is determined according to at least one candidate similar sub-content vector corresponding to each content category of the candidate similar content in the candidate similar content set, and at least one target sub-content vector corresponding to each content category of the target content, and the multi-modal similarity between the candidate similar content and the target content is determined according to at least one candidate similar sub-content vector corresponding to each content category of the candidate similar content in the candidate similar content set, and at least one target sub-content vector corresponding to the same content category of the target content.

[0053] For each candidate similar content in the candidate similar content set, a similarity score corresponding to each content category of the candidate similar content is calculated according to the distance between at least one candidate similar sub-content vector corresponding to each content category of the candidate similar content and at least one target sub-content vector corresponding to the same content category of the target content.

[0054] The similarity scores corresponding to all the content categories of the candidate similar content are taken as the multi-modal similarity between the candidate similar content and the target content.

[0055] In some embodiments, specifically, the similarity score corresponding to each content category of the candidate similar content is determined by the distance between the at least one candidate similar sub-content vector corresponding to the content category of the candidate similar content and the at least one target sub-content vector corresponding to the same content category of the target content, thereby determining the similarity score of the candidate similar content in each content category relative to the target content; for all candidate similar contents in the candidate similar content set, the similarity score of each candidate similar content in each content category relative to the target content is obtained, thereby obtaining the multi-modal similarity between each candidate similar content and the target content.

[0056] Further, the similarity score corresponding to each content category of the candidate similar content is calculated according to the distance between the at least one candidate similar sub-content vector corresponding to the content category of the candidate similar content and the at least one target sub-content vector corresponding to the same content category of the target content for each candidate similar content in the candidate similar content set, including:

[0057] generating a text confidence parameter according to the text properties of all texts in the target content;

[0058] for each candidate similar content, taking the distance between the target text vector and the candidate similar text vector corresponding to all texts of the candidate similar content as a text vector distance, and taking the product of the text vector distance and the text confidence parameter as the similarity score corresponding to the texts in the candidate similar content and the target content;

[0059] wherein the candidate similar text vector is a vector obtained by vectorizing all texts in the candidate similar content. Preferably, the candidate similar text vector is a target text vector corresponding to all texts in the historical target content corresponding to the candidate similar content.

[0060] In some embodiments, the text attribute includes but is not limited to the field to which the text belongs and / or the length of the text, etc. The text confidence parameter can be determined according to the strictness or leniency of the requirements of the target content according to the field to which the text belongs. The more strict the field requirements, the longer the text length, the less accurate the distance calculation, and the larger the text confidence parameter. For example, in the news field, a few hundred word text news, whose main content is in the text, has a confidence parameter of 2, and other fields may have a confidence parameter of 1. For another example, if the text length is less than 60 characters, the text confidence parameter is 0.5, if the text length is less than 100 characters, the text confidence parameter is 1, and if the text length is greater than 100 characters, the text confidence parameter is 2. The specific value of the text confidence parameter can be obtained by debugging or fitting the collected historical data. The similar text vector corresponding to the text of the candidate similar content is obtained using the same vectorization method as the text of the target content. Each candidate similar content with text content corresponds to a similar score corresponding to the text.

[0061] The embodiments of the present application have the following technical effects: By adjusting the distance calculation result of the similar text vector and the target text vector according to the application scenario through the text confidence parameter, the similar score corresponding to the text is more accurate.

[0062] Further, for each candidate similar content in the candidate similar content set, the distance between the at least one candidate similar sub-content vector corresponding to each content category of the candidate similar content and the at least one target sub-content vector corresponding to the same content category of the target content is calculated to obtain a similar score corresponding to the content category of the candidate similar content.

[0063] For each candidate similar content, for each target picture vector corresponding to a picture in the target content, the distance between the target picture vector and each candidate similar picture vector corresponding to a picture in the candidate similar content is calculated, and the minimum distance in the calculated distances is selected as the target picture distance initial value corresponding to the target picture vector. The first picture confidence parameter corresponding to the target picture vector is generated according to the picture attribute of the picture corresponding to the target picture vector, and the product of the target picture distance initial value and the first picture confidence parameter is used as the target picture distance corresponding to the target picture vector.

[0064] For each candidate similar content, for each candidate similar picture vector corresponding to each picture in the candidate similar content, a distance between the candidate similar picture vector and a target picture vector corresponding to each picture in the target content is calculated, and a minimum distance in the calculated distances is selected as a candidate similar picture distance initial value corresponding to the candidate similar picture vector, and a second picture confidence parameter is generated according to a picture attribute of a picture in the target content corresponding to the minimum distance, and a product of the candidate similar picture distance initial value and the second picture confidence parameter is used as a candidate similar picture distance corresponding to the candidate similar picture vector;

[0065] For each candidate similar content, a quantity difference parameter is determined according to a number of pictures in the target content and a number of pictures in the candidate similar content.

[0066] For each candidate similar content, a similarity score corresponding to all pictures of the candidate similar content is calculated according to the following formula (1):

[0067]

[0068] Wherein: pic_score is the similarity score corresponding to all pictures of the candidate similar content; param num is the quantity difference parameter; pic_dis i is a target picture distance corresponding to a target picture vector corresponding to the i th picture in the target content; n is the number of pictures in the target content; recall_pic_dis j is a candidate similar picture distance corresponding to a candidate similar picture vector corresponding to the j th picture in the candidate similar content; m is the number of pictures in the candidate similar content.

[0069] Wherein, the candidate similar picture vector corresponding to each picture in the candidate similar content is a vector obtained by vectorizing the picture in the candidate similar content; preferably, the candidate similar picture vector is a target picture vector corresponding to a picture in a historical target content corresponding to the candidate similar content.

[0070] Further, the calculation of the similarity score corresponding to the content category of the candidate similar content according to the distance between the at least one candidate similar sub-content vector corresponding to each content category of the candidate similar content and the at least one target sub-content vector corresponding to the same content category of the target content includes:

[0071] For each candidate similar content, the distance between each candidate similar video segment vector in the candidate similar video segment vector sequence corresponding to the video in the candidate similar content and each target video segment vector in the target video segment vector sequence corresponding to the video in the target content is calculated; and

[0072] The target video segment vector corresponding to the distance greater than or equal to the video segment similarity threshold value is selected as the similar segment from the calculated distances, and the longest continuous similar segment corresponding to the target content is selected as the similar subject corresponding to the candidate similar content; and

[0073] The maximum value in the calculated distances corresponding to the target video segment vector in the similar subject corresponding to the candidate similar content is taken as the similarity score of the video in the candidate similar content.

[0074] The candidate similar video segment vector sequence corresponding to the video in the candidate similar content is the video segment vector sequence obtained by vectorizing the video in the candidate similar content; preferably, the candidate similar video segment vector sequence is the target video segment vector sequence corresponding to the video in the historical target content corresponding to the candidate similar content.

[0075] Further, the similar content matching the target content is determined from the candidate similar content set according to the multi-modal similarity between each candidate similar content in the candidate similar content set and the target content, comprising:

[0076] For each candidate similar content, the similarity scores corresponding to all content categories of the candidate similar content are weighted and summed to obtain the overall similarity of the candidate similar content.

[0077] The similar content of the target content is determined according to the overall similarity of each candidate similar content and a preset similarity threshold value.

[0078] In some embodiments, the similarity scores corresponding to each of the content categories are comprehensively considered, specifically, for each candidate similar content, the text corresponding similarity score, the picture corresponding similarity score and / or the video corresponding similarity score of the candidate similar content are weighted and biased to obtain the overall similarity of the candidate similar content. The overall similarity reflects the similarity between the candidate similar content and the target content as a whole, and by setting a preset similarity threshold, one or more candidate similar contents most similar to the target content are selected as the similar content of the target content. The weight corresponding to each similarity score can be determined according to the analysis of historical data of a specific industry or obtained by training a training data set constructed by historical data of a specific industry. For example, in microblog, more attention can be paid to text, so that the weight of text is larger, in advertising, more attention can be paid to pictures and videos, so that the weights of pictures and videos are larger, in news, more attention can be paid to text and video, so that the weights of text and video are larger, and so on. The accuracy of determining the similar content is improved by comprehensively considering the similarity scores corresponding to each of the content categories and controlling the weight of each content category through the weight.

[0079] Further, the determining the similar content of the target content according to the overall similarity of each of the candidate similar content and the preset similarity threshold comprises:

[0080] If there is a candidate similar content with an overall similarity less than the preset similarity threshold among all candidate similar contents, the candidate similar content with the overall similarity less than the preset similarity threshold is selected as a preliminary similar content; and the candidate similar content with the minimum overall similarity in the preliminary similar content is selected as the similar content of the target content; otherwise,

[0081] The target content itself is set as the similar content of the target content.

[0082] Further, the target content and the candidate similar content both have an id and a similar feature id attribute;

[0083] The selecting the candidate similar content with the minimum overall similarity in the preliminary similar content as the similar content of the target content comprises:

[0084] The similar feature id of the candidate similar content with the minimum overall similarity in the preliminary similar content is assigned to the similar feature id of the target content.

[0085] The setting the target content itself as the similar content of the target content comprises:

[0086] The id of the target content is assigned to the similar feature id of the target content.

[0087] In some embodiments, by setting the similar feature id of the target content to the similar feature id of the candidate similar content similar to its content or its own id, all candidate similar contents with similar content and the target content can have the same similar feature id, so that all candidate similar contents with similar content and the target content can be found according to the similar feature id, and all similar contents corresponding to the specified similar feature id can be quickly retrieved from all historical target contents according to the specified similar feature id.

[0088] Further, the method for generating the candidate similar content set comprises:

[0089] Setting all candidate sub-content vectors corresponding to all candidate sub-contents corresponding to the same content category of all candidate contents in the candidate content set to the candidate sub-content vector set corresponding to the content category;

[0090] For each candidate sub-content vector set corresponding to a content category, using Faiss to cluster the candidate sub-content vector set corresponding to the content category to obtain at least one cluster center of the candidate sub-content vector set corresponding to the content category;

[0091] For each target sub-content vector corresponding to a content category of the target content, calculating the distance value between the target sub-content vector and each of the at least one cluster center of the candidate sub-content vector set corresponding to the same content category, and selecting the K cluster centers with the smallest distance value as the K cluster centers corresponding to the target sub-content vector; and,

[0092] For each cluster center in the K cluster centers corresponding to the target sub-content vector, calculating the distance between each candidate sub-content vector in the cluster corresponding to the cluster center and the target sub-content vector, and taking the candidate content corresponding to the candidate sub-content vector corresponding to the distance meeting the specified recall condition among the calculated distances as the candidate similar content;

[0093] Wherein, the candidate content is a historical target content that has been determined to be similar; and the at least one candidate sub-content vector corresponding to each content category of the candidate content is at least one target sub-content vector corresponding to the same content category of the historical target content.

[0094] The candidate sub-content vector set corresponding to the content category is composed of candidate sub-content vectors corresponding to all candidate sub-contents corresponding to the same content category of all candidate contents.

[0095] In some embodiments, the generation of the cluster centers is generated by the Faiss reference kmeans clustering algorithm. For example, 1000 candidate picture vectors are set, a total of 100 cluster centers are generated, 10 cluster centers closest to the target picture vector are selected, 10 is K, the determination of the value of K depends on the trade-off between retrieval efficiency and retrieval effect, but K must be less than or equal to the total number of cluster centers, and the total number of cluster centers must be less than or equal to the number of candidate picture vectors; and in the process of selecting K cluster centers corresponding to each target sub-content vector, the target sub-content vectors have no influence on each other, for example, there are 2 pictures under the target content, each picture is determined to correspond to K cluster centers independently, for example, picture 1 and 100 cluster centers are calculated to find the nearest K cluster centers corresponding to picture 1; picture 2 and 100 cluster centers are calculated to find the nearest K cluster centers corresponding to picture 2, and the target pictures do not affect each other

[0096] As time goes by, a large number of historical target contents will be accumulated to form a large candidate content set. By determining the most relevant cluster of the target content and further determining the most relevant candidate similar content from the most relevant cluster, the number of similar content comparisons can be significantly reduced to improve the efficiency while ensuring the accuracy of the comparison.

[0097] Further, the method further comprises:

[0098] The target content for which the similar content is determined is added to the candidate content set as a candidate content.

[0099] The candidate content in the candidate content set that is more than a preset time length from the current time is removed from the candidate content set.

[0100] In some embodiments, the candidate content is updated by a preset time length of the nearest time. In some content change trend scenarios, such as the change of some hot topics in micro blogs, the candidate content can follow the latest trend.

[0101] On the other hand, as Figure 2 shown, the embodiment of the present application provides a similar content determination device, comprising:

[0102] A target sub-content acquisition unit 21 is configured to acquire at least one target sub-content corresponding to each content category from the target content according to the content category.

[0103] A target sub-content vectorization unit 22 is configured to vectorize at least one target sub-content corresponding to each content category to obtain at least one target sub-content vector corresponding to the content category.

[0104] The multi-modal similarity determination unit 23 is configured to determine a multi-modal similarity between the candidate similar content and the target content according to at least one candidate similar sub-content vector corresponding to each content category of the candidate similar content in the candidate similar content set and at least one target sub-content vector corresponding to each content category of the target content.

[0105] The similar content determination unit 24 is configured to determine a similar content matched with the target content from the candidate similar content set according to the multi-modal similarity between each candidate similar content in the candidate similar content set and the target content.

[0106] The candidate similar content set includes at least one candidate similar content, at least one candidate similar sub-content corresponding to each content category is obtained from each candidate similar content according to the content category, and at least one candidate similar sub-content vector corresponding to the content category is obtained by vectorizing at least one candidate similar sub-content corresponding to the content category of each candidate similar content.

[0107] Further, the content category includes text, picture, and video.

[0108] The target sub-content vectorization unit 22 includes:

[0109] The text vectorization module is configured to vectorize all text in the target content into a corresponding target text vector when the target content includes text.

[0110] The picture vectorization module is configured to correspondingly vectorize at least one picture in the target content into at least one target picture vector when the target content includes picture; and / or,

[0111] The video vectorization module is configured to vectorize a video in the target content into a corresponding target video segment vector sequence when the target content includes video.

[0112] The target sub-content vector includes a target text vector, a target picture vector, and / or a target video segment vector sequence.

[0113] Further, the multi-modal similarity determination unit 23 includes:

[0114] The similarity score determination module is configured to calculate a similarity score corresponding to each content category of the candidate similar content according to a distance between at least one candidate similar sub-content vector corresponding to each content category of the candidate similar content and at least one target sub-content vector corresponding to the same content category of the target content for each candidate similar content in the candidate similar content set.

[0115] a multi-modal similarity determination module configured to determine a multi-modal similarity between the candidate similar content and the target content by taking the similarity score corresponding to each content category of the candidate similar content as the multi-modal similarity.

[0116] Further, the similarity score determination module comprises:

[0117] a text confidence parameter generation module configured to generate a text confidence parameter according to text attributes of all texts in the target content;

[0118] a text similarity score calculation module configured to, for each candidate similar content, take a distance between the target text vector and a candidate similar text vector corresponding to all texts in the candidate similar content as a text vector distance, and take a product of the text vector distance and the text confidence parameter as a similarity score corresponding to texts in the candidate similar content and the target content;

[0119] wherein the candidate similar text vector is a vector obtained by vectorizing all texts in the candidate similar content, and preferably the candidate similar text vector is a target text vector corresponding to all texts in a historical target content corresponding to the candidate similar content.

[0120] Further, the similarity score determination module comprises:

[0121] a target picture distance calculation module configured to, for each candidate similar content, calculate a distance between a target picture vector corresponding to each picture in the target content and a candidate similar picture vector corresponding to each picture in the candidate similar content, select a minimum distance among the calculated distances as a target picture distance initial value corresponding to the target picture vector, generate a first picture confidence parameter corresponding to the target picture vector according to a picture attribute of the picture corresponding to the target picture vector, and use a product of the target picture distance initial value and the first picture confidence parameter as a target picture distance corresponding to the target picture vector.

[0122] a candidate similar picture distance calculation module configured to, for each candidate similar content, calculate a distance between a candidate similar picture vector corresponding to each picture in the candidate similar content and a target picture vector corresponding to each picture in the target content, select a minimum distance among the calculated distances as a candidate similar picture distance initial value corresponding to the candidate similar picture vector, generate a second picture confidence parameter according to a picture attribute of the picture in the target content corresponding to the minimum distance, and use a product of the candidate similar picture distance initial value and the second picture confidence parameter as a candidate similar picture distance corresponding to the candidate similar picture vector.

[0123] a quantity difference parameter determination module configured to determine, for each candidate similar content, a quantity difference parameter according to a number of pictures of the target content and a number of pictures of the candidate similar content;

[0124] a picture score calculation module configured to calculate, for each candidate similar content, a similarity score corresponding to all pictures of the candidate similar content according to the following formula (1);

[0125] wherein a candidate similar picture vector corresponding to each picture in the candidate similar content is a vector obtained by vectorizing the picture in the candidate similar content; preferably, the candidate similar picture vector is a target picture vector corresponding to a picture in the historical target content corresponding to the candidate similar content.

[0126] Further, the similarity score determination module comprises:

[0127] a video segment distance calculation module configured to calculate, for each candidate similar video segment vector sequence corresponding to a video in each candidate similar content, a distance between each candidate similar video segment vector in the candidate similar video segment vector sequence and each target video segment vector in a target video segment vector sequence corresponding to a video in the target content; and,

[0128] a similar subject determination module configured to select, from the calculated distances, a target video segment vector corresponding to a distance greater than or equal to a video segment similarity threshold as a similar segment, and select a longest continuous similar segment corresponding to the target content as a similar subject corresponding to the candidate similar content; and,

[0129] a video score determination module configured to take a maximum value in the calculated distances corresponding to target video segment vectors in the similar subject corresponding to the candidate similar content as a similarity score of a video in the candidate similar content;

[0130] wherein a candidate similar video segment vector sequence corresponding to a video in the candidate similar content is a video segment vector sequence obtained by vectorizing the video in the candidate similar content; preferably, the candidate similar video segment vector sequence is a target video segment vector sequence corresponding to a video in the historical target content corresponding to the candidate similar content.

[0131] Further, the similar content determination unit 24 comprises:

[0132] an overall similarity determination module configured to, for each candidate similar content, weight and sum similarity scores corresponding to all content categories of the candidate similar content to obtain an overall similarity of the candidate similar content;

[0133] The similarity comparison determines a similar content module for determining the similar content of the target content according to the overall similarity of each candidate similar content and a preset similarity threshold.

[0134] Further, the similarity comparison determines a similar content module, comprising:

[0135] The similarity comparison module is arranged to trigger the first similar content determination module if there is a candidate similar content with an overall similarity less than the preset similarity threshold among all candidate similar contents, otherwise trigger the second similar content determination module.

[0136] The first similar content determination module is arranged to determine the candidate similar content with an overall similarity less than the preset similarity threshold as a preliminary similar content, and determine the candidate similar content with the minimum overall similarity in the preliminary similar content as the similar content of the target content.

[0137] The second similar content determination module is arranged to set the target content itself as the similar content of the target content.

[0138] Further, the target content and the candidate similar content both have an id and a similar feature id attribute.

[0139] The first similar content determination module is further arranged to assign the similar feature id of the candidate similar content with the minimum overall similarity in the preliminary similar content to the similar feature id of the target content.

[0140] The second similar content determination module is further arranged to assign the id of the target content to the similar feature id of the target content.

[0141] Further, the device further comprises a candidate similar content set generation unit.

[0142] The candidate similar content set generation unit comprises:

[0143] The candidate sub-content vector set generation module is arranged to set all candidate sub-content vectors corresponding to all candidate sub-contents of the same content category of all candidate contents in the candidate content set as the candidate sub-content vector set corresponding to the content category.

[0144] The cluster center determination module is arranged to utilize Faiss to cluster the candidate sub-content vector set corresponding to each content category to obtain at least one cluster center of the candidate sub-content vector set corresponding to the content category.

[0145] The clustering center selection module is configured to, for at least one target sub-content vector corresponding to each content category of the target content, calculate a distance value between the target sub-content vector and each of at least one clustering center of a candidate sub-content vector set corresponding to the same content category, and select K clustering centers with minimum distance values as K clustering centers corresponding to the target sub-content vector; and

[0146] The candidate similar content determination module is configured to, for each of the K clustering centers corresponding to the target sub-content vector, calculate a distance between each candidate sub-content vector in a cluster corresponding to the clustering center and the target sub-content vector, and determine a candidate content corresponding to a candidate sub-content vector corresponding to a distance meeting a specified recall condition as a candidate similar content.

[0147] The candidate content is a historical target content that has been determined to be similar content; and at least one candidate sub-content vector corresponding to each content category of the candidate content is at least one target sub-content vector corresponding to the same content category of the historical target content.

[0148] The candidate sub-content vector set corresponding to the content category is composed of candidate sub-content vectors corresponding to all candidate sub-content vectors corresponding to the same content category of all candidate contents.

[0149] Further, the apparatus further comprises:

[0150] The first candidate content set update module is configured to add the target content that has been determined to be similar content to the candidate content set as a candidate content.

[0151] The second candidate content set update module is configured to remove, from the candidate content set, a candidate content in the candidate content set that is away from a current time by more than a preset time length.

[0152] The embodiments of this invention have the following technical effects: By calculating the multimodal similarity between each candidate similar content and the target content, the similar content of the target content is further determined. This achieves the determination of similar content by combining multiple similarity dimensions, avoiding the need to determine similar content by comparing each dimension individually with a preset threshold, thereby significantly reducing the misjudgment rate of similar content analysis. Specifically, by calculating the similarity scores of each candidate similar content and the target content in each content category, and combining the similarity scores of each content category of the candidate similar content to obtain the overall similarity of each candidate similar content, the similar content of the target content is determined based on the overall similarity. This achieves the determination of similar content by combining multiple similarity dimensions to obtain an overall similarity, avoiding the need to determine similar content by comparing each dimension individually with a preset threshold. This application uses a comprehensive overall similarity to significantly reduce the misjudgment rate of similar content analysis. Specifically, this application obtains an overall similarity by combining text, images, and videos, thereby fully reflecting the correlation between the information recorded in the text, images, and videos. Identifying similar content based on the overall similarity obtained thereby significantly reduces the misjudgment rate. Furthermore, before performing similarity judgment, candidate content is filtered to obtain candidate similar content, which significantly reduces the workload of similarity comparison and improves the efficiency of identifying similar content. Furthermore, by limiting the timeliness of candidate content and using target content from the most recent period as candidate content, similar content identification can be adaptively performed to target content that changes over time or with social trends.

[0153] The technical solutions of the present invention will be described in detail below with reference to specific application examples. For technical details not described in the implementation process, please refer to the relevant descriptions above.

[0154] like Figure 3 As shown, the embodiments of the present invention will be described below:

[0155] Vector generation:

[0156] The BERT model, the picture vector generation model, and the video vector generation model are used to generate vectors corresponding to the text, all pictures, and all video segments of the target content (such as a microblog containing only text content, a microblog containing multiple pictures and text content, or a microblog containing video content and text content). Three candidate sets are generated based on the text, all pictures, and all video segment vectors of the historical target content. The target content is the target content currently being calculated for similar content, and the historical target content is the target content that has been calculated for similar content in the past. The candidate content set is formed by the historical target content, and the text, all pictures, and / or all video segments contained in each candidate content are vectorized into vectors and saved in the corresponding vector candidate sets of the text, all pictures, and / or video segments. The candidate content and the corresponding vectors of the text, all pictures, and / or video segments it contains are associated with each other, and the corresponding vectors of the text, all pictures, and / or video segments it contains can be retrieved by the candidate content, and the corresponding candidate content can also be retrieved by the corresponding vectors of the text, all pictures, and / or video segments it contains.

[0157] Similar vector recall:

[0158] For the three candidate sets, the indexivfflat function in Faiss is used to cluster the candidate set vectors, generating a corresponding number of cluster centers. The K nearest cluster centers are selected, and the distance between the target vector and all vectors under the K cluster centers is calculated (Euclidean distance for images and text, and cosine distance for video segments). For distances that meet the recall conditions (less than the first preset distance threshold for images and text, and greater than or equal to the second preset distance for video segments), they are considered as candidate similar vectors of the target vector. The candidate similar vectors of the target vector, the corresponding historical content, and the corresponding distance are recorded.

[0159] Text, picture, and video similarity determination:

[0160] Text similarity text_score determination:

[0161] The text confidence parameter text_param is generated mainly based on the distance text_dis between the target text vector and the candidate similar vector, and with reference to the text length of the target content; the text confidence parameter can be determined strictly or loosely according to the field, the more strict the field requirement is (for example, the text of news is several hundred words, and the main content is in the text, the confidence parameter is set to 2, and the confidence parameter of other fields can be 1), the longer the text length is, the less accurate the distance calculation is, and the larger the text confidence parameter is; for example, if the text length is less than 60, the text confidence parameter is 0.5, if the text length is less than 100, the confidence parameter is 1, and if the text length is greater than 100, the text confidence parameter is 2; in a specific project, the text confidence parameter can be obtained by debugging historical data. The text corresponding score of each candidate similar content is determined according to the following formula (2):

[0162] text_score = text_dis * text_param (2)

[0163] Wherein, text_score is the text corresponding score of the candidate similar content; text_dis is the distance between the candidate similar text vector corresponding to the candidate similar content and the target text vector; text_param is the text confidence parameter determined according to the text attribute of the target content;

[0164] According to the above method, each candidate similar content corresponds to a text corresponding similarity score.

[0165] Determination of picture similarity pic_score:

[0166] For each picture under the target content, the distance pic_dis between the target picture vector corresponding to the picture under the target content and the candidate similar picture vector corresponding to each picture under the candidate similar content is calculated i , and the first picture confidence parameter pic_param is determined according to the picture type (long picture, word picture, expression package picture, etc.) of the picture under the target content i (for example, the first picture confidence parameter is 1 under normal circumstances, and is set to 5 to 10 when the picture under the target content is a long picture, a word picture or an expression package), and the distance pic_dis i is corrected using the first picture confidence parameter as formula (3):

[0167] pic_dis i = pic_dis i * pic_param i (3)

[0168] Through the above process, a group of pic_dis iThe distance value, i.e., the target picture distance, corresponds to each picture under the target content; assuming that the number of pictures in the target content is n, i takes a value from 1 to n;

[0169] The distance calculation described above is also performed on each picture under the candidate similar content and each picture under the target content, and the distance recall_pic_dis closest to the picture is recorded j For one candidate similar content, a group of recall_pic_dis j The distance value, i.e., the candidate similar picture distance, corresponds to each picture under the candidate similar content; assuming that the number of pictures in the candidate similar content is m, j takes a value from 1 to m;

[0170] The quantity difference parameter param is determined according to the number n of pictures of the target content and the number m of pictures of the candidate similar content num The quantity difference parameter can be fitted using historical data, and can be corrected according to the running effect in combination with the requirements of specific application scenarios, for example, in some scenarios, if the number of pictures of the target content is less than the number of pictures of the candidate content, the quantity difference parameter is set to 0.7, otherwise, it is set to 0.3.

[0171] The picture similarity pic_score of the candidate similar content is calculated as shown in formula (1).

[0172] According to the above method, each candidate similar content corresponds to a picture corresponding similarity score.

[0173] Determination of video similarity video_score:

[0174] The video segments of the target content and the candidate similar content are compared with the collected video systematized beginning and end segment library, and the identified systematized video beginning and end segments are deleted; the cosine distance of the vectors of the remaining video segments of the target content and the video segments of the candidate similar content is calculated, and when the distance video_dis is greater than or equal to the video segment similarity threshold, the two segments are considered to be similar segments; the longest continuous similar segment in the target content is selected as the similar main body of the target content and the candidate similar content; the maximum distance pic_dis in the video_dis corresponding to all video segments of the similar main body max , i.e., the video similarity, is the video corresponding similarity score of the candidate content, i.e., the value of pic_dis max is taken as the value of video_score.

[0175] The text, picture, and video similarity are fused to obtain the overall similarity final_score:

[0176] Determine the content type of the target content, such as pure picture, picture-text, video-text, etc. According to the target content type and the above-mentioned text, picture, video similarity, determine the text weight xtext, picture weight xpic, video weight xvideo and compensation parameter b of each type of content in the participation of fusion calculation; The text weight, picture weight, video weight and compensation parameter can be fitted according to the historical data in the specific application scene.

[0177] According to the text length, picture type (long picture, word picture, emoticon picture, etc.), the difference between the number of target content and candidate similar content pictures, whether it belongs to certain special field, whether it has field intersection with the candidate similar target, etc. Adjust the weight proportion xtext, xpic, xvideo and compensation parameter b; For example, in some fields, adjust the weight according to the importance, the text is more important in the news field, take out the data of the news field to adjust the weight, use test data to verify whether the similar text can be calculated out, so as to determine the text weight, picture weight, video weight and compensation parameter.

[0178] The overall similarity final_score is the weighted value of the similarity (or similarity score) of each content type, as shown in formula 4:

[0179] final_score = xtext*text_score + xpic*pic_score + xvideo*(1-video_score) + b (4)

[0180] Wherein, final_score is the overall similarity of the candidate similar content; text_score is the similarity score corresponding to the text of the candidate similar content; pic_score is the similarity score corresponding to the picture of the candidate similar content; video_score is the similarity score corresponding to the video of the candidate similar content; xtext is the text weight; xpic is the picture weight; xvideo is the video weight; b is the compensation parameter.

[0181] Determination and clustering of repeated content

[0182] Screening of candidate similar content of target content, for the candidate similar content with overall similarity final_score less than the preset similarity threshold, it is considered that the target content is similar content, if there is no similar content that meets the condition, it is considered that the target content is original content, and the similar feature is the id of the target content itself; For the target content with similar content, sort all candidate similar contents of the target content according to the overall similarity final_score, take the similar feature id of the candidate similar content with the minimum final_score as the similar feature id of the target content.

[0183] Any two contents with the same similar feature id are considered as duplicate contents.

[0184] Expiry of candidate set

[0185] For historical target content, when the historical content and the target content retrieval time are greater than the preset time t, the text, picture and video vector corresponding to the historical content should be expired from the candidate content set, so as to ensure that the candidate content set is always maintained in a stable order, and the accuracy of clustering effect and the efficiency of information processing are guaranteed.

[0186] The technical method has the following technical effects: the prior art often ignores the text information contained in the video content when judging the video content, resulting in that part of the text information-based video content cannot be correctly similar to the text or graphic content. The application introduces a text similarity distance for video-text content in the overall similarity judgment stage, fills the gap in the similarity judgment of text content and video content, and improves the accuracy of target content in similarity judgment.

[0187] Further, the prior art often faces the trade-off between effect and efficiency when judging whether there is similar content - too large candidate content set will reduce the search efficiency and also waste computing resources; too small candidate content set cannot effectively contain similar content, resulting in incorrect judgment. The application effectively utilizes the vector retrieval tool Faiss and the function of similar content similar feature id transmission, uses the function of Faiss clustering and matching based on similar clustering center, expands the candidate content set in an appropriate amount, ensures the accuracy of the search result; at the same time, the function of candidate content set expiration reduces unnecessary computing amount, improves the search efficiency; at the same time, the function of similar content similar feature id transmission ensures that even if the original content no longer exists in the candidate content set, the target content can also match the original content based on the similar content in the candidate set; at the same time, based on this function, when there are multiple similar similar contents, the feature similar id can successfully judge that the contents are similar to each other.

[0188] Further, the prior art is too rigid in requiring a certain number of similar picture groups when determining similarity for multi-picture content, resulting in some similar content that may only exist in a small difference above the threshold on a partial picture group, but due to the strict requirement of the number of similar picture groups, is misjudged as not similar. The present application breaks the concept of similar picture groups when calculating picture similarity, only focusing on the minimum distance of the picture to all pictures in the candidate similar content, and in the final calculation of picture similarity, the misjudgment risk caused by a small number of threshold critical pictures is shared by average calculation; At the same time, the number difference parameter is introduced, which can weight the reliability of the picture similarity when there is a difference in the number of pictures between the target content and the candidate similar content.

[0189] In summary, the present application introduces text content to improve the accuracy of similarity judgment when judging video content similarity; using Faiss, candidate set expiration function, similar content similar feature id transmission function to efficiently and accurately cluster large-scale, long-time, large amount of repeated content, improve efficiency on the basis of maintaining effect; For improving the effect of multi-picture content similarity judgment.

[0190] It should be understood that the specific order or hierarchy of steps in the processes disclosed is an example. Based upon design preferences, it should be understood that the specific order or hierarchy of steps in the processes can be re-arranged while remaining within the scope of the present disclosure. The accompanying method claims present elements of the various steps in a sample order, and are not meant to be limited to the specific order or hierarchy presented.

[0191] In the above detailed description, various features are grouped together in single embodiments for the purpose of streamlining the disclosure. Such disclosed methods should not be interpreted as reflecting an intention that the claimed embodiments require more features than are explicitly recited in each claim. On the contrary, as indicated above, the inventiveness lies in less than all features of the disclosed single embodiments. Accordingly, the claims are hereby expressly incorporated into this detailed description, with each claim standing on its own as a separate preferred embodiment of the present application.

[0192] In order for any person skilled in the art to implement or use the present application, the above discloses the disclosed embodiments. Various modifications of these embodiments are obvious to those skilled in the art, and the general principles defined herein can also be applied to other embodiments without departing from the spirit and protection scope of the present disclosure. Therefore, the present disclosure is not limited to the embodiments given herein, but is consistent with the broadest scope of the principles and novel features disclosed in the present disclosure.

[0193] The above description includes examples of one or more embodiments. Of course, not all possible combinations of components or methods described above can be implemented to make an embodiment, but a person of ordinary skill in the art will recognize from the above description that a variety of further combinations and permutations of the embodiments described above are possible. Accordingly, the embodiments described above are intended to embrace all such alterations, modifications and variations going to the scope of the appended claims. Further, in the description of the specification or claims, the phrase "comprising" includes the embodiments in which additional steps or components are "included in" or "consist of" the described embodiments. Additionally, the use of "or" in the claims is meant to encompass both "and" and "or" unless specifically indicated otherwise.

[0194] Those of skill would further appreciate that the various illustrative logical blocks, modules, and steps described in connection with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or combinations of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans can implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present embodiments.

[0195] The various illustrative logical blocks, modules, and steps described in connection with the embodiments disclosed herein can be implemented or performed with a general purpose processor, a digital signal processor, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general purpose processor can be a microprocessor, but in the alternative, the general purpose processor can be any conventional processor, controller, microcontroller, or state machine. A processor can also be implemented as a combination of computing devices, e.g., a combination of a digital signal processor and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a digital signal processor core, or any other such configuration.

[0196] The steps of a method or algorithm described in connection with the embodiments disclosed herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module can reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor such the processor can read information from, and write information to, the storage medium. In the alternative, the storage medium can be integral to the processor. The processor and the storage medium can reside in an ASIC. The ASIC can reside in a user terminal. In the alternative, the processor and the storage medium can reside as discrete components in a user terminal.

[0197] In one or more exemplary designs, the functions described can be implemented in hardware, software, firmware, or any combination thereof. If implemented in software, the functions can be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media include both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. Storage media can be any available media that can be accessed by a general purpose or special purpose computer. By way of example, such computer-readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code means in the form of instructions or data structures and that can be accessed by a general-purpose or special-purpose computer, or a general-purpose or special-purpose processor. Also, any connection is properly termed a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or data

[0198] The above detailed description describes the purpose, technical solutions and advantages of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the scope of the present application.

Claims

1. A similar content determination method characterized by comprising: The method comprises the following steps: obtaining at least one target sub-content corresponding to each content category of the target content according to the content category; vectorizing at least one target sub-content corresponding to each content category to obtain at least one target sub-content vector corresponding to the content category; determining the multi-modal similarity between the candidate similar content and the target content according to at least one candidate similar sub-content vector corresponding to each content category of the candidate similar content and at least one target sub-content vector corresponding to each content category of the target content; determining the similar content matching the target content from the candidate similar content set according to the multi-modal similarity between each candidate similar content and the target content in the candidate similar content set; wherein the candidate similar content set comprises at least one candidate similar content; at least one candidate similar sub-content corresponding to each content category of each candidate similar content is obtained according to the content category; and at least one candidate similar sub-content vector corresponding to each content category is obtained by vectorizing at least one candidate similar sub-content corresponding to the content category for each content category of each candidate similar content; The method for generating the candidate similar content set comprises the following steps: setting all candidate sub-content vectors corresponding to all candidate sub-contents corresponding to the same content category of all candidate contents in the candidate content set as a candidate sub-content vector set corresponding to the content category; performing clustering on the candidate sub-content vector set corresponding to the content category by using Faiss to obtain at least one cluster center of the candidate sub-content vector set corresponding to the content category for each content category; calculating the distance value between the target sub-content vector and each cluster center of the at least one cluster center of the candidate sub-content vector set corresponding to the same content category for each content category of the target content, and selecting K cluster centers with the smallest distance value as K cluster centers corresponding to the target sub-content vector; and calculating the distance between each candidate sub-content vector in the cluster corresponding to each cluster center and the target sub-content vector for each cluster center in the K cluster centers corresponding to the target sub-content vector, and selecting the candidate content corresponding to the candidate sub-content vector corresponding to the distance meeting the specified recall condition from the calculated distances as the candidate similar content; wherein the candidate content is a historical target content of the determined similar content; and at least one candidate sub-content vector corresponding to each content category of the candidate content is at least one target sub-content vector corresponding to the same content category of the historical target content; The candidate sub-content vector set corresponding to the content category is composed of candidate sub-content vectors corresponding to all candidate sub-contents corresponding to the same content category of all candidate contents.

2. The similar content determination method of claim 1, wherein The content category comprises text, picture and video. The at least one target sub-content corresponding to each content category of the target content is vectorized to obtain at least one target sub-content vector corresponding to the content category, including: In the case where the target content includes text, all text in the target content is vectorized into a corresponding target text vector; In the case where the target content includes pictures, at least one picture in the target content is respectively vectorized into at least one target picture vector; In the case where the target content includes video, the video in the target content is vectorized into a corresponding target video segment vector sequence; The target sub-content vector includes a target text vector, a target picture vector, and / or a target video segment vector sequence.

3. The similar content determination method of claim 2, wherein The multi-modal similarity between the candidate similar content and the target content is determined according to at least one candidate similar sub-content vector corresponding to each content category of the candidate similar content in the candidate similar content set and at least one target sub-content vector corresponding to each content category of the target content, including: For each candidate similar content in the candidate similar content set, a similarity score corresponding to each content category of the candidate similar content is calculated according to the distance between at least one candidate similar sub-content vector corresponding to the content category of the candidate similar content and at least one target sub-content vector corresponding to the same content category of the target content; The similarity scores corresponding to all content categories of the candidate similar content are taken as the multi-modal similarity between the candidate similar content and the target content.

4. The similar content determination method according to Claim 3, characterized by, The similarity score corresponding to each content category of the candidate similar content is calculated according to the distance between at least one candidate similar sub-content vector corresponding to the content category of the candidate similar content and at least one target sub-content vector corresponding to the same content category of the target content, including: A text confidence parameter is generated according to the text attribute of all text in the target content; For each candidate similar content, the distance between the target text vector and a candidate similar text vector corresponding to all text of the candidate similar content is taken as a text vector distance, and the product of the text vector distance and the text confidence parameter is taken as a similarity score corresponding to the text in the candidate similar content and the target content; The candidate similar text vector is obtained by vectorizing all text in the candidate similar content.

5. The similar content determination method of claim 3, wherein The similarity score corresponding to each content category of the candidate similar content is calculated according to the distance between at least one candidate similar sub-content vector corresponding to the content category of the candidate similar content and at least one target sub-content vector corresponding to the same content category of the target content, including: For each candidate similar content, for each target picture vector corresponding to a picture in the target content, a distance between the target picture vector and a candidate similar picture vector corresponding to a picture in the candidate similar content is calculated, and a minimum distance among the calculated distances is selected as a target picture distance initial value corresponding to the target picture vector, and a first picture confidence parameter corresponding to the target picture vector is generated according to a picture attribute of the picture in the target content corresponding to the minimum distance, and a product of the target picture distance initial value and the first picture confidence parameter is used as a target picture distance corresponding to the target picture vector; For each candidate similar content, for each candidate similar picture vector corresponding to a picture in the candidate similar content, a distance between the candidate similar picture vector and a target picture vector corresponding to a picture in the target content is calculated, and a minimum distance among the calculated distances is selected as a candidate similar picture distance initial value corresponding to the candidate similar picture vector, and a second picture confidence parameter is generated according to a picture attribute of the picture in the target content corresponding to the minimum distance, and a product of the candidate similar picture distance initial value and the second picture confidence parameter is used as a candidate similar picture distance corresponding to the candidate similar picture vector; For each candidate similar content, a quantity difference parameter is determined according to a number of pictures of the target content and a number of pictures of the candidate similar content; For each candidate similar content, a similarity score corresponding to all pictures of the candidate similar content is calculated according to the following formula: wherein: is a similarity score corresponding to all pictures of the candidate similar content as a whole; is a quantity difference parameter; is a target picture distance corresponding to a target picture vector corresponding to the i-th picture in the target content; n is the number of pictures in the target content; is a candidate similar picture distance corresponding to a candidate similar picture vector corresponding to the j-th picture in the candidate similar content; m is the number of pictures in the candidate similar content; Wherein, a candidate similar picture vector corresponding to each picture in the candidate similar content is a vector obtained by vectorizing the picture in the candidate similar content.

6. The similar content determination method of claim 3, wherein, The calculating, for each candidate similar content in the set of candidate similar contents, a similarity score corresponding to each content category of the candidate similar content according to distances between at least one candidate similar sub-content vector corresponding to the content category of the candidate similar content and at least one target sub-content vector corresponding to a same content category of the target content, comprises: For each candidate similar video segment vector sequence corresponding to a video in each candidate similar content, a distance between each candidate similar video segment vector in the candidate similar video segment vector sequence and each target video segment vector in a target video segment vector sequence corresponding to a video in the target content is calculated; and, From the calculated distances, a target video segment vector corresponding to a distance greater than or equal to a video segment similarity threshold value is selected as a similar segment, and a longest continuous similar segment corresponding to the target content is selected as a similar main body corresponding to the candidate similar content; and, A maximum value among the calculated distances corresponding to target video segment vectors in the similar main body corresponding to the candidate similar content is used as a similarity score of a video in the candidate similar content; Wherein, a candidate similar video segment vector sequence corresponding to a video in the candidate similar content is a video segment vector sequence obtained by vectorizing the video in the candidate similar content.

7. The similar content determination method of claim 3, wherein The determining the similar content matched with the target content from the candidate similar content set according to the multi-modal similarity between each of the candidate similar content and the target content comprises: For each candidate similar content, the similar scores of all content categories of the candidate similar content are weighted and summed to obtain an overall similarity of the candidate similar content; The similar content of the target content is determined according to the overall similarity of each of the candidate similar content and a preset similarity threshold.

8. The similar content determination method of claim 7, wherein, The determining the similar content of the target content according to the overall similarity of each of the candidate similar content and a preset similarity threshold comprises: If there is a candidate similar content with an overall similarity less than the preset similarity threshold in all candidate similar contents, the candidate similar content with the overall similarity less than the preset similarity threshold is taken as a preliminary selected similar content; and the candidate similar content with the minimum overall similarity in the preliminary selected similar content is taken as the similar content of the target content; otherwise, The target content itself is set as the similar content of the target content.

9. A similar content determination apparatus characterized by comprising: Comprise: A target sub-content acquisition unit is configured to acquire at least one target sub-content corresponding to each content category from the target content according to the content category; A target sub-content vectorization unit is configured to vectorize at least one target sub-content corresponding to each content category to obtain at least one target sub-content vector corresponding to the content category for each content category; A multi-modal similarity determination unit is configured to determine the multi-modal similarity between the candidate similar content and the target content according to at least one candidate similar sub-content vector corresponding to each content category of the candidate similar content in the candidate similar content set and at least one target sub-content vector corresponding to each content category of the target content; A similar content determination unit is configured to determine the similar content matched with the target content from the candidate similar content set according to the multi-modal similarity between each of the candidate similar content and the target content; The candidate similar content set comprises at least one candidate similar content; at least one candidate similar sub-content corresponding to each content category is acquired from each candidate similar content according to the content category; and for each content category of each candidate similar content, at least one candidate similar sub-content vector corresponding to the content category is vectorized to obtain at least one candidate similar sub-content vector corresponding to the content category; The device further comprises a candidate similar content set generation unit; The candidate similar content set generation unit comprises: A candidate sub-content vector set generation module is configured to set all candidate sub-content vectors corresponding to all candidate sub-contents of the same content category of all candidate contents in the candidate content set as a candidate sub-content vector set corresponding to the content category. a cluster center determination module configured to, for each candidate sub-content vector set corresponding to a content category, perform clustering on the candidate sub-content vector set corresponding to the content category by using Faiss to obtain at least one cluster center of the candidate sub-content vector set corresponding to the content category; a cluster center selection module configured to, for each target sub-content vector corresponding to a content category of the target content, calculate a distance value between the target sub-content vector and each cluster center of at least one cluster center of a candidate sub-content vector set corresponding to the same content category, and select K cluster centers with the smallest distance values as K cluster centers corresponding to the target sub-content vector; and a candidate similar content determination module configured to, for each cluster center of the K cluster centers corresponding to the target sub-content vector, calculate a distance between each candidate sub-content vector in a cluster corresponding to the cluster center and the target sub-content vector, and determine a candidate content corresponding to a candidate sub-content vector corresponding to a distance meeting a specified recall condition as a candidate similar content. The candidate content is a historical target content that has been determined to be similar. At least one candidate sub-content vector corresponding to each content category of the candidate content is at least one target sub-content vector corresponding to the same content category of a historical target content. The candidate sub-content vector set corresponding to the content category is composed of candidate sub-content vectors corresponding to all candidate sub-content vectors corresponding to the same content category of all candidate contents.

Images