A method and related apparatus for content search

Abstract

Embodiments of the present application disclose a content search method and related device, which are applied to artificial intelligence or vehicle-mounted scenarios, convert a delivery feature value feature of a candidate content in a candidate content set into a delivery feature, interactively fuse a first content encoding vector obtained by the delivery feature and a content feature of the encoded candidate content into a second content encoding vector, at least one dimension component in the second content encoding vector representing a fusion feature of the delivery feature and the content feature, search a target content in the candidate content set through a target object encoding vector obtained by the second content encoding vector and an object feature of an encoded target object. The method enhances the influence effect of the delivery feature value of the candidate content in the content search process, effectively improves the sensitivity of the content search effect to the delivery feature value of the candidate content, and thus improves the content search effect. The content is, for example, an advertisement, and the delivery feature value is, for example, a delivery bid.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular to a method and related apparatus for content search. Background Technology

[0002] After a content creator submits content to a content platform, the platform stores multiple pieces of content. When an audience browses the content platform, the platform needs to consider the audience's characteristics to search for multiple pieces of content, so that it can subsequently display content that is highly relevant to the audience and thus obtain their feedback; therefore, the effectiveness of content search is particularly important.

[0003] In related technologies, content search methods typically encode object features and content features separately to obtain two vectors of equal dimensions. Based on these two vectors, multiple pieces of content are searched to obtain content that is more relevant to the object.

[0004] In reality, when content advertisers place content, the bid price is an important piece of information that needs to be reflected in the content search results. However, the content search methods mentioned above do not consider the impact of the bid price on the content search results. In other words, the content search results are not sensitive to the bid price, resulting in poor content search performance. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a content search method and related apparatus that enhances the influence of candidate content delivery feature values ​​during the content search process. This facilitates the effective reflection of candidate content delivery feature values ​​in the content search results, thereby improving the sensitivity of the content search results to the delivery feature values ​​of candidate content and ultimately enhancing the content search performance.

[0006] The embodiments of this application disclose the following technical solutions:

[0007] On the one hand, this application provides a method for content search, the method comprising:

[0008] The delivery feature values ​​of the candidate content in the candidate content set are subjected to feature transformation processing to obtain the delivery feature of the candidate content;

[0009] The first content encoding vector of the candidate content and the delivery feature are interactively fused to obtain the second content encoding vector of the candidate content; the first content encoding vector is obtained by encoding the content features of the candidate content, and at least one dimension of the second content encoding vector represents the fusion feature of the delivery feature and the content feature.

[0010] Based on the second content encoding vector and the target object encoding vector of the target object, candidate content is searched in the candidate content set to obtain the target content; the target object encoding vector is obtained by encoding the object features of the target object.

[0011] On the other hand, this application provides a content search apparatus, the apparatus comprising: a feature conversion unit, a fusion unit, and a search unit;

[0012] The feature conversion unit is used to perform feature conversion processing on the delivery feature values ​​of candidate content in the candidate content set to obtain the delivery features of the candidate content.

[0013] The fusion unit is used to perform interactive fusion processing on the delivery feature and the first content encoding vector of the candidate content to obtain the second content encoding vector of the candidate content; the first content encoding vector is obtained by encoding the content features of the candidate content, and at least one dimension component in the second content encoding vector represents the fusion feature of the delivery feature and the content feature;

[0014] The search unit is used to search for candidate content in the candidate content set based on the second content encoding vector and the target object encoding vector of the target object, and obtain the target content; the target object encoding vector is obtained by encoding the object features of the target object.

[0015] On the other hand, this application provides a device for content search, the device including a processor and a memory:

[0016] The memory is used to store program code and transmit the program code to the processor;

[0017] The processor is used to execute the content search method described above according to the instructions in the program code.

[0018] On the other hand, embodiments of this application provide a computer-readable storage medium for storing a computer program that, when executed by a processor, performs the content search method described above.

[0019] On the other hand, embodiments of this application provide a computer program product, which includes a computer program or instructions; when the computer program or instructions are executed by a processor, the content search method described above is performed.

[0020] As can be seen from the above technical solution, the delivery feature values ​​of candidate content in the candidate content set are transformed into delivery features; the delivery features and the first content encoding vector of the candidate content are interactively fused into a second content encoding vector, wherein the first content encoding vector is obtained by encoding the content features of the candidate content, and at least one dimension of the second content encoding vector represents the fusion feature of the delivery features and the content features; the target content is obtained by searching for candidate content in the candidate content set through the second content encoding vector and the target object encoding vector of the target object, wherein the target object encoding vector is obtained by encoding the object features of the target object. Based on this, the method interactively fuses the delivery features obtained by transforming the delivery feature values ​​of candidate content with the first content encoding vector obtained by encoding the content features of candidate content, so that the second content encoding vector after interactive fusion can strengthen the integration of delivery features; thereby enhancing the influence of the delivery feature values ​​of candidate content in the content search process, making it easier for the delivery feature values ​​of candidate content to be effectively reflected in the content search results, so as to effectively improve the sensitivity of the content search results to the delivery feature values ​​of candidate content, thereby improving the content search results. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A schematic diagram of the structure of the advertising search method in the related technology provided in the embodiments of this application;

[0023] Figure 2 This is a schematic diagram illustrating an application scenario of a content search method provided in an embodiment of this application;

[0024] Figure 3 A flowchart illustrating a content search method provided in this application embodiment;

[0025] Figure 4 A schematic diagram illustrating the interactive fusion processing of discrete delivery features and a first advertising encoding vector, as provided in an embodiment of this application;

[0026] Figure 5 A schematic diagram illustrating the interactive fusion processing of continuous delivery features and a first advertising encoding vector provided in an embodiment of this application;

[0027] Figure 6 A schematic diagram of a content search device provided in an embodiment of this application;

[0028] Figure 7 This application provides a schematic diagram of the structure of a server according to an embodiment of the present application.

[0029] Figure 8 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. Detailed Implementation

[0030] The embodiments of this application will now be described with reference to the accompanying drawings.

[0031] Currently, content creators can place content on content platforms, just as advertisers can place ads on advertising platforms. (See [link to relevant documentation]). Figure 1 The provided schematic diagram of the advertising search method in the related technology shows that the object features are encoded to obtain the object encoding vector and the advertising features are encoded to obtain the advertising encoding vector. The inner product operation is performed on the object encoding vector and the advertising encoding vector to obtain the inner product. Based on the inner product, multiple advertisements are searched according to the maximum inner product search algorithm to obtain advertisements with good relevance to the object.

[0032] Research has revealed that when advertisers place ads, the bid price is a crucial piece of information that needs to be reflected in ad search performance. However, the aforementioned ad search methods do not consider the impact of ad bid price on ad search performance; that is, ad search performance is not sensitive to ad bid price, resulting in poor ad search performance.

[0033] In view of this, this application proposes a content search method and related apparatus, which interactively fuses the placement features obtained from the bid of candidate ads and the first ad encoding vector obtained from the ad features of the encoded candidate ads, so that the second ad encoding vector after interactive fusion can strengthen the integration of placement features; thereby enhancing the influence of the bid of candidate ads in the ad search process, so that the bid of candidate ads can be effectively reflected in the ad search results, thereby effectively improving the sensitivity of ad search results to the bid of candidate ads, and thus improving ad search results.

[0034] To facilitate understanding of the technical solution of this application, the content search method provided in the embodiments of this application will be introduced below in conjunction with actual application scenarios.

[0035] See Figure 2 , Figure 2 This is a schematic diagram illustrating an application scenario of a content search method provided in an embodiment of this application. Figure 2 The application scenario shown includes a terminal device 201 and a server 202, wherein the terminal device 201 is used as the target device and the server 202 is used as the content search device.

[0036] The target user uses terminal device 201 to browse the content platform. Server 202 performs feature transformation processing on the delivery feature values ​​of the candidate content in the candidate content set to obtain the delivery features of the candidate content. The candidate content in the candidate content set is stored by the content platform after the content advertiser has placed the ad. For example, if the target user is object A, the content is an advertisement, and the delivery feature value is the bid price, and object A uses terminal device 201 to browse the advertising platform, server 202 can perform feature transformation on the bid prices of the candidate advertisements in the candidate advertisement set to obtain the delivery features of the candidate advertisements.

[0037] Server 202 performs interactive fusion processing on the delivery features and the first content encoding vector of the candidate content to obtain a second content encoding vector of the candidate content. The first content encoding vector is obtained by encoding the content features of the candidate content, and at least one dimension of the second content encoding vector represents the fusion feature of the delivery features and content features. For example, based on the above example, server 202 can interactively fuse the delivery features of the candidate advertisement and the first advertisement encoding vector of the candidate advertisement into a second advertisement encoding vector of the candidate advertisement, wherein the first advertisement encoding vector is obtained by encoding the advertisement features of the candidate advertisement, and at least one dimension of the second advertisement encoding vector represents the fusion feature of the delivery features and advertisement features.

[0038] Server 202 searches for candidate content in the candidate content set based on the second content encoding vector and the target object encoding vector of the target object to obtain the target content; the target object encoding vector is obtained by encoding the object features of the target object. Correspondingly, server 202 sends the target content to terminal device 201 so that terminal device 201 can display the target content through the content platform. For example, based on the above example, server 202 can search for candidate advertisements in the candidate advertisement set to obtain the target advertisement using the second content encoding vector of the candidate advertisement and the object encoding vector of object A, wherein the object encoding vector of object A is obtained by encoding the object features of object A.

[0039] It is evident that by interactively fusing the delivery features obtained from the delivery feature values ​​of candidate content and the first content encoding vector obtained from the content features of the candidate content, the second content encoding vector after interactive fusion can strengthen the integration of delivery features. This enhances the influence of the delivery feature values ​​of candidate content during the content search process, making it easier for the delivery feature values ​​of candidate content to be effectively reflected in the content search results. This effectively improves the sensitivity of the content search results to the delivery feature values ​​of candidate content, thereby improving the content search results.

[0040] That is, the delivery features obtained from the bid of the candidate ads and the first ad encoding vector obtained from the ad features of the candidate ads are interactively fused, so that the second ad encoding vector after interactive fusion can strengthen the integration of delivery features; thereby enhancing the influence of the bid of the candidate ads in the ad search process, so that the bid of the candidate ads can be effectively reflected in the ad search results, thus effectively improving the sensitivity of the ad search results to the bid of the candidate ads, thereby improving the ad search results.

[0041] It is understood that the object characteristics of the target object in the content search method provided in this application involve user and other related data. When the above embodiments of this application are applied to specific products or technologies, it is necessary to obtain separate permission or consent from the user, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0042] The content search method provided in this application is based on Artificial Intelligence (AI). AI is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and generate new intelligent machines that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities.

[0043] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing (NLP), machine learning (ML) / deep learning, autonomous driving, and intelligent transportation.

[0044] In the embodiments of this application, the main artificial intelligence software technologies involved include the aforementioned natural language processing technologies and machine learning / deep learning. For example, it may involve text processing and semantic understanding technologies in natural language processing, and may also involve various artificial neural networks in machine learning / deep learning.

[0045] The content search method provided in this application can be applied to devices with data processing capabilities, such as servers and terminal devices. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services, but is not limited to these. Terminal devices include, but are not limited to, mobile phones, tablets, computers, smart cameras, smart voice interaction devices, smart home appliances, vehicle terminals, and aircraft, but are not limited to these. Terminal devices and servers can be directly or indirectly connected via wired or wireless communication, which is not restricted herein.

[0046] The content search device is capable of performing natural language processing (NLP), an important area within computer science and artificial intelligence. It studies various theories and methods for enabling effective communication between humans and computers using natural language. NLP is a science integrating linguistics, computer science, and mathematics. Therefore, research in this field involves natural language, i.e., the language people use in daily life, and thus it is closely related to linguistic research. NLP technologies typically include text processing, semantic understanding, machine translation, question answering, and knowledge graphs. In this embodiment, the content search device can use text processing and semantic understanding techniques within NLP to encode the content features of candidate content to obtain a first content encoding vector, and encode the object features of the target object to obtain a target object encoding vector, etc.

[0047] The content search device can possess machine learning capabilities. Machine learning is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, and many other disciplines. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning. In the embodiments of this application, the encoding processing in the content search method mainly involves the application of artificial neural networks, using artificial neural networks to achieve encoding.

[0048] The content search method provided in this application embodiment may also involve blockchain, wherein data such as processing parameters involved in interactive fusion processing and encoding processing can be stored on the blockchain.

[0049] The content search method provided in this application can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, or in-vehicle scenarios.

[0050] The following describes in detail the content search method provided in the embodiments of this application, using a server as the content search device.

[0051] See Figure 3 This figure is a flowchart of a content search method provided in an embodiment of this application. Figure 3 As shown, the method for searching this content includes the following steps:

[0052] S301: Perform feature transformation processing on the delivery feature values ​​of candidate content in the candidate content set to obtain the delivery features of candidate content.

[0053] In related technologies, after content providers publish content on a content platform, the object-based content search method simply encodes the object features and content features separately to obtain two equal-dimensional vectors. Based on these two equal-dimensional vectors, multiple content items are searched to obtain content that is more relevant to the object.

[0054] Research has revealed that when content advertisers place content, the bid price is a crucial piece of information that needs to be reflected in the content search results. However, the aforementioned content search methods do not consider the impact of the bid price on the content search results; that is, the content search results are not sensitive to the bid price, resulting in poor content search performance.

[0055] In this embodiment of the application, based on the above-mentioned content search method, it is also necessary to consider the impact of content bidding on content search results, so that the content search results are sensitive to the content bidding. Based on this, after the content publisher publishes content and the published content forms a candidate content set as candidate content, based on the bidding price of the candidate content as the bidding feature value of the candidate content, it is first necessary to convert the bidding feature value of the candidate content into a feature that is easy to encode, and use it as the bidding feature of the candidate content.

[0056] Since the delivery feature values ​​of candidate content can be divided into discrete and continuous types according to data type, in the specific implementation of S301, the delivery feature values ​​of candidate content can be transformed into discrete delivery features or into continuous delivery features. The specific implementation method is as follows:

[0057] The first specific implementation method involves converting the delivery feature values ​​of candidate content into discrete delivery features. This essentially means: First, obtaining the delivery feature values ​​of the candidate content and its delivery mode. The delivery mode refers to display mode, click mode, or conversion mode. Further, the conversion mode is divided into shallow conversion mode and deep conversion mode. Display mode means the candidate content is charged based on the delivery feature value based on impressions; click mode means the candidate content is charged based on the delivery feature value based on clicks; conversion mode means the delivery feature value is charged upon reaching the conversion target; shallow conversion mode means the delivery feature value is charged upon reaching a shallow conversion target; and deep conversion mode means the delivery feature value is charged upon reaching a deep conversion target. Then, given different partition sets corresponding to different delivery modes, the delivery feature values ​​of the candidate content are assigned to the corresponding partitions according to the partition set corresponding to the delivery mode, thus obtaining discrete delivery features, which serve as the delivery features of the candidate content. Therefore, this application provides a possible implementation method where the data type of the delivery feature is discrete. S301 may include, for example, the following S3011-S3012:

[0058] S3011: Obtain the delivery feature values ​​and delivery mode of candidate content.

[0059] In the specific implementation of S3011, considering that the different delivery feature values ​​of different candidate content generally follow a long-tail distribution, that is, among different delivery feature values, the higher delivery feature values ​​are few but important; based on this, in order to avoid the adverse effects caused by the higher delivery feature values ​​actually being abnormal delivery feature values, the delivery feature values ​​set by the content delivery person for the candidate content are used as the initial delivery feature values. After obtaining the initial delivery feature values ​​of the candidate content, a second transformation process is needed to obtain the delivery feature values ​​of the candidate content. The volatility between the delivery feature values ​​of candidate content in the candidate content set is less than the volatility between the initial delivery feature values ​​of candidate content in the candidate content set. Therefore, this application provides a possible implementation method, and the step of obtaining the delivery feature values ​​of candidate content in S3011 may include, for example, the following S1-S2:

[0060] S1: Obtain the initial delivery feature values ​​of the candidate content.

[0061] S2: Perform a second transformation on the initial delivery feature value to obtain the delivery feature value; the volatility between the delivery feature values ​​of candidate content in the candidate content set is less than the volatility between the initial delivery feature values ​​of candidate content in the candidate content set.

[0062] The second transformation process can be, for example, a log(1+x) data transformation process, where x represents the initial delivery feature value of the candidate content. That is, S2 can include, for example, performing a log(1+x) data transformation on the initial delivery feature value x of the candidate content to obtain the delivery feature value of the candidate content. Of course, in this embodiment, the second transformation process can also be implemented in other ways, such as other log function transformation processes, etc., which will not be elaborated here.

[0063] As an example, when the candidate content is a candidate ad and the delivery feature is the delivery bid, the initial delivery bid of the candidate ad is obtained as x. The initial delivery bid x of the candidate ad is transformed by log(1+x) to obtain the delivery bid of the candidate ad as x'.

[0064] As another example, when the candidate content is a candidate ad and the delivery feature is the delivery bid, the initial delivery bids of the candidate ad are obtained as x1 and x2. The initial delivery bids x1 and x2 of the candidate ad are transformed by log(1+x) respectively to obtain the delivery bids of the candidate ad as x1' and x2'.

[0065] In the specific implementation of S3011, the delivery mode for candidate content includes display mode, click mode, or conversion mode. Conversion mode includes shallow conversion mode or deep conversion mode. Display mode can be, for example, a cost-per-click (CPC) model; click mode can be, for example, a cost-per-thousand-impression (CPM) model; and conversion mode can be, for example, an optimization cost-per-thousand-impression (OCPM) or an optimization cost-per-click (OCPC) model optimized based on conversion goals. Correspondingly, shallow conversion mode can be, for example, an OCPM or OCPC model based on shallow conversion goals, and deep conversion mode can be, for example, an OCPM or OCPC model based on deep conversion goals.

[0066] S3012: Based on the partition set corresponding to the delivery mode, the delivery feature values ​​are divided to obtain the delivery features.

[0067] In the specific implementation of S3012, due to the differences in the mean and variance of the candidate content in the candidate content set under different delivery modes, the partition granularity of different partition sets corresponding to different delivery modes is different. In order to divide the delivery feature values ​​of candidate content into more accurate partitions, the partition granularity of the partition set corresponding to the display mode can be determined to be smaller than that of the partition set corresponding to the click mode based on the actual situation, and the partition granularity of the partition set corresponding to the conversion mode is uneven.

[0068] As an example, the partition granularity of the partition set corresponding to the CPM mode is smaller than that of the partition set corresponding to the CPC mode, and the partition granularity of the partition sets corresponding to the OCPM and OCPC modes is uneven. If the candidate ad's delivery mode is CPM mode (or CPC mode, OCPM based on shallow conversion goals, or OCPC based on shallow conversion goals), and the bid of the obtained candidate ad is x', the bid x' of the candidate ad is divided using the partition set corresponding to the CPM mode (or CPC mode, OCPM based on shallow conversion goals, or OCPC based on shallow conversion goals) to obtain the delivery feature y of the candidate ad. If the candidate ad's delivery mode is OCPM based on deep conversion goals (or OCPC based on deep conversion goals), and the bids of the obtained candidate ads are x1' and x2', the bids of the candidate ads x1' and x2' are divided using the partition set corresponding to OCPM based on deep conversion goals (or OCPC based on deep conversion goals) to obtain delivery features y1 and y2.

[0069] The second specific implementation method involves converting the delivery feature values ​​of candidate content into continuous delivery features. This essentially means: first, obtaining the delivery feature values ​​of the candidate content; then determining whether the number of delivery feature values ​​is single or multiple. If it is single, the delivery feature value can be directly transformed into a continuous delivery feature as the delivery feature of the candidate content; if it is multiple, the multiple delivery feature values ​​of the candidate content need to be merged and then transformed into a continuous delivery feature as the delivery feature of the candidate content. Therefore, this application provides a possible implementation method where the data type of the delivery feature is continuous. S301 may include, for example, the following S3013-S3015:

[0070] S3013: Obtain the delivery feature values ​​of candidate content.

[0071] Similarly, referring to the specific implementation of the step of obtaining the delivery feature value of candidate content in S3011 above, when S3013 is specifically implemented, S3013 may include, for example: obtaining the initial delivery feature value of candidate content; performing a second transformation process on the initial delivery feature value to obtain the delivery feature value; the volatility between the delivery feature values ​​of candidate content in the candidate content set is less than the volatility between the initial delivery feature values ​​of candidate content in the candidate content set.

[0072] S3014: If the delivery feature value of the candidate content is a single value, perform the first transformation process on the delivery feature value to obtain the delivery feature.

[0073] Wherein, when the delivery mode of the candidate content is display mode, click mode, or shallow conversion mode, and the delivery feature value of the candidate content corresponds to a single target, then the delivery feature value of the candidate content is single. Therefore, this application provides a possible implementation method in which the delivery feature value of the candidate content is single when the delivery mode of the candidate content is display mode, click mode, or shallow conversion mode.

[0074] The first transformation process can be, for example, a continuous-type standardization transformation or a log function transformation; that is, S3014 can include, for example, performing a continuous-type standardization transformation or a log function transformation on the candidate content's delivery feature value if the candidate content's delivery feature value is single, to obtain the candidate content's delivery feature. Of course, in this embodiment, the second transformation process can also be implemented in other ways, such as other function transformations, etc., which will not be elaborated here.

[0075] S3015: If there are multiple delivery feature values ​​for candidate content, perform fusion processing and first transformation processing on the multiple delivery feature values ​​to obtain delivery features.

[0076] In this case, when the candidate content is delivered in a deep conversion mode, the delivery feature value of the candidate content corresponds to multiple targets, thus the candidate content has multiple delivery feature values. Therefore, this application provides a possible implementation method in which the candidate content has multiple delivery feature values ​​when the delivery mode is deep conversion mode.

[0077] The fusion process can be, for example, a weighted process; that is, S3015 may include, for example, if there are multiple delivery feature values ​​for candidate content, first performing a weighted process on the multiple delivery feature values, and then performing a continuous-type standardized transformation or a log function transformation to obtain the delivery features of the candidate content. Of course, in this embodiment of the application, the fusion process can also be implemented in other ways, which will not be elaborated here.

[0078] As an example, when candidate content is candidate ads and the delivery feature is the bid, if the candidate ad's delivery mode is CPM, CPC, OCPM based on shallow conversion goals, or OCPC based on shallow conversion goals, the obtained bid for the candidate ad is x'. The bid x' is then subjected to continuous standardization or logarithmic transformation to obtain the candidate ad's delivery feature y. If the candidate ad's delivery mode is OCPM or OCPC based on deep conversion goals, the obtained bids for the candidate ads are x1' and x2'. The bids x1' and x2' are first weighted, then subjected to continuous standardization or logarithmic transformation to obtain the candidate ad's delivery feature y. The weighting process can, for example, use the formula αx1' + βx2', where the values ​​of α and β are adjusted according to actual application needs; for example, α = 0.5, β = 0.5.

[0079] S302: Perform interactive fusion processing on the first content encoding vector of the delivery feature and the candidate content to obtain the second content encoding vector of the candidate content; the first content encoding vector is obtained by encoding the content features of the candidate content, and at least one dimension of the second content encoding vector represents the fusion feature of the delivery feature and the content feature.

[0080] In this embodiment, after obtaining the delivery features of candidate content in S301, it is necessary to perform interactive fusion processing on the delivery features of candidate content and the first content encoding vector obtained by encoding the content features of candidate content to obtain the interactively fused second content encoding vector. In the interactively fused second content encoding vector, at least one dimension represents the fusion feature of the delivery features and the content features. Compared to the vector obtained by simply concatenating the delivery features of candidate content with the first content encoding vector, where each dimension only represents either the delivery feature or the content feature, the interactively fused second content encoding vector can strengthen the integration of delivery features. This enhances the influence of the delivery feature values ​​of candidate content during the content search process, making it easier for the delivery feature values ​​of candidate content to be effectively reflected in the content search results, thereby effectively improving the sensitivity of the content search results to the delivery feature values ​​of candidate content and ultimately improving the content search results.

[0081] Corresponding to the specific implementation of S301 above, the delivery feature value of the candidate content can be transformed into a discrete delivery feature, or it can be transformed into a continuous delivery feature. In the specific implementation of S302, the interaction and fusion processing method between the discrete delivery feature and the first content encoding vector is different from that between the continuous delivery feature and the first content encoding vector. The specific implementation method is as follows:

[0082] The first specific implementation method, the interactive fusion processing of discrete delivery features and the first content encoding vector, refers to: first, encoding discrete delivery features to obtain a delivery feature encoding vector, the dimension of which needs to match the dimension of the first content encoding vector; then, fusing the delivery feature encoding vector to the first content encoding vector through vector fusion to obtain the second content encoding vector. In this method, the second content encoding vector is obtained by interactively fusing the delivery feature encoding vector and the first content encoding vector, whereby the delivery feature encoding vector is obtained by encoding the delivery features of the candidate content, and the first content encoding vector is obtained by encoding the content features of the candidate content; therefore, the second content encoding vector interactively fuses the delivery features and content features of the candidate content. That is, this application provides a possible implementation method, and S302 may include, for example, the following S3021-S3022:

[0083] S3021: Encode the delivery features to obtain the delivery feature encoding vector; the dimension of the delivery feature encoding vector matches the dimension of the first content encoding vector.

[0084] In the specific implementation of S3021, firstly, the discrete delivery features need to be subjected to simple encoding processing, such as embedding, to obtain multiple delivery feature embedding vectors. Similarly, the dimension of the delivery feature embedding vectors needs to match the dimension of the first content encoding vector. Then, the multiple delivery feature embedding vectors need to undergo further complex encoding processing, such as feature interaction processing and pooling. First, feature interaction processing is performed on the multiple delivery feature embedding vectors to obtain delivery feature interaction vectors, and then pooling is performed on the delivery feature interaction vectors to obtain delivery feature encoding vectors. Therefore, this application provides a possible implementation method, and S3021 may include, for example, the following S3-S5:

[0085] S3: Embed the delivery features to obtain multiple delivery feature embedding vectors; the dimension of the delivery feature embedding vectors matches the dimension of the first content encoding vector.

[0086] S4: Perform feature interaction processing on multiple delivery feature embedding vectors to obtain delivery feature interaction vectors; the dimension of the delivery feature interaction vectors matches the dimension of the first content encoding vector.

[0087] S5: Perform pooling on the delivery feature interaction vector to obtain the delivery feature encoding vector.

[0088] In order to improve the accuracy of the delivery feature encoding vector, pooling can be weighted pooling, for example, where the weights are set and adjusted according to the actual application requirements, such as by adaptively adjusting the model parameters during the training of the delivery feature encoding model for candidate content.

[0089] S3022: Perform vector fusion processing on the delivery feature encoding vector and the first content encoding vector to obtain the second content encoding vector, and the dimension of the second content encoding vector matches the dimension of the first content encoding vector.

[0090] The vector fusion process can be, for example, vector summation; that is, S3022 can include, for example, adding the delivery feature encoding vector and the first content encoding vector bit by bit to obtain a second content encoding vector, where each component of the second content encoding vector represents the fusion feature of the delivery feature and the content feature. Of course, in this embodiment, the vector fusion process can also be implemented in other ways, such as bit-by-bit multiplication, etc., which will not be elaborated here.

[0091] As an example, when candidate content is candidate ads and the delivery feature is the bid, see [link to relevant documentation]. Figure 4 This diagram illustrates the interactive fusion processing of discrete delivery features and a first ad encoding vector. The right-hand model is the candidate ad delivery feature encoding model, which includes an embedding layer, a feature interaction layer, and a pooling layer. After converting the candidate ad's bid features into discrete delivery features, the discrete delivery features are input into the embedding layer for embedding, outputting multiple delivery feature embedding vectors. These vectors are then input into the feature interaction layer for feature interaction, outputting delivery feature interaction vectors. Finally, the interaction vectors are input into the pooling layer, outputting the candidate ad's delivery feature encoding vector. The intermediate model is the candidate ad ad feature encoding model, which includes an embedding layer, a pre-defined network layer, and a fully connected layer. The candidate ad's ad features are input into this model, passing through the embedding layer, the pre-defined network layer, and the fully connected layer, outputting the candidate ad's first ad encoding vector. Above these two models, the delivery feature encoding vector and the first ad encoding vector are fused, for example, by bitwise addition, to obtain the candidate ad's second ad encoding vector. Each component in the second ad encoding vector represents the fused features of the delivery features and the ad features.

[0092] The second specific implementation method, the interactive fusion processing method of continuous delivery features and the first content encoding vector, refers to: firstly, fusing the delivery features and the first content encoding vector through scalar fusion processing to obtain a scalar, which is used as the fusion scalar; then concatenating the fusion scalar to the first content encoding vector to obtain the second content encoding vector. In this method, the second content encoding vector is obtained by concatenating the fusion scalar and the first content encoding vector, the fusion scalar is obtained by interactively fusing the delivery features and the first content encoding vector, and the first content encoding vector is obtained by encoding the content features of the candidate content; therefore, the second content encoding vector interactively fuses the delivery features and content features of the candidate content. Therefore, this application provides a possible implementation method, and S302 may include, for example, the following S3023-S3024:

[0093] S3023: Perform scalar fusion processing on the delivery features and the first content encoding vector to obtain a fused scalar.

[0094] In the specific implementation of S3023, firstly, the first content encoding vector needs to be processed into a content scalar through a scalar-based transformation process; then, based on the content scalar, the content scalar and the delivery feature are fused to obtain a fused scalar. Therefore, this application provides a possible implementation method, and S3023 may include, for example, the following S6-S7:

[0095] S6: Perform scalar-based transformation on the first content encoding vector to obtain the content scalar.

[0096] The scalar-based transformation process can be, for example, vector compression; that is, S6 can include, for example, performing vector compression on the first content encoding vector using a logistic regression model to obtain a content scalar. Of course, the scalar-based transformation process in this embodiment can also be implemented in other ways, such as bitwise averaging, etc., which will not be elaborated here.

[0097] S7: Merge the content scalar and delivery features to obtain a merged scalar.

[0098] The fusion process can be, for example, a summation process, or further, a weighted process; that is, S7 can include, for example, weighting the content scalar and the delivery features to obtain a fused scalar. Of course, the scalar-based transformation process in this embodiment can also be implemented in other ways, such as product processing, etc., which will not be elaborated here.

[0099] S3024: Concatenate the fused scalar and the first content encoding vector to obtain the second content encoding vector.

[0100] In the specific implementation of S3024, for example, the dimension of the first content encoding vector is k, and the fusion scalar is concatenated to the first content encoding vector to obtain the second content encoding vector, which has a dimension of k+1. The one-dimensional component concatenated in the second content encoding vector represents the fusion feature of the delivery feature and the content feature.

[0101] As an example, when candidate content is candidate ads and the delivery feature is the bid, see [link to relevant documentation]. Figure 5 This diagram illustrates the interactive fusion processing of continuous delivery features and a first ad encoding vector. The model on the right is the ad feature encoding model for candidate ads, which includes an embedding layer, a pre-defined network layer, and a fully connected layer. The ad features of the candidate ads are input into this model, passing through the embedding layer, the pre-defined network layer, and the fully connected layer, outputting the first ad encoding vector of the candidate ads. After converting the bid features of the candidate ads into continuous delivery features, the first content encoding vector is compressed using a logistic regression-like model to obtain an ad scalar. The ad scalar and the continuous delivery features are weighted to obtain a fused scalar. The fused scalar and the first ad encoding vector are concatenated to obtain the second ad encoding vector of the candidate ads. The concatenated one-dimensional component in the second ad encoding vector represents the fused feature of the delivery features and the ad features.

[0102] S303: Based on the second content encoding vector and the target object encoding vector of the target object, search for candidate content in the candidate content set to obtain the target content; the target object encoding vector is obtained by encoding the object features of the target object.

[0103] In this embodiment, after obtaining the second content encoding vector of the candidate content in S302, since the second content encoding vector of the candidate content can enhance the integration of delivery features, it is necessary to use the second content encoding vector of the candidate content to replace the first content encoding vector of the candidate content. Combined with the target object encoding vector obtained by encoding the object features of the target object, candidate content in the candidate content set is searched. This enhances the influence of the delivery features of the candidate content during the content search process, thereby obtaining the target content. This method facilitates the effective reflection of the bid price of candidate ads in the ad search results, effectively improving the sensitivity of ad search results to the bid price of candidate ads, thus improving ad search results.

[0104] In the specific implementation of S303, firstly, a preset operation can be performed on the second content encoding vector and the target object encoding vector of the target object using a preset operation method, and the result of the operation is used as the operation result corresponding to the candidate content; then, based on the preset search conditions for the operation result, the target content can be obtained by searching from the candidate content set using the operation result and the preset search conditions. Therefore, this application provides a possible implementation method, and S303 may include, for example, the following S3031-S3032:

[0105] S3031: Perform a preset operation based on the second content encoding vector and the target object encoding vector of the target object to obtain the operation result corresponding to the candidate content.

[0106] The preset operation can be, for example, an inner product operation; that is, S3031 can include, for example, performing an inner product operation based on the second content encoding vector and the target object encoding vector of the target object to obtain the inner product corresponding to the candidate content. Of course, in this embodiment, the second transformation process can also be implemented in other ways, such as vector distance operation, etc., which will not be elaborated here.

[0107] S3032: Based on the calculation results and preset search conditions, search for candidate content in the candidate content set to obtain the target content.

[0108] The preset search conditions may include, for example, selecting the candidate content corresponding to the top N operation results after sorting from high to low; that is, S3032 may include, for example, sorting the operation results corresponding to the candidate content in the candidate content set from high to low; selecting the candidate content corresponding to the top N operation results after sorting as the target content; N is a positive integer, N≥2.

[0109] In addition, in this embodiment of the application, the preset search condition may be, for example, selecting candidate content that is greater than or equal to a preset threshold; that is, S3032 may include, for example, searching for candidate content in the candidate content set that is greater than or equal to the preset threshold as target content based on the calculation result and the preset threshold.

[0110] As an example, when the candidate content is a candidate advertisement, based on the example regarding S302 above, Figure 4 and Figure 5The model on the left is the target object encoding model for the target object. This model includes an embedding layer, a pre-defined network layer, and a fully connected layer. The object features of the target object are input into this model, and after passing through the embedding layer, the pre-defined network layer, and the fully connected layer, the target object encoding vector of the target object is output. Based on this, the inner product operation is performed on the second ad encoding vector of the candidate ad and the target object encoding vector of the target object to obtain the inner product corresponding to the candidate ad. The inner products corresponding to the candidate ads in the candidate ad set are sorted from high to low. The candidate ads corresponding to the top N inner products after sorting are selected as the target ads; N is a positive integer, N≥2.

[0111] It should be noted that, in the specific implementation of S302 when the data type of the delivery feature is continuous, S6-S7 and S3024, regarding the step of obtaining the target object encoding vector, need to refer to the step of obtaining the second content encoding vector of the aforementioned candidate content. First, the object feature is encoded to obtain the initial object encoding vector; then, the initial object encoding vector needs to be processed into an object scalar through a scalar-based transformation; then, based on the object scalar, the object scalar is concatenated to the initial object encoding vector to obtain the target object encoding vector. Therefore, this application provides a possible implementation, and the steps for obtaining the target object encoding vector may include, for example, the following S8-S10:

[0112] S8: Encode the object features of the target object to obtain the initial object encoding vector.

[0113] S9: Perform scalar-based transformation on the initial object encoding vector to obtain the object scalar.

[0114] S10: Concatenate the above object scalar and the initial object encoding vector to obtain the target object encoding vector.

[0115] For example, if the initial object encoding vector has a dimension of k, concatenating the object scalar to the initial object encoding vector yields the target object encoding vector, which has a dimension of k+1.

[0116] Furthermore, in this embodiment, to further avoid significant differences between the fused scalar value in the second content encoding vector obtained in S3024 and the values ​​of other dimensions in the second content encoding vector, and to further avoid significant differences between the object scalar value in the target object encoding vector obtained in S10 and the values ​​of other dimensions in the target object encoding vector, it is necessary to perform regularization processing on the second content encoding vector and the target object encoding vector, so that both the regularized second content encoding vector and the regularized target object encoding vector conform to regularization constraints. Based on this, when implementing S303, it is necessary to use the regularized second content encoding vector, combined with the regularized target object encoding vector, to search for candidate content in the candidate content set to obtain the target content. Therefore, this application provides a possible implementation method, which may further include the following S11-S12:

[0117] S11: Perform regularization on the second content encoding vector to obtain the regularized second content encoding vector.

[0118] S12: Perform regularization processing on the target object encoding vector to obtain the regularized target object encoding vector.

[0119] Corresponding to S11-S12 above, S303 may include, for example, searching for candidate content in the candidate content set based on the regularized second content encoding vector and the regularized target object encoding vector to obtain the target content.

[0120] The content search method provided in the above embodiments transforms the delivery feature values ​​of candidate content in the candidate content set into delivery features; it then interactively fuses the delivery features and the first content encoding vector of the candidate content into a second content encoding vector, wherein the first content encoding vector is obtained by encoding the content features of the candidate content, and at least one dimension of the second content encoding vector represents the fusion feature of the delivery features and the content features; and it searches for candidate content in the candidate content set to obtain target content through the second content encoding vector and the target object encoding vector of the target object, wherein the target object encoding vector is obtained by encoding the object features of the target object. Based on this, the method interactively fuses the delivery features obtained by transforming the delivery feature values ​​of candidate content with the first content encoding vector obtained by encoding the content features of candidate content, so that the second content encoding vector after interactive fusion can strengthen the integration of delivery features; thereby enhancing the influence of the delivery feature values ​​of candidate content during the content search process, making it easier for the delivery feature values ​​of candidate content to be effectively reflected in the content search results, thus effectively improving the sensitivity of the content search results to the delivery feature values ​​of candidate content, and ultimately improving the content search results.

[0121] The content search method provided in this application embodiment can be applied to an advertising system. The specific implementation of the advertising search method is as follows:

[0122] One specific implementation method is as follows: First, for candidate ads in the candidate ad set, obtain the initial bid and the ad delivery mode of the candidate ads. Perform a log(1+x) data transformation on the initial bid of the candidate ads to obtain the bid of the candidate ads, where x represents the initial bid. When the ad delivery mode of the candidate ads is CPM mode (or CPC mode, OCPM based on shallow conversion goals, OCPC based on shallow conversion goals, OCPM based on deep conversion goals, or OCPC based on deep conversion goals), divide the bid of the candidate ads based on the partition set corresponding to the CPM mode (or CPC mode, OCPM based on shallow conversion goals, OCPC based on shallow conversion goals, OCPM based on deep conversion goals, or OCPC based on deep conversion goals), and obtain the delivery characteristics of the candidate ads.

[0123] Then, the delivery features of the candidate ads are input into the embedding layer for embedding processing, and multiple delivery feature embedding vectors are output. The multiple delivery feature embedding vectors are input into the feature interaction layer for feature interaction processing, and delivery feature interaction vectors are output. The delivery feature interaction vectors are input into the pooling layer, and delivery feature encoding vectors of the candidate ads are output. The delivery feature encoding vectors of the candidate ads and the first ad encoding vector of the candidate ads are added bitwise to obtain the second ad encoding vector of the candidate ads. The first ad encoding vector of the candidate ads is obtained by inputting the ad features of the candidate ads into the embedding layer, the preset network layer and the fully connected layer for encoding processing, and the output of the second ad encoding vector. Each component in the second ad encoding vector represents the fusion feature of delivery features and ad features.

[0124] Finally, the inner product operation is performed on the second ad encoding vector of the candidate ad and the target object encoding vector of the target object to obtain the inner product corresponding to the candidate ad. The target object encoding vector of the target object is obtained by inputting the object features of the target object into the embedding layer, the preset network layer and the fully connected layer for encoding processing and output. The inner products corresponding to the candidate ads in the candidate ad set are sorted from high to low. The candidate ads corresponding to the top N inner products after sorting are selected as the target ads. N is a positive integer, N≥2.

[0125] Another specific implementation method is as follows: First, for candidate ads in the candidate ad set, obtain the initial bid of the candidate ad and perform a log(1+x) data transformation on the initial bid of the candidate ad to obtain the bid of the candidate ad, where x represents the initial bid. If the delivery mode of the candidate ad is CPM mode, CPC mode, OCPM based on shallow conversion goals, or OCPC based on shallow conversion goals, perform a continuous standardization transformation or a log function transformation on the bid of the candidate ad to obtain the delivery characteristics of the candidate ad. If the delivery mode of the candidate ad is OCPM based on deep conversion goals or OCPC based on deep conversion goals, perform a weighted processing on the bid of the candidate ad first, and then perform a continuous standardization transformation or a log function transformation to obtain the delivery characteristics of the candidate ad.

[0126] Then, after encoding the advertising features of the candidate ads into the embedding layer, the preset network layer, and the fully connected layer to obtain the first advertising encoding vector of the candidate ads, the first content encoding vector is compressed using a logistic regression-like model to obtain the advertising scalar. The advertising scalar and the continuous delivery features are weighted to obtain the fusion scalar. The fusion scalar and the first advertising encoding vector are concatenated to obtain the second advertising encoding vector of the candidate ads. The concatenated one-dimensional components in the second advertising encoding vector all represent the fusion features of the delivery features and the advertising features.

[0127] Finally, the target object encoding vector is obtained by inputting the object features of the target object into the embedding layer, the preset network layer, and the fully connected layer for encoding, outputting an initial object encoding vector. This initial object encoding vector is then transformed using scalar transformation to obtain an object scalar. The object scalar and the initial object encoding vector are then concatenated to obtain the target object encoding vector. Next, the second ad encoding vector of the candidate ads needs to be regularized to obtain a regularized second ad encoding vector. The target object encoding vector is then regularized to obtain a regularized target object encoding vector. An inner product operation is performed on the regularized second ad encoding vector and the regularized target object encoding vector to obtain the inner product corresponding to the candidate ads. The inner products of the candidate ads in the candidate ad set are sorted from highest to lowest. The candidate ads corresponding to the top N sorted inner products are selected as the target ads; N is a positive integer, N≥2.

[0128] In addition to the content search method provided in the above embodiments, this application also provides a content search apparatus.

[0129] See Figure 6 , Figure 6 This is a schematic diagram of a content search device provided in an embodiment of this application. Figure 6As shown, the content search device 600 includes: a feature conversion unit 601, a fusion unit 602, and a search unit 603;

[0130] The feature conversion unit 601 is used to perform feature conversion processing on the delivery feature values ​​of candidate content in the candidate content set to obtain the delivery features of candidate content;

[0131] The fusion unit 602 is used to perform interactive fusion processing on the first content encoding vector of the delivery feature and the candidate content to obtain the second content encoding vector of the candidate content; the first content encoding vector is obtained by encoding the content features of the candidate content, and at least one dimension of the second content encoding vector represents the fusion feature of the delivery feature and the content feature.

[0132] Search unit 603 is used to search for candidate content in the candidate content set to obtain target content based on the second content encoding vector and the target object encoding vector of the target object; the target object encoding vector is obtained by encoding the object features of the target object.

[0133] As one possible implementation, the data type of the delivered features is discrete, and the feature conversion unit 601 includes: a first acquisition subunit and a division subunit;

[0134] The first acquisition subunit is used to acquire the delivery feature values ​​and delivery patterns of candidate content;

[0135] Sub-units are used to divide the delivery feature values ​​based on the partition set corresponding to the delivery mode, so as to obtain the delivery features;

[0136] The fusion unit 602 includes: a first coding subunit and a first fusion subunit;

[0137] The first encoding subunit is used to encode the delivery features to obtain the delivery feature encoding vector; the dimension of the delivery feature encoding vector matches the dimension of the first content encoding vector.

[0138] The first fusion subunit is used to perform vector fusion processing on the delivery feature encoding vector and the first content encoding vector to obtain the second content encoding vector, wherein the dimension of the second content encoding vector matches the dimension of the first content encoding vector.

[0139] As one possible implementation, the encoding subunit includes: an embedding module, a feature interaction module, and a pooling module;

[0140] The embedding module is used to embed the delivery features to obtain multiple delivery feature embedding vectors; the dimension of the delivery feature embedding vectors matches the dimension of the first content encoding vector.

[0141] The feature interaction module is used to perform feature interaction processing on multiple delivery feature embedding vectors to obtain delivery feature interaction vectors; the dimension of the delivery feature interaction vector is matched with the dimension of the first content encoding vector.

[0142] The pooling module is used to pool the delivery feature interaction vector to obtain the delivery feature encoding vector.

[0143] As one possible implementation method, the delivery mode includes display mode, click mode, or conversion mode. The conversion mode includes shallow conversion mode or deep conversion mode. The partition granularity of the partition set corresponding to the display mode is smaller than that of the partition set corresponding to the click mode, and the partition granularity of the partition set corresponding to the conversion mode is uneven.

[0144] As one possible implementation, the data type of the delivered feature is continuous, and the feature conversion unit 601 includes: a second acquisition subunit, a first transformation subunit, and a second transformation subunit;

[0145] The second acquisition subunit is used to acquire the delivery feature values ​​of candidate content;

[0146] The first transformation subunit is used to perform a first transformation process on the delivery feature value if the delivery feature value of the candidate content is a single value, so as to obtain the delivery feature.

[0147] The second transformation subunit is used to perform fusion processing and first transformation processing on multiple delivery feature values ​​if the candidate content has multiple delivery feature values, in order to obtain delivery features.

[0148] The fusion unit 602 includes: a second fusion subunit and a first splicing subunit;

[0149] The second fusion subunit is used to perform scalar fusion processing on the delivery features and the first content encoding vector to obtain a fused scalar.

[0150] The first splicing subunit is used to splice the fused scalar and the first content encoding vector to obtain the second content encoding vector.

[0151] As one possible implementation, the second fusion subunit includes: a conversion module and a fusion module;

[0152] The conversion module is used to perform scalar-based conversion processing on the first content encoding vector to obtain the content scalar.

[0153] The fusion module is used to fuse content scalars and delivery features to obtain a fused scalar.

[0154] As one possible implementation, the device further includes: an encoding unit; the encoding unit includes: a second encoding subunit, a conversion subunit, and a second splicing subunit;

[0155] The second encoding subunit is used to encode the object features of the target object to obtain the initial object encoding vector;

[0156] The transformation subunit is used to perform scalar-based transformation processing on the initial object encoding vector to obtain the object scalar.

[0157] The second splicing subunit is used to splice the above object scalar and the initial object encoding vector to obtain the target object encoding vector.

[0158] As one possible implementation, the device further includes: a first regularization unit and a second regularization unit;

[0159] The first regularization unit is used to perform regularization processing on the second content encoding vector to obtain the regularized second content encoding vector.

[0160] The second regularization unit is used to perform regularization processing on the target object encoding vector to obtain the regularized target object encoding vector;

[0161] Search unit 603 is used for:

[0162] Based on the regularized second content encoding vector and the regularized target object encoding vector, candidate content is searched in the candidate content set to obtain the target content.

[0163] As one possible implementation, when the candidate content is delivered in display mode, click mode, or shallow conversion mode, the candidate content has a single delivery feature value; when the delivery mode is deep conversion mode, the candidate content has multiple delivery feature values.

[0164] As one possible implementation, the first or second acquisition subunit includes: an acquisition module and a transformation module;

[0165] The acquisition module is used to obtain the initial delivery feature values ​​of candidate content;

[0166] The transformation module is used to perform a second transformation on the initial delivery feature value to obtain the delivery feature value; the volatility between the delivery feature values ​​of the candidate content in the candidate content set is less than the volatility between the initial delivery feature values ​​of the candidate content in the candidate content set.

[0167] As one possible implementation, the search unit 603 includes: an operation subunit and a search subunit;

[0168] The operation subunit is used to perform a preset operation based on the second content encoding vector and the target object encoding vector of the target object to obtain the operation result corresponding to the candidate content;

[0169] The search subunit is used to search for candidate content in the candidate content set based on the calculation results and preset search conditions, and obtain the target content.

[0170] The content search apparatus provided in the above embodiments transforms the delivery feature values ​​of candidate content in a candidate content set into delivery features; it then interactively fuses the delivery features and a first content encoding vector of the candidate content into a second content encoding vector, wherein the first content encoding vector is obtained by encoding the content features of the candidate content, and at least one dimension of the second content encoding vector represents the fusion feature of the delivery features and the content features; and it searches for candidate content in the candidate content set to obtain target content through the second content encoding vector and the target object encoding vector of the target object, wherein the target object encoding vector is obtained by encoding the object features of the target object. Based on this, the method interactively fuses the delivery features obtained by transforming the delivery feature values ​​of candidate content into delivery features with the first content encoding vector obtained by encoding the content features of candidate content, so that the second content encoding vector after interactive fusion can strengthen the integration of delivery features; thereby enhancing the influence of the delivery feature values ​​of candidate content during the content search process, making it easier for the delivery feature values ​​of candidate content to be effectively reflected in the content search results, thus effectively improving the sensitivity of the content search results to the delivery feature values ​​of candidate content, and ultimately improving the content search results.

[0171] In addition to the content search method described above, this application also provides a device for content search, so that the above content search method can be implemented and applied in practice. The computer device provided in this application will be described below from the perspective of hardware physicalization.

[0172] See Figure 7 , Figure 7 This is a schematic diagram of a server structure provided in an embodiment of this application. The server 700 can vary significantly due to different configurations or performance. It may include one or more Central Processing Units (CPUs) 722 (e.g., one or more processors) and a memory 732, and one or more storage media 730 (e.g., one or more mass storage devices) for storing application programs 742 or data 744. The memory 732 and storage media 730 can be temporary or persistent storage. The program stored in the storage media 730 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the server. Furthermore, the CPU 722 may be configured to communicate with the storage media 730 and execute the series of instruction operations stored in the storage media 730 on the server 700.

[0173] Server 700 may also include one or more power supplies 726, one or more wired or wireless network interfaces 750, one or more input / output interfaces 758, and / or one or more operating systems 741, such as Windows Server. TM Mac OS X TM Unix TM Linux TM FreeBSD TM etc.

[0174] The steps performed by the server in the above embodiments can be based on this Figure 7 The server structure shown.

[0175] The CPU 722 is used to perform the following steps:

[0176] The delivery feature values ​​of the candidate content in the candidate content set are processed by feature transformation to obtain the delivery features of the candidate content;

[0177] The first content encoding vector of the delivery feature and the candidate content are interactively fused to obtain the second content encoding vector of the candidate content; the first content encoding vector is obtained by encoding the content features of the candidate content, and at least one component in the second content encoding vector represents the fusion feature of the delivery feature and the content feature.

[0178] Based on the second content encoding vector and the target object encoding vector of the target object, candidate content is searched in the candidate content set to obtain the target content; the target object encoding vector is obtained by encoding the object features of the target object.

[0179] Optionally, the CPU 1222 may also execute method steps of any specific implementation of the content search method in the embodiments of this application.

[0180] See Figure 8 , Figure 8 This is a schematic diagram of a terminal device provided in an embodiment of this application. For ease of explanation, only the parts related to the embodiment of this application are shown; for specific technical details not disclosed, please refer to the method section of the embodiment of this application. The terminal device can be any terminal device including mobile phones, tablets, PDAs, etc. Taking a mobile phone as an example:

[0181] Figure 8 This diagram illustrates a partial structural representation of a mobile phone related to the terminal device provided in this embodiment. (Reference) Figure 8The mobile phone includes components such as a radio frequency (RF) circuit 810, a memory 820, an input unit 830, a display unit 840, a sensor 850, an audio circuit 860, a Wi-Fi module 870, a processor 880, and a power supply 890. Those skilled in the art will understand that... Figure 8 The mobile phone structure shown does not constitute a limitation on the mobile phone and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0182] The following is combined Figure 8 A detailed introduction to each component of a mobile phone:

[0183] RF circuit 810 can be used for receiving and transmitting signals during information transmission or calls. Specifically, it receives downlink information from the base station and processes it with processor 880; additionally, it transmits uplink data to the base station. Typically, RF circuit 810 includes, but is not limited to, an antenna, at least one amplifier, a transceiver, a coupler, a low-noise amplifier (LNA), and a duplexer. Furthermore, RF circuit 810 can also communicate wirelessly with networks and other devices. The aforementioned wireless communication can use any communication standard or protocol, including but not limited to Global System for Mobile Communications (GSM), General Packet Radio Service (GPRS), Code Division Multiple Access (CDMA), Wideband Code Division Multiple Access (WCDMA), Long Term Evolution (LTE), email, and Short Messaging Service (SMS).

[0184] The memory 820 can be used to store software programs and modules. The processor 880 runs the software programs and modules stored in the memory 820 to realize various functions and data processing of the mobile phone. The memory 820 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory 820 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0185] The input unit 830 can be used to receive input numerical or character information, and to generate key signal inputs related to user settings and function control of the mobile phone. Specifically, the input unit 830 may include a touch panel 831 and other input devices 832. The touch panel 831, also known as a touch screen, can collect touch operations performed by the user on or near it (such as operations performed by the user using a finger, stylus, or any suitable object or accessory on or near the touch panel 831), and drive the corresponding connection devices according to a pre-set program. Optionally, the touch panel 831 may include two parts: a touch detection device and a touch controller. The touch detection device detects the user's touch position and the signal generated by the touch operation, and transmits the signal to the touch controller; the touch controller receives touch information from the touch detection device, converts it into touch point coordinates, and sends it to the processor 880, and can also receive and execute commands sent by the processor 880. In addition, the touch panel 831 can be implemented using various types such as resistive, capacitive, infrared, and surface acoustic wave. In addition to the touch panel 831, the input unit 830 may also include other input devices 832. Specifically, other input devices 832 may include, but are not limited to, one or more of the following: physical keyboard, function keys (such as volume control buttons, power buttons, etc.), trackball, mouse, joystick, etc.

[0186] The display unit 840 can be used to display information input by the user or information provided to the user, as well as various menus of the mobile phone. The display unit 840 may include a display panel 841, which may optionally be configured as a Liquid Crystal Display (LCD), Organic Light-Emitting Diode (OLED), or similar display panel. Furthermore, a touch panel 831 may cover the display panel 841. When the touch panel 831 detects a touch operation on or near it, it transmits the information to the processor 880 to determine the type of touch event. Subsequently, the processor 880 provides corresponding visual output on the display panel 841 based on the type of touch event. Although in Figure 8 In this embodiment, the touch panel 831 and the display panel 841 are two separate components to realize the input and output functions of the mobile phone. However, in some embodiments, the touch panel 831 and the display panel 841 can be integrated to realize the input and output functions of the mobile phone.

[0187] The mobile phone may also include at least one sensor 850, such as a light sensor, a motion sensor, and other sensors. Specifically, the light sensor may include an ambient light sensor and a proximity sensor. The ambient light sensor can adjust the brightness of the display panel 841 according to the ambient light level, and the proximity sensor can turn off the display panel 841 and / or backlight when the phone is moved to the ear. As a type of motion sensor, an accelerometer sensor can detect the magnitude of acceleration in various directions (generally three axes). When stationary, it can detect the magnitude and direction of gravity, which can be used for applications that recognize the phone's posture (such as landscape / portrait switching, related games, magnetometer posture calibration), vibration recognition-related functions (such as pedometer, taps), etc. Other sensors that may be configured in the mobile phone, such as gyroscopes, barometers, hygrometers, thermometers, and infrared sensors, will not be described in detail here.

[0188] Audio circuit 860, speaker 861, and microphone 862 provide an audio interface between the user and the mobile phone. Audio circuit 860 converts received audio data into electrical signals and transmits them to speaker 861, where speaker 861 converts them into sound signals for output. On the other hand, microphone 862 converts collected sound signals into electrical signals, which are received by audio circuit 860, converted into audio data, and then output to processor 880 for processing. The audio data is then transmitted via RF circuit 810 to, for example, another mobile phone, or output to memory 820 for further processing.

[0189] WiFi is a short-range wireless transmission technology. Through the WiFi module 870, mobile phones can help users send and receive emails, browse web pages, and access streaming media, providing users with wireless broadband internet access. Although Figure 8The WiFi module 870 is shown, but it is understood that it is not an essential component of a mobile phone and can be omitted as needed without changing the essence of the invention.

[0190] The processor 880 is the control center of the mobile phone, connecting various parts of the phone through various interfaces and lines. It performs various functions and processes data by running or executing software programs and / or modules stored in the memory 820, and by calling data stored in the memory 820, thereby controlling the phone as a whole. Optionally, the processor 880 may include one or more processing units; preferably, the processor 880 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor 880.

[0191] The mobile phone also includes a power supply 890 (such as a battery) that supplies power to various components. Preferably, the power supply can be logically connected to the processor 880 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system.

[0192] Although not shown, mobile phones may also include a camera, Bluetooth module, etc., which will not be described in detail here.

[0193] In this embodiment of the application, the memory 820 included in the mobile phone can store program code and transmit the program code to the processor.

[0194] The processor 880 included in the mobile phone can execute the content search method provided in the above embodiments according to the instructions in the program code.

[0195] This application also provides a computer-readable storage medium for storing a computer program for executing the content search method provided in the above embodiments.

[0196] This application also provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the content search method provided in the various optional implementations of the above aspects.

[0197] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium can be at least one of the following media: read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, etc., and other media capable of storing program code.

[0198] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device and system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0199] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for content search, characterized in that, The method includes: The delivery feature value of candidate content in the candidate content set is transformed to obtain the delivery feature of the candidate content. This includes: obtaining the delivery feature value and the delivery mode of the candidate content, where the delivery mode includes display mode, click mode, or conversion mode, and the conversion mode includes shallow conversion mode or deep conversion mode; when the obtained delivery feature is discrete, the delivery feature value is partitioned based on the partition set corresponding to the delivery mode to obtain discrete delivery features, where the partition granularity of the partition set corresponding to the display mode is smaller than the partition granularity of the partition set corresponding to the click mode. The partition granularity of the partition set corresponding to the conversion mode is uneven; or, when the acquired delivery feature is continuous, the delivery feature value is determined to be single or multiple based on the delivery mode. If the delivery mode is display mode, click mode, or shallow conversion mode, then the delivery feature value is single, and the delivery feature value is subjected to a first transformation process to obtain continuous delivery features. If the delivery mode is deep conversion mode, then the delivery feature value is multiple, and the multiple delivery feature values ​​are subjected to fusion processing and the first transformation process to obtain continuous delivery features; the delivery feature value of the candidate content includes the delivery bid of the candidate content; The method involves interactively fusing the delivery feature and the first content encoding vector of the candidate content to obtain a second content encoding vector of the candidate content. This includes: when the data type of the delivery feature is discrete, encoding the delivery feature to obtain a delivery feature encoding vector; the dimension of the delivery feature encoding vector matches the dimension of the first content encoding vector; performing vector fusion processing on the delivery feature encoding vector and the first content encoding vector to obtain a second content encoding vector, where the dimension of the second content encoding vector matches the dimension of the first content encoding vector; or, when the data type of the delivery feature is continuous, performing scalar fusion processing on the delivery feature and the first content encoding vector to obtain a fused scalar; concatenating the fused scalar and the first content encoding vector to obtain the second content encoding vector; where the first content encoding vector is obtained by encoding the content features of the candidate content, and the concatenated one-dimensional component in the second content encoding vector represents the fusion feature of the delivery feature and the content feature. Based on the second content encoding vector and the target object encoding vector of the target object, candidate content is searched in the candidate content set to obtain the target content; the target object encoding vector is obtained by encoding the object features of the target object.

2. The method according to claim 1, characterized in that, The step of encoding the delivery features to obtain the delivery feature encoding vector includes: The delivery feature is embedded to obtain multiple delivery feature embedding vectors; the dimension of the delivery feature embedding vector matches the dimension of the first content encoding vector. The multiple delivery feature embedding vectors are subjected to feature interaction processing to obtain delivery feature interaction vectors; the dimension of the delivery feature interaction vectors matches the dimension of the first content encoding vector. The delivery feature interaction vector is pooled to obtain the delivery feature encoding vector.

3. The method according to claim 1, characterized in that, The step of performing scalar fusion processing on the delivery features and the first content encoding vector to obtain a fused scalar includes: The first content encoding vector is subjected to scalar-based transformation to obtain the content scalar. The content scalar and the delivery feature are fused to obtain the fused scalar.

4. The method according to claim 3, characterized in that, The steps for obtaining the target object encoding vector include: The object features are encoded to obtain an initial object encoding vector; The initial object encoding vector is subjected to scalar-based transformation to obtain the object scalar; The object scalar and the initial object encoding vector are concatenated to obtain the target object encoding vector.

5. The method according to claim 4, characterized in that, The method further includes: The second content encoding vector is regularized to obtain the regularized second content encoding vector; The target object encoding vector is regularized to obtain the regularized target object encoding vector; The step of searching for candidate content in the candidate content set based on the second content encoding vector and the target object encoding vector of the target object to obtain the target content includes: Based on the regularized second content encoding vector and the regularized target object encoding vector, candidate content is searched in the candidate content set to obtain the target content.

6. The method according to claim 1, characterized in that, The step of obtaining the delivery feature value of the candidate content includes: Obtain the initial delivery feature values ​​of the candidate content; The initial delivery feature value is subjected to a second transformation process to obtain the delivery feature value; the volatility among the delivery feature values ​​of the candidate content in the candidate content set is less than the volatility among the initial delivery feature values ​​of the candidate content in the candidate content set.

7. The method according to claim 1, characterized in that, The step of searching for candidate content in the candidate content set based on the second content encoding vector and the target object encoding vector of the target object to obtain the target content includes: Based on the second content encoding vector and the target object encoding vector of the target object, a preset operation is performed to obtain the operation result corresponding to the candidate content; Based on the calculation results and preset search conditions, candidate content in the candidate content set is searched to obtain the target content.

8. A content search device, characterized in that, The device includes: a feature conversion unit, a fusion unit, and a search unit; The feature conversion unit is used to perform feature conversion processing on the delivery feature values ​​of candidate content in the candidate content set to obtain the delivery features of the candidate content, wherein the delivery feature values ​​of the candidate content include the delivery bid of the candidate content. The fusion unit is used to perform interactive fusion processing on the delivery feature and the first content encoding vector of the candidate content to obtain the second content encoding vector of the candidate content; the first content encoding vector is obtained by encoding the content features of the candidate content, and the one-dimensional component concatenated in the second content encoding vector represents the fusion feature of the delivery feature and the content feature; The search unit is used to search for candidate content in the candidate content set based on the second content encoding vector and the target object encoding vector of the target object, and to obtain the target content; the target object encoding vector is obtained by encoding the object features of the target object. When the data type of the delivery feature is discrete, the feature conversion unit includes: a first acquisition subunit and a division subunit; The first acquisition subunit is used to acquire the delivery feature values ​​and delivery patterns of candidate content; Sub-units are used to divide the delivery feature values ​​based on the partition set corresponding to the delivery mode to obtain delivery features; the delivery mode includes display mode, click mode or conversion mode, and the conversion mode includes shallow conversion mode or deep conversion mode; the partition granularity of the partition set corresponding to the display mode is smaller than that of the partition set corresponding to the click mode, and the partition granularity of the partition set corresponding to the conversion mode is uneven. Alternatively, when the data type of the delivery feature is continuous, the feature conversion unit includes: a second acquisition subunit, a first transformation subunit, and a second transformation subunit; The second acquisition subunit is used to acquire the delivery feature value of the candidate content and the delivery mode of the candidate content; based on the delivery mode, the delivery feature value is determined to be single or multiple. When the delivery mode of the candidate content is display mode, click mode or shallow conversion mode, the delivery feature value of the candidate content is single; when the delivery mode is deep conversion mode, the delivery feature value of the candidate content is multiple. The first transformation subunit is used to perform a first transformation process on the delivery feature value if the delivery feature value of the candidate content is a single value, so as to obtain the delivery feature. The second transformation subunit is used to perform fusion processing and first transformation processing on multiple delivery feature values ​​if the candidate content has multiple delivery feature values, in order to obtain delivery features. Wherein, when the data type of the delivery feature is discrete, the fusion unit includes: a first encoding subunit and a first fusion subunit; The first encoding subunit is used to encode the delivery feature to obtain a delivery feature encoding vector; the dimension of the delivery feature encoding vector matches the dimension of the first content encoding vector. The first fusion subunit is used to perform vector fusion processing on the delivery feature encoding vector and the first content encoding vector to obtain the second content encoding vector, wherein the dimension of the second content encoding vector matches the dimension of the first content encoding vector; Alternatively, when the data type of the delivery feature is continuous, the fusion unit includes: a second fusion subunit and a first splicing subunit; The second fusion subunit is used to perform scalar fusion processing on the delivery feature and the first content encoding vector to obtain a fused scalar; The first splicing subunit is used to splice the fused scalar and the first content encoding vector to obtain the second content encoding vector.

9. The content search apparatus according to claim 8, characterized in that, The first encoding subunit includes: an embedding module, a feature interaction module, and a pooling module; The embedding module is used to embed the delivery features to obtain multiple delivery feature embedding vectors; the dimension of the delivery feature embedding vectors matches the dimension of the first content encoding vector. The feature interaction module is used to perform feature interaction processing on multiple delivery feature embedding vectors to obtain delivery feature interaction vectors; the dimension of the delivery feature interaction vector is matched with the dimension of the first content encoding vector. The pooling module is used to pool the delivery feature interaction vector to obtain the delivery feature encoding vector.

10. The content search apparatus according to claim 8, characterized in that, The second fusion subunit includes: a conversion module and a fusion module; The conversion module is used to perform scalar-based conversion processing on the first content encoding vector to obtain the content scalar. The fusion module is used to fuse content scalars and delivery features to obtain a fused scalar.

11. The content search apparatus according to claim 10, characterized in that, The device also includes: an encoding unit; the encoding unit includes: a second encoding subunit, a conversion subunit, and a second splicing subunit; The second encoding subunit is used to encode the object features of the target object to obtain the initial object encoding vector; The transformation subunit is used to perform scalar-based transformation processing on the initial object encoding vector to obtain the object scalar. The second splicing subunit is used to splice the above object scalar and the initial object encoding vector to obtain the target object encoding vector.

12. The content search apparatus according to claim 11, characterized in that, The device further includes: a first regularization unit and a second regularization unit; The first regularization unit is used to perform regularization processing on the second content encoding vector to obtain the regularized second content encoding vector. The second regularization unit is used to perform regularization processing on the target object encoding vector to obtain the regularized target object encoding vector; The search unit is used to search for candidate content in the candidate content set based on the regularized second content encoding vector and the regularized target object encoding vector, and obtain the target content.

13. The content search apparatus according to claim 8, characterized in that, The first or second acquisition subunit includes: an acquisition module and a transformation module; The acquisition module is used to obtain the initial delivery feature values ​​of candidate content; The transformation module is used to perform a second transformation on the initial delivery feature value to obtain the delivery feature value; the volatility between the delivery feature values ​​of the candidate content in the candidate content set is less than the volatility between the initial delivery feature values ​​of the candidate content in the candidate content set.

14. A computer device, characterized in that, The device includes a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is configured to execute the content search method according to any one of claims 1-7 according to the instructions in the program code.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program, which, when executed by a processor, performs the content search method according to any one of claims 1-7.

16. A computer program product, characterized in that, Includes a computer program or instructions; when the computer program or instructions are executed by a processor, the content search method described in any one of claims 1-7 is performed.

Images