Object retrieval method and device, electronic equipment and storage medium

Abstract

The invention provides an object retrieval method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining a retrieval target which comprises at least one seed object; extracting semantic features of each seed object from the description text of each seed object; according to the semantic features of the seed objects, a candidate object set similar to the retrieval target semantics is recalled from a candidate object library; under the condition that the data volume of the candidate object library is smaller than a threshold value, sorting the recalled candidate object sets by utilizing the attribute tags of the seed objects; and determining a similar object recommendation list corresponding to the retrieval target from the candidate object set according to the sorting result. According to the method, the deep semantics of the unstructured text and the concrete index of the structured tag are integrated, so that the problems of single matching dimension and insufficient accuracy caused by only depending on single tag matching are avoided, and the accuracy and comprehensiveness of similar object retrieval are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, and in particular, to an object retrieval method and device, electronic equipment and storage medium. BACKGROUND

[0002] With the development of short videos and social media, network celebrities have become an important channel for brand marketing, and efficient retrieval and accurate recommendation of similar network celebrities have become a key problem in the industry.

[0003] The existing similar network celebrity retrieval method mainly relies on manual screening or simple sorting mechanism based on label classification. The brand party usually manually finds the candidate according to the basic data such as the number of fans, the number of plays, and the number of interactions provided by the platform, combined with experience, or performs single index matching through the label system of the platform. However, this method has poor matching accuracy, and only supports similar retrieval of a single network celebrity, which is difficult to meet the diversified application requirements such as group substitution and combination placement. SUMMARY

[0004] The present application provides an object retrieval method, device, electronic equipment and storage medium to solve the defects in the prior art.

[0005] The present application provides an object retrieval method, comprising the following steps: Obtaining a retrieval target, the retrieval target comprising at least one seed object; Extracting semantic features of various seed objects from the description text of various seed objects; According to the semantic features of various seed objects, recalling a candidate object set similar in semantics to the retrieval target from a candidate object library; In the case where the data amount of the candidate object library is less than a threshold value, sorting the recalled candidate object set by using the attribute labels of various seed objects; According to the sorting result, determining a similar object recommendation list corresponding to the retrieval target from the candidate object set.

[0006] According to the object retrieval method provided by the present application, the candidate object set similar in semantics to the retrieval target is recalled from the candidate object library according to the semantic features of various seed objects, comprising: According to the similarity between the semantic features of each candidate object in the candidate object library and the semantic features of various seed objects, determining the semantic similarity between each candidate object in the candidate object library and the retrieval target; Adding the candidate object with a semantic similarity greater than a threshold value to the candidate object set as a rough screening object.

[0007] According to the object retrieval method provided by the application, the semantic similarity between each candidate object in the candidate object library and the retrieval target is determined according to the similarity between the semantic features of each candidate object in the candidate object library and the semantic features of various sub-objects, and the method comprises the following steps: In the case that the seed objects in the retrieval target are all positive seed objects, the maximum value of the similarity between the semantic features of each candidate object and the semantic features of various sub-objects is taken as the semantic similarity between the corresponding candidate object and the retrieval target, or the average value of the similarity between the semantic features of each candidate object and the semantic features of various sub-objects is taken as the semantic similarity between the corresponding candidate object and the retrieval target. In the case that the seed objects in the retrieval target include positive seed objects and negative seed objects, the semantic similarity between each rough screening object and the retrieval target is determined according to the difference between the semantic similarity between each rough screening object and the positive seed object and the semantic similarity between each rough screening object and the negative seed object.

[0008] According to the object retrieval method provided by the application, the candidate object set recalled is sorted by using the attribute labels of various sub-objects, and the method comprises the following steps: The label similarity between each rough screening object and the retrieval target is determined according to the matching degree between each rough screening object in the candidate object set and various sub-objects in attribute labels. The label similarity and the semantic similarity of each rough screening object are fused to obtain the comprehensive similarity between each rough screening object and the retrieval target. The candidate object set recalled is sorted according to the comprehensive similarity of each rough screening object.

[0009] According to the object retrieval method provided by the application, the label similarity between each rough screening object and the retrieval target is determined according to the matching degree between each rough screening object in the candidate object set and various sub-objects in attribute labels, and the method comprises the following steps: The intersection quantity and the union quantity between the attribute label set of each rough screening object and the attribute label set of various sub-objects are calculated, wherein the attribute label set is a set composed of attribute labels with a frequency greater than a preset frequency among all attribute labels of the object; The label matching degree between each rough screening object and various sub-objects is determined according to the ratio of the intersection quantity corresponding to each rough screening object to the union quantity. The label similarity between each rough screening object and the retrieval target is determined according to the label matching degree between each rough screening object and various sub-objects. According to the object retrieval method provided by the application, the label similarity of each rough screening object with the retrieval target is determined according to the label matching degree of each rough screening object with various sub-objects, and the method comprises the following steps: In the case that the seed objects in the retrieval target are all positive seed objects, the maximum value of the label matching degree of each rough screening object with various sub-objects is taken as the label similarity of the corresponding rough screening object with the retrieval target, or the average value of the label matching degree of each rough screening object with various sub-objects is taken as the label similarity of the corresponding rough screening object with the retrieval target. In the case that the seed objects in the retrieval target include positive seed objects and negative seed objects, the label similarity of each rough screening object with the retrieval target is determined according to the difference between the label similarity of each rough screening object with the positive seed objects and the label similarity of each rough screening object with the negative seed objects.

[0010] According to the object retrieval method provided by the application, the label similarity of each rough screening object with the retrieval target is determined according to the label matching degree of each rough screening object with various sub-objects, and the method comprises the following steps: The semantic weight coefficient corresponding to the semantic similarity of each rough screening object and the label weight coefficient corresponding to the label similarity of each rough screening object are determined. The semantic similarity of each rough screening object is corrected according to the semantic weight coefficient of each rough screening object and the label weight coefficient of each rough screening object to obtain the corrected semantic similarity of each rough screening object. The label similarity of each rough screening label is corrected according to the label weight coefficient of each rough screening object to obtain the corrected label similarity of each rough screening object. The comprehensive similarity of each rough screening object is obtained by fusing the corrected semantic similarity and the corrected label similarity of each rough screening object.

[0011] According to the object retrieval method provided by the application, the candidate object set similar in semantics to the retrieval target is recalled from the candidate object library, and then the method further comprises the following steps: In the case that the data quantity of the candidate object library is greater than or equal to the threshold value, the retrieval target, the attribute label and the semantic feature of each rough screening object in the candidate object set are input into the sorting model to obtain the similar object recommendation list corresponding to the retrieval target output by the sorting model. The sorting model is trained based on the historical retrieval target of each candidate object in the candidate object library, the semantic feature, the attribute label and the historical recommendation list corresponding to the historical retrieval target of each candidate object.

[0012] The application further provides an object retrieval device comprising the following modules: An acquisition unit is configured to acquire a search target, the search target comprising at least one seed object; An extraction unit is configured to extract semantic features of various sub-objects from description texts of the various sub-objects; A recall unit is configured to recall a set of candidate objects similar in semantics to the search target from a candidate object library according to the semantic features of the various sub-objects; A sorting unit is configured to sort the set of recalled candidate objects according to attribute labels of the various sub-objects when a data volume of the candidate object library is less than a threshold value; A determination unit is configured to determine a list of similar objects corresponding to the search target from the set of candidate objects according to a sorting result.

[0013] The application further provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the object retrieval method according to any one of the above when executing the program.

[0014] The application further provides a non-transitory computer-readable storage medium having a computer program stored thereon, and the computer program is executable on a processor to implement the object retrieval method according to any one of the above.

[0015] The application further provides a computer program product comprising a computer program, and the computer program is executable on a processor to implement the object retrieval method according to any one of the above.

[0016] The object retrieval method, device, electronic device, and storage medium provided by the application realize a multi-dimensional and staged similar object retrieval mechanism that takes into account deep semantic matching and key business index matching by first performing large-scale recall using semantic features that can deeply represent the intrinsic characteristics of objects and then performing supplementary sorting in combination with structured attribute labels. Since the application integrates deep semantics of unstructured texts and objectified indexes of structured labels, the problem of single matching dimension and insufficient accuracy caused by relying on only single label matching is avoided, and the precision and comprehensiveness of similar object retrieval are significantly improved. In addition, the application supports alternative object searching for a single seed object and group expansion retrieval for multiple seed objects, and flexibly meets diversified application requirements. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the application or the prior art, the following will briefly introduce the drawings needed in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0018] Figure 1 is a flowchart of the object retrieval method provided by the present application.

[0019] Figure 2 is a flowchart of another object retrieval method provided by the present application.

[0020] Figure 3 is a structural diagram of the object retrieval device provided by the present application.

[0021] Figure 4 is a structural diagram of the electronic device provided by the present application. DETAILED DESCRIPTION

[0022] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0023] With the rapid development of short videos and social media, network celebrities have become an important channel for brand promotion, product marketing and user reach. In practical applications, brand parties not only need to select cooperative network celebrities that meet their needs, but also need to quickly find potential candidates similar to target network celebrities in characteristics for alternative screening, audience expansion and multi-channel promotion. Therefore, efficient retrieval and accurate recommendation of similar network celebrities have become a key problem in the industry.

[0024] Currently, similar network celebrity retrieval mainly relies on manual screening. Brand parties usually manually find candidates according to platform-provided basic data such as fan count, play count, interaction count and sales record, combined with experience. Although some platforms provide tag classification and simple sorting mechanisms, they can only match based on a single indicator or a small number of dimensions, lack comprehensive modeling of network celebrity content style, audience structure, sales ability and other multi-dimensional characteristics, resulting in insufficient comprehensive matching. At the same time, this method generally only supports single-network celebrity similarity retrieval and does not have the ability to retrieve similar multiple network celebrities or positive and negative multiple network celebrities, with low overall efficiency and insufficient accuracy.

[0025] To this end, the present application provides an object retrieval method, aiming to perform candidate object recall by extracting semantic features of seed objects, and to perform ranking optimization on the recall results in combination with attribute labels, to construct a multi-dimensional fusion retrieval mechanism, support single object and multi-object similarity retrieval, realize efficient, accurate and personalized similar object recommendation, and meet the actual needs of brands in diversified marketing scenarios. Among them, the object retrieval method provided by the present application can be applied to similar network celebrity retrieval scenarios, similar product recommendation scenarios, and other scenarios that require similar matching based on object features, such as similar content recommendation, similar user search, similar enterprise retrieval, etc. The present application does not make specific limitations on specific application scenarios. In order to explain the technical solutions of the present application, the following embodiments are described by taking the application in similar network celebrity retrieval scenarios as an example.

[0026] wherein, Figure 1 is a flowchart of the object retrieval method provided by the present application, as shown in Figure 1 The method comprises steps 110, 120, 130, 140 and 150.

[0027] Step 110, obtaining a retrieval target, the retrieval target comprising at least one seed object.

[0028] Here, the retrieval target can be understood as a defined set of similar objects that the user wants to find, which is used to clearly input the retrieval task. The retrieval target comprises at least one seed object. The seed object is one or more specific instances specified by the user as a similarity comparison reference. For example, in the similar network celebrity retrieval scenario, a seed object can be a specific network celebrity. The retrieval target comprises at least one seed object, meaning that the present embodiment supports both single object similarity retrieval and multi-object similarity retrieval.

[0029] When the retrieval target only includes one seed object, its application scenario can be to find alternative or replacement candidates for a network celebrity who has cooperated or intends to cooperate. When the retrieval target includes multiple seed objects, its application scenario can be to find more people with the same characteristics (e.g. similar style, high audience overlap) for a group of network celebrities with the same characteristics, to perform group replacement or expand the marketing matrix.

[0030] As an optional embodiment, the retrieval target can be obtained by receiving user input through a user interface. For example, the user can specify one or more seed objects by inputting the ID, nickname of the network celebrity, or pasting the personal homepage link in the retrieval interface. After receiving the user's input, the specified one or more network celebrities are determined as the seed objects of the present retrieval task, and are packaged into the retrieval target.

[0031] Step 120, extracting semantic features of various sub-objects from the description text of various sub-objects.

[0032] Specifically, the description text can be understood as unstructured text information that can describe the intrinsic characteristics of the seed object, which can be one or a combination of the following: a network celebrity's personal profile, the title of historical published content, the text of historical published content, the subtitle text of video content, user comments on the content, etc. The description text can comprehensively reflect the content style, professional field, language habit, and other deep information of the network celebrity.

[0033] The semantic feature can be understood as a high-dimensional vector representation obtained by analyzing and transforming the description text. The vector represents the deep semantic information of the description text in the mathematical space, rather than just the literal keywords.

[0034] As an optional embodiment, a pre-trained deep learning language model (such as the BERT model) can be used to extract semantic features. Specifically, the description text of various sub-objects obtained is input into the language model, and the language model outputs a fixed-dimensional vector as the semantic feature of the seed object. In order to improve the performance in a specific field, the language model can also be a model obtained by fine-tuning on general corpus using professional corpus of network celebrities (such as various network celebrity profiles, hit scripts, etc.), so that the extracted semantic features can more accurately capture the unique expressions and styles in the field of network celebrities.

[0035] Step 130, according to the semantic features of various sub-objects, recalling a set of candidate objects similar in semantic to the retrieval target from the candidate object library.

[0036] Specifically, the candidate object library can be understood as a pre-stored mass object information. In this embodiment, the candidate object library can be a database pre-stored with mass network celebrity information. Each object in the database has its own semantic feature vector extracted and stored in a similar way as step 120.

[0037] Recalling refers to filtering out a subset that is initially matched in semantics with the retrieval target from the mass candidate object library, i.e., the candidate object set. The candidate object set includes multiple objects, which can be understood as objects similar to the retrieval target after preliminary filtering.

[0038] As an optional embodiment, the recall process can be performed based on vector similarity calculation. Specifically, when the retrieval target contains multiple seed objects, semantic feature vectors of the multiple seed objects are first aggregated to form a central vector representing the overall semantic direction of the retrieval target. Subsequently, an approximate nearest neighbor search algorithm is used to retrieve the Top-K candidate objects with the highest cosine similarity or the closest Euclidean distance to the central vector from the candidate object library, thereby obtaining the candidate object set.

[0039] It should be noted that the embodiment first recalls the candidate object set through semantic features, which can maximize the scope of all potential objects that match the seed objects in terms of core characteristics such as content style and domain tone, avoiding missing high-quality objects that are highly matched in content essence due to premature filtering of hard indicators such as fan quantity and region.

[0040] Step 140, in the case where the data quantity of the candidate object library is less than the threshold value, the recalled candidate object set is sorted using the attribute labels of the various seed objects.

[0041] Considering that training a complex sorting model for a candidate object library with a small data quantity may have overfitting risks or may not be cost-effective, the embodiment formulates a more lightweight, efficient and stable strategy based on attribute label sorting.

[0042] The attribute label can be understood as structured and discrete feature information describing the object, which can be one or a combination of the following: the fan level of a network celebrity, the average play quantity, the interaction rate, the affiliated MCN agency, the product category, the product carrying capacity level, the joint brand, the fan portrait label (such as gender, age, and regional distribution), etc.

[0043] As an optional embodiment, sorting the candidate object set using attribute labels can be achieved by calculating the similarity score of each candidate object and the seed object in the attribute label. For example, different weights can be set for different attribute labels, and then the comprehensive score is calculated by weighting. For numerical labels (such as fan quantity), the normalized difference can be calculated; for category labels (such as product carrying category), the overlap degree can be calculated. In this way, on the basis of semantic similarity, objects that are also more matched in key business indicators (such as product carrying capacity and fan level) can be further filtered out.

[0044] It should be noted that if a first round of rough screening is performed using a simple attribute label (such as the number of fans and region) to reduce the calculation scale, and then a complex semantic similarity calculation is performed on the screened subset, a large number of potential objects that do not completely meet the business indicators but are excellent matches in content style may be irreversibly excluded too early, resulting in low recall rate and poor diversity of the final recommendation result. The embodiment prioritizes ensuring the high relevance of the recall set in content semantics, and then performs fine sorting in key business indicators using attribute labels, ensuring that each object in the final recommendation list is highly matched with the seed object in the core style. On this basis, quantitative evaluation and reordering are performed according to the attribute label, thereby ensuring the accuracy of the recommendation while achieving high recall rate of all potential similar objects, and achieving the technical effect of simultaneous improvement of recommendation accuracy and recall rate.

[0045] Step 150, determining a similar object recommendation list corresponding to the search target from the candidate object set according to the sorting result.

[0046] Specifically, the similar object recommendation list is the final output of this search task, which can be an object list arranged from high to low according to the similarity with the search target.

[0047] As an optional embodiment, the top N objects in the sorting result can be directly intercepted as the final similar object recommendation list and presented to the user through the user interface. The list can display the network celebrity's avatar, nickname, key attribute label, and comprehensive similarity score with the search target for the user to refer to and select.

[0048] The object search method provided in the embodiment realizes a multi-dimensional and phased similar object search mechanism that takes into account deep semantic matching and key business indicator matching by first using semantic features that can deeply represent the intrinsic characteristics of objects for large-scale recall, and then combining structured attribute labels for supplementary sorting. Since the embodiment combines the deep semantics of unstructured text and the objectified indicators of structured labels, it avoids the problem of single matching dimension and insufficient accuracy caused by relying solely on single label matching, significantly improving the accuracy and comprehensiveness of similar object search. In addition, the embodiment supports alternative object search for a single seed object and group expansion search for multiple seed objects, flexibly meeting diverse application needs.

[0049] Based on the above embodiment, step 130 recalls a candidate object set similar in semantics to the search target from the candidate object library according to the semantic features of various sub-objects, including: Step 131, determining the semantic similarity between each candidate object in the candidate object library and the search target according to the similarity between the semantic features of each candidate object in the candidate object library and the semantic features of various sub-objects.

[0050] Here, the semantic similarity can be understood as a quantitative value, which is used to accurately represent the matching degree of the candidate object with the search target in the deep semantic level of content style, field tone, etc. The higher the value, the stronger the similarity in the semantic level.

[0051] As an optional embodiment, when the search target includes multiple seed objects, first, the semantic feature vectors of the multiple seed objects are aggregated to obtain a center vector that can represent the overall semantic direction of the search target. Then, for any candidate object in the candidate object library, the pre-stored semantic feature vector of the candidate object is obtained. Subsequently, the semantic similarity between the candidate object and the search target is determined by a vector similarity calculation function. The semantic similarity can be determined by calculating the cosine similarity between the two vectors.

[0052] Step 132, adding the candidate object with a semantic similarity greater than a threshold value to the candidate object set as a rough screening object.

[0053] Here, the threshold value is a pre-set, configurable value, such as 0.8, 0.85, or 0.9. The threshold value is used as a quality control threshold to screen candidate objects with high enough semantic similarity.

[0054] After calculating the semantic similarity between each candidate object in the candidate object library and the search target, the similarity value is compared with the pre-set threshold value. If the semantic similarity of a certain candidate object is greater than the threshold value, it indicates that the candidate object has a high degree of matching with the search target in the semantic level of content style, field tone, etc., and thus the candidate object can be determined as a rough screening object and added to the candidate object set. In this way, it can be ensured that all candidate objects entering the next ranking stage have a high degree of relevance to the seed object in the content semantic level.

[0055] The present embodiment introduces a controllable threshold value as a screening standard, achieving precise and standardized control of the recall process. Compared with the Top-K method alone, the present embodiment can ensure that all rough screening objects entering the candidate object set meet the minimum semantic relevance standard, avoiding the possibility of including objects with low relevance into the candidate set in the case of sparse similar objects, thereby ensuring the overall quality of the recall set, providing higher quality input for the subsequent fine ranking step, and ultimately improving the overall precision rate of the similar object recommendation list.

[0056] Based on any of the above embodiments, the semantic similarity between each candidate object in the candidate object library and the search target is determined according to the similarity between the semantic features of each candidate object in the candidate object library and the semantic features of the seed objects, comprising: In the case that the seed objects in the retrieval target are all positive seed objects, the maximum value of the similarity between the semantic feature of each candidate object and the semantic feature of each seed object is taken as the semantic similarity between the corresponding candidate object and the retrieval target; or the average value of the similarity between the semantic feature of each candidate object and the semantic feature of each seed object is taken as the semantic similarity between the corresponding candidate object and the retrieval target. In the case that the seed objects in the retrieval target include positive seed objects and negative seed objects, the semantic similarity between each coarse screening object and the retrieval target is determined according to the difference between the semantic similarity between each coarse screening object and the positive seed object and the semantic similarity between each coarse screening object and the negative seed object.

[0057] It is considered that, in the case that the seed objects in the retrieval target are all positive seed objects, the target of the retrieval task is to converge to the feature space defined by one or more positive targets, and therefore the focus in determining the semantic similarity between each candidate object in the candidate object library and the retrieval target is the degree of approximation of the candidate object to the group of positive seed objects. In the case that the seed objects in the retrieval target include positive seed objects and negative seed objects, the target of the retrieval task is to approach the positive target while avoiding the feature space defined by the negative target, and therefore the focus in determining the semantic similarity between each candidate object in the candidate object library and the retrieval target is the comprehensive performance of the candidate object between positive attraction and negative repulsion. That is, the technical problems and optimization targets corresponding to the case that the seed objects in the retrieval target are all positive seed objects and the case that the seed objects in the retrieval target include positive seed objects and negative seed objects are different.

[0058] The positive seed object can be understood as a reference object that the user explicitly specifies to find similar objects. For example, the user wants to find other people who are similar in style to network celebrities A, B and C, at which time the network celebrities A, B and C are all positive seed objects. The negative seed object can be understood as a reference object that the user explicitly specifies to exclude and avoid similar objects. For example, the user wants to find a type similar to the makeup celebrities A and B (positive seed objects) in style, but wants to explicitly exclude the funny style celebrity C (negative seed object).

[0059] In the case that the seed objects in the retrieval target are all positive seed objects, the embodiment provides at least two optional ways to calculate the semantic similarity: One is to take the maximum value of the similarity between the semantic feature of each candidate object and the semantic feature of each seed object as the semantic similarity between the corresponding candidate object and the retrieval target. The consideration of using this way to calculate the semantic similarity is to meet the retrieval requirement of being highly similar to any member of the seed object group, and this strategy tends to expand the diversity of the recall result, which is suitable for the scene where the seed objects represent multiple different aspects of the desired style.

[0060] Specifically, for any candidate object in the candidate object library, the cosine similarity between its semantic feature vector and the semantic feature vector of each positive seed object in the retrieval target is calculated respectively, obtaining a set of similarity values. Then, the maximum one is selected from the set of similarity values as the final semantic similarity between the candidate object and the retrieval target.

[0061] Another is to take the average of the similarity between the semantic features of each candidate object and the semantic features of various sub-objects as the semantic similarity between the corresponding candidate object and the retrieval target. The consideration of using this way to calculate the semantic similarity is to find similar objects that can comprehensively reflect the overall characteristics of the seed object group. This strategy tends to improve the central focus of the recall result and is suitable for scenarios where the seed objects collectively point to a very clear and concentrated style.

[0062] Specifically, for any candidate object in the candidate object library, the cosine similarity between its semantic feature vector and the semantic feature vector of each positive seed object in the retrieval target is calculated respectively, obtaining a set of similarity values. Then, the maximum one is selected from the set of similarity values as the final semantic similarity between the candidate object and the retrieval target.

[0063] In the case where the seed objects in the retrieval target include positive seed objects and negative seed objects, the semantic similarity between each candidate object and the retrieval target is determined according to the difference between the semantic similarity of each candidate object and the positive seed object and the semantic similarity of each candidate object and the negative seed object. The consideration of using this way to calculate the semantic similarity is to realize a constrained similarity retrieval, that is, to maximize the similarity with the positive target while minimizing the similarity with the negative target, thereby realizing more fine and personalized retrieval control.

[0064] As an optional embodiment, the average semantic similarity of the candidate object with all positive seed objects can be calculated first . Then, the average semantic similarity of the candidate object with all negative seed objects is calculated . Finally, the final semantic similarity is calculated by a fusion formula , which can be: wherein, is a weight coefficient for controlling the influence balance of positive and negative constraints. The intuitive meaning of this formula is that an ideal candidate object should be as close as possible to the semantics of the positive seed group and as far as possible from the semantics of the negative seed group. When the value increases, it means that the weight of the consideration of similarity with the positive seed object is increased, while the weight of the consideration of similarity with the negative seed object is correspondingly reduced.

[0065] Based on any of the above embodiments, step 140 sorts the recalled candidate object set using the attribute labels of various sub-objects, including: Step 141, according to the matching degree of each coarse screening object in the candidate object set with various sub-objects in the attribute labels, determines the label similarity of each coarse screening object with the search target.

[0066] Specifically, the label similarity is a quantitative value that represents the matching degree of a coarse screening object with the search target in terms of structured attributes such as fan size, sales ability, and audience portrait.

[0067] As an optional embodiment, the process of determining the label similarity can be a weighted sum of the matching degrees of multiple attribute labels. Specifically, when the search target includes multiple seed objects, the attribute labels can be aggregated first, for example, the average value is taken for numerical labels (such as fan size), and the union is taken for categorical labels (such as sales categories). Subsequently, for any coarse screening object in the candidate object set, the matching degree of each attribute label of the coarse screening object with the corresponding attribute label of the search target is calculated, for example, for numerical labels (such as fan size, interaction rate), the matching degree can be calculated according to the normalized difference; for categorical labels (such as sales categories, content fields), the matching degree can be calculated using Jaccard similarity coefficient. Finally, different weights are assigned to different attribute labels, and the final label similarity of the coarse screening object with the search target is obtained by weighted sum. The weights can be determined according to specific business requirements, or can be obtained by adaptive learning based on historical user behavior or through a machine learning model, which is not limited in this embodiment.

[0068] Step 142, fuse the label similarity and the semantic similarity of each coarse screening object to obtain the comprehensive similarity of each coarse screening object with the search target.

[0069] Specifically, the comprehensive similarity is the final evaluation score obtained by organically combining the similarity evaluation results of two different dimensions (i.e. semantic level and attribute label level) of the coarse screening object. The comprehensive similarity can be calculated using weighted linear fusion, or the semantic similarity and the label similarity can be input as features into a machine learning model (such as a logistic regression model, a gradient boosting decision tree model, etc.) to train a model that can predict the comprehensive similarity, which is not limited in this embodiment.

[0070] Step 143, according to the comprehensive similarity of each coarse screening object, sorts the recalled candidate object set.

[0071] After the comprehensive similarity of each rough screening object in the candidate object set is calculated, all rough screening objects in the candidate object set can be ranked in descending order according to their comprehensive similarity scores. The rough screening object with the highest score is ranked first, indicating that it has the highest overall similarity to the search target after combining the content semantics and attribute label dimensions. The ranking result is the basis for the final recommendation list.

[0072] The embodiment realizes comprehensive evaluation and ranking of similarity from the two dimensions of semantic level and attribute label by respectively calculating semantic similarity and label similarity and then performing weighted fusion through an adjustable fusion coefficient. Since the embodiment not only considers the deep semantic association between the inherent characteristics of objects but also takes into account the matching degree of the objectified indicators and allows the user to dynamically adjust the importance of the two dimensions according to the requirements, the accuracy and business relevance of the ranking result are significantly improved, the one-sidedness caused by single-dimensional ranking is avoided, a more comprehensive, flexible, and highly customized ranking result of similar objects can be provided for the user, and the personalized adjustment requirements of the user for the similarity emphasis in different business scenarios are met.

[0073] Based on any of the above embodiments, step 141 determines the label similarity between each rough screening object and the search target according to the matching degree between the attribute label set of each rough screening object and the attribute label set of each seed object, including: Step 1411 calculates the intersection quantity and union quantity between the attribute label set of each rough screening object and the attribute label set of each seed object. The attribute label set is a set of attribute labels with a frequency greater than a preset frequency among all attribute labels of an object.

[0074] Here, the attribute label set can be understood as a high-frequency label set that represents the core and stable characteristics of an object (rough screening object or seed object) selected from all attribute labels of the object. When determining the attribute label set, the preset frequency is introduced as a screening threshold, and the purpose is to filter out occasional, non-core, or noisy labels.

[0075] For example, the category type attribute labels associated with seed object A in historical content include: “makeup” (occurred 15 times), “skincare” (occurred 12 times), “review” (occurred 10 times), and “daily Vlog” (occurred 3 times). If the preset frequency is 5, the attribute label set of the seed object A is {makeup, skincare, review}. Similarly, for rough screening object B, the labels include “makeup” (occurred 20 times), “outfit” (occurred 15 times), “review” (occurred 8 times), and “food” (occurred 2 times), and the attribute label set of the rough screening object B is {makeup, outfit, review}.

[0076] Subsequently, the number of intersections and the number of unions between the attribute label set of the rough screening object B and the attribute label set of the seed object A are calculated. The intersection between the two is {makeup, review}, so the number of intersections is 2; the union is {makeup, skin care, review, dressing}, so the number of unions is 4.

[0077] Step 1412, according to the ratio of the intersection number to the union number corresponding to each rough screening object, determine the label matching degree of each rough screening object with each seed object respectively.

[0078] Step 1413, according to the label matching degree of each rough screening object with each seed object respectively, determine the label similarity of each rough screening object with the search target.

[0079] Specifically, the label matching degree is used to quantify the degree of coincidence between a single rough screening object and a single seed object in the core attribute label layer. The larger the ratio of the intersection number to the union number, the higher the similarity of the two objects in the core attribute characteristics, and the higher the determined label matching degree.

[0080] After determining the label matching degree of a single rough screening object and a single seed object, if there are multiple seed objects in the search target, the label matching degrees of the rough screening object and each seed object need to be aggregated to determine the final label similarity of the rough screening object and the entire search target.

[0081] As an optional embodiment, when the search target contains seed object A and seed object C, the label matching degree 1 of rough screening object B with seed object A and the label matching degree 2 of rough screening object B with seed object C can be calculated first. Then, the average of the label matching degree 1 and the label matching degree 2 can be taken as the label similarity of rough screening object B with the search target, or the larger value of the label matching degree 1 and the label matching degree 2 can be taken as the label similarity of rough screening object B with the search target.

[0082] Based on any of the above embodiments, according to the label matching degree of each rough screening object with each seed object respectively, the label similarity of each rough screening object with the search target is determined, including: In the case where the seed objects in the search target are all positive seed objects, the maximum value of the label matching degree of each rough screening object with each seed object is taken as the label similarity of the corresponding rough screening object with the search target; or the average value of the label matching degree of each rough screening object with each seed object is taken as the label similarity of the corresponding rough screening object with the search target; In the case where the seed objects in the search target include positive seed objects and negative seed objects, according to the difference between the label similarity of each rough screening object with the positive seed object and the label similarity of each rough screening object with the negative seed object, the label similarity of each rough screening object with the search target is determined.

[0083] If all seed objects in the retrieval target are positive seed objects, the goal of the retrieval task is to find objects in the coarse-screened objects that are as close as possible to the feature space defined by the group of positive seed objects. Therefore, when determining the label similarity between each coarse-screened object and the retrieval target, the focus is on the aggregated similarity of the coarse-screened object and the group of positive seed objects in terms of attribute labels. If the seed objects in the retrieval target include both positive and negative seed objects, the goal of the retrieval task is to find objects with features similar to positive seed objects while ensuring that their features are significantly different from those of negative seed objects. Therefore, when determining the label similarity between each coarse-screened object and the retrieval target, the focus is on the comprehensive performance of the coarse-screened object in terms of both approaching the features of positive seed objects and moving away from the features of negative seed objects.

[0084] In this context, positive seed objects can be understood as benchmark objects that the user explicitly specifies they want to find that are similar to them. For example, if a user wants to find other candidates whose styles are similar to those of online influencers A, B, and C, then influencers A, B, and C are all positive seed objects. Negative seed objects can be understood as benchmark objects that the user explicitly specifies they want to exclude or avoid. For example, if a user wants to find beauty influencers A and B (positive seed objects) with similar styles, but wants to explicitly exclude comedic influencer C (negative seed object), they want to find similar types.

[0085] When all seed objects in the search target are positive seed objects, this embodiment provides at least two optional methods for calculating tag similarity: One approach is to use the maximum value of the tag matching scores between each coarsely screened object and various sub-objects as the tag similarity between the corresponding coarsely screened object and the search target. The consideration for using this method to calculate tag similarity is that it is applicable to search scenarios with "OR" logic, that is, as long as a coarsely screened object is highly similar to any positive seed object, it should be considered a strongly related object, thereby expanding the diversity of search results and finding objects that match different branch styles in the seed group.

[0086] Specifically, for any coarsely screened object, the tag matching degree between its attribute tag set and the attribute tag set of each positive seed object in the search target is calculated to obtain a set of tag matching degrees. Then, the largest tag matching degree in this set is selected as the final tag similarity between the coarsely screened object and the search target.

[0087] Another way is to take the average of the matching degrees of each coarse screening object with the labels of various sub-objects as the label similarity of the corresponding coarse screening object with the search target. The consideration of this way of calculating the label similarity is that it is suitable for the search scene of the "and" logic, that is, the similar objects that are expected to be found should comprehensively reflect the common characteristics of all positive seed objects, so as to recommend core objects that are more consistent with the overall style of the seed group.

[0088] Specifically, for any coarse screening object, the label matching degrees between its attribute label set and the attribute label set of each positive seed object in the search target are calculated respectively to obtain a set of label matching degrees. Then, the average of the set of label matching degrees is taken as the final label similarity of the coarse screening object with the search target.

[0089] In the case where the seed objects in the search target include positive seed objects and negative seed objects, the label similarity of each coarse screening object with the search target is determined according to the difference between the label matching degree of each coarse screening object with the positive seed object and the label matching degree of each coarse screening object with the negative seed object. The consideration of this way of calculating the label similarity is that it can integrate the two search targets of "tending to positive" and "moving away from negative" into a unified scoring framework. By taking the similarity with the positive object as a gain term and the similarity with the negative object as a penalty term, it can quantitatively evaluate whether a coarse screening object meets both conditions, so as to accurately screen out objects that meet complex and exclusionary search intentions.

[0090] As an optional embodiment, the average label matching degree of the coarse screening object with all positive seed objects can be calculated first . Then, the average label matching degree of the coarse screening object with all negative seed objects is calculated . Finally, the final label similarity is calculated by a fusion formula , which can be: wherein, is a weight coefficient for controlling the balance of the influence of positive and negative constraints. The intuitive meaning of this formula is that an ideal coarse screening object should be as close as possible to the semantics of the positive seed group and as far as possible from the semantics of the negative seed group. When the value increases, it means that the weight of the consideration of the similarity with the positive seed object is increased, and the weight of the consideration of the similarity with the negative seed object is correspondingly reduced when calculating the final label similarity.

[0091] Based on any of the above embodiments, step 142 fuses the label similarity and the semantic similarity of each coarse screening object to obtain the comprehensive similarity of each coarse screening object, including: Step 1421, determine the semantic weight coefficient corresponding to the semantic similarity of each rough screening object, and the label weight coefficient corresponding to the label similarity of each rough screening object.

[0092] Specifically, the semantic weight coefficient can be understood as a value for quantifying the importance of the semantic similarity in the final comprehensive evaluation. The larger the coefficient, the more emphasis on the matching degree of the internal characteristics such as content style and field tone between objects. The semantic weight coefficient can be determined by receiving user settings on the interactive interface. For example, a slider with a range of [0, 1] can be provided, one end of the slider is marked as "emphasis on content style" and the other end is marked as "emphasis on key business indicators". The position of the slider dragged by the user determines the value of the semantic weight coefficient.

[0093] The label weight coefficient can be understood as a value for quantifying the importance of the label similarity in the final comprehensive evaluation. The larger the coefficient, the more emphasis on the matching degree of external quantitative indicators such as fan quantity, goods carrying capacity, and audience portrait between objects. The label weight coefficient can be determined in conjunction with the semantic weight coefficient. For example, the sum of the label weight coefficient and the semantic weight coefficient can be set to 1. When the user sets the semantic weight coefficient, the label weight coefficient is automatically determined.

[0094] Step 1422, according to the semantic weight coefficient of each rough screening object and the label weight coefficient of each rough screening object, the semantic similarity of each rough screening object is corrected to obtain the corrected semantic similarity of each rough screening object.

[0095] Considering that the semantic similarity of each rough screening object is a purely technical index calculated, it may not completely align the value size with the relevance of the final search target due to not combining the user's current business preferences. If the final comprehensive similarity is directly determined according to the semantic similarity, the final ranking result may not fully reflect the user's personalized weight preference, reducing the business relevance of the recommendation result.

[0096] Further, considering that the semantic weight coefficient can correct the basic contribution value of the semantic similarity from the importance allocation level, and the label weight coefficient can correct the relative influence of the semantic similarity from the balance constraint level, therefore, according to the semantic weight coefficient of each rough screening object and the label weight coefficient of each rough screening object, the semantic similarity of each rough screening object is corrected to obtain the corrected semantic similarity of each rough screening object. The corrected semantic similarity can be understood as a measurement value that more reflects the real contribution in a specific business scenario after the original semantic similarity is nonlinearly adjusted according to the user's double weight preference.

[0097] As an optional embodiment, the semantic similarity of each rough screening object can be modified by multiplying the semantic similarity with a correction factor composed of a semantic weight coefficient and a label weight coefficient, to obtain a modified semantic similarity of each rough screening object.

[0098] Step 1423, modifying the label similarity of each rough screening label according to the label weight coefficient of each rough screening object, to obtain a modified label similarity of each rough screening object.

[0099] It is considered that the label similarity of each rough screening object is a comprehensive index calculated based on multiple discrete attributes, which may not be directly comparable with other dimensions (such as the semantic dimension) due to not considering the importance of the user to the overall dimension. If the final comprehensive similarity is directly determined according to the label similarity, the contribution of the index may not meet the user's expectation, affecting the rationality of the final ranking.

[0100] Further, considering that the label weight coefficient can modify the overall contribution of the label similarity from the importance allocation level, the embodiment modifies the label similarity of each rough screening object according to the label weight coefficient of each rough screening object, to obtain a modified label similarity of each rough screening object. The modified label similarity can be understood as a measurement value that can be fairly integrated with other dimensions after linear scaling of the original label similarity according to the importance preference set by the user.

[0101] As an optional embodiment, the label similarity of each rough screening object can be modified by multiplying the label similarity with the label weight coefficient, to obtain a modified label similarity of each rough screening object. It should be noted that the label similarity itself integrates the quantitative scores of multiple specific business indicators, and its internal physical meaning is relatively independent and clear. Its role in the final fusion process is mainly to provide an evaluation orthogonal to the semantic dimension. Therefore, only the contribution scaling of the label similarity itself according to its own label weight coefficient is needed, without introducing a semantic weight coefficient for cross-correction, which can maintain the clarity and independence of the modification logic of each dimension, and avoid unnecessary coupling and complex interaction effects.

[0102] Step 1424, fusing the modified semantic similarity and the modified label similarity of each rough screening object to obtain a comprehensive similarity of each rough screening object.

[0103] Specifically, the modified semantic similarity has reflected the content style matching degree under the user preference, the modified label similarity has reflected the business index matching degree under the user preference, and then the two are fused, so that the two similarity measurements of different natures which have been independently adjusted by weights can be combined fairly, and the comprehensive similarity obtained can more accurately reflect the comprehensive evaluation standard of the user under the specific business preference. Wherein, the comprehensive similarity of each rough screening object can be obtained by performing addition operation on the modified semantic similarity and the modified label similarity.

[0104] Based on any of the above embodiments, the set of candidate objects similar to the retrieval target semantics is recalled from the candidate object library, and then further includes: In the case where the data quantity of the candidate object library is greater than or equal to the threshold value, the retrieval target, the attribute label and the semantic feature of each rough screening object in the set of candidate objects are input into the ranking model to obtain the similar object recommendation list corresponding to the retrieval target output by the ranking model. The ranking model is trained based on the historical retrieval target of each candidate object in the candidate object library, the semantic feature and the attribute label of each candidate object, and the historical recommendation list corresponding to the historical retrieval target.

[0105] Specifically, if the data quantity of the candidate object library is greater than or equal to the threshold value, it indicates that the data quantity is sufficient to support the effective training of a complex model, and can avoid the overfitting problem caused by data sparsity. At this time, the data in the candidate object library is sufficient to train a ranking model with good generalization ability and stable performance. In this case, the attribute label and the semantic feature of each rough screening object in the set of candidate objects are extracted, and the two and the feature of the retrieval target are input into the ranking model. Each rough screening object is scored by the ranking model, and each rough screening object is sorted in descending order of score to obtain the similar object recommendation list. The feature of the retrieval target can be understood as a feature representation aggregated by the semantic feature and the attribute label of one or more seed objects contained therein, for example, it can be a combination of the aggregated central semantic vector, the averaged numerical label and the merged category label.

[0106] In addition, the ranking model is trained based on historical search targets of each candidate object in the candidate object library, semantic features of each candidate object, attribute labels, and historical recommendation lists corresponding to the historical search targets. The historical search target refers to a query initiated by a user in a historical operation, i.e., a historical seed object specified by the user. The historical recommendation list refers to a recommendation result displayed to the user for a certain historical search target, and an interaction behavior of the user on the result, such as clicking, viewing details, collecting, and initiating cooperation. By analyzing the positive interaction behavior of the user, different relevance levels can be marked for different objects in the recommendation list, such as marking an object that is clicked as "relevant", marking an object that is collected as "very relevant", and marking an object that is not interacted as "irrelevant".

[0107] During training, the semantic features and attribute labels of each historical search target and the candidate objects recommended at that time are used as input features of the model, and the historical recommendation list is used as a learning goal of the model. By learning a large amount of historical data, the ranking model can automatically and intelligently learn the complex and nonlinear combination relationship between the semantic features and the multi-dimensional attribute labels, and the mapping between the combination relationship and the final user preference, without manually setting complex fusion rules or weights.

[0108] As can be seen, in the condition of sufficient data, the ranking model trained based on historical user behavior data is introduced to replace the ranking method based on fixed rules, the adaptive learning and automatic optimization of the ranking strategy are realized, the real preference of the user can be more accurately fitted, and therefore the ranking quality and the individualization degree of the final recommendation list are significantly improved.

[0109] Figure 2 is a flowchart of another object retrieval method provided by the present application, as shown in Figure 2 First, before retrieval, a candidate object library is constructed. Specifically, based on Internet public data, related information of the candidate object can be collected and structured, a multi-dimensional portrait is constructed for each candidate object, the portrait is stored in Elasticsearch (ES), and one candidate object corresponds to one document storing the portrait.

[0110] Further, the image includes at least basic data, description text and attribute label. Among them, the basic data can include information such as the number of fans, the number of plays, the interaction rate and the cost per thousand exposures; the description text can be a natural language text summarizing the identity and content style of the candidate object, which is used to express its overall creative positioning, for example, the description text can be "pet care knowledge blogger, video content mainly covers daily pet care, healthy diet matching and disease prevention knowledge explanation", and the description text also stores corresponding semantic features in the ES, which can be obtained by a bag-of-words model (such as TF-IDF, LDA model) or a pre-trained language model (such as BERT model). The attribute label can include a content classification label and a word cloud label, wherein the content classification label can be understood as the theme category of the candidate object's created video, and each content classification label is accompanied by a frequency of occurrence, which reflects the proportion of this category in the created content, such as the content classification label can be {makeup: 70%, pet health: 30%}; the word cloud label can be understood as a set of high-frequency keywords in the candidate object's created content, and each word cloud label is accompanied by a frequency of occurrence, such as the word cloud label can be {dog food: 50%, pest control: 30%, vaccine: 20%}.

[0111] After the candidate object library is constructed, a user-input retrieval task can be received, the retrieval task including a retrieval target, a description text of the retrieval target and basic retrieval constraint conditions. Among them, the retrieval target can include one seed object, or multiple seed objects, or at least one positive seed object and at least one negative seed object. Among them, the retrieval constraint conditions can include the number of similar objects in the object recommendation list, the retrieval platform, the target area, the language preference, the fan quantity interval, the interaction rate range and other basic constraint conditions.

[0112] (1) Single seed object similarity retrieval, that is, the retrieval target includes one seed object, and the seed object is a positive seed object: First, it is judged whether the seed object exists in the candidate object library, if yes, the semantic features and attribute labels of the seed object are directly obtained from the candidate object library. If not, the seed object is determined to be a new seed object, at which time the semantic features are extracted from the description text of the seed object, and the attribute labels of the seed object are obtained.

[0113] Then, according to the basic retrieval constraint conditions, the first round of filtering is performed on the candidate object library to eliminate the candidate objects that do not meet the basic retrieval constraint conditions, and the preliminary screening objects are obtained. Subsequently, the semantic similarity between the semantic features of the seed object and the semantic features of the preliminary screening objects is calculated, which can be measured by cosine similarity, Euclidean distance, Manhattan distance, etc. Taking the cosine similarity as an example, the semantic similarity can be calculated based on the following formula: in, Represents the seed object semantic features Indicates preliminary screening targets semantic features Represents the seed object Compared with the initial screening targets Semantic similarity.

[0114] The preliminary screening objects are sorted from high to low according to the semantic similarity between the seed object and each preliminary screening object, and the Top N preliminary screening objects are selected as the candidate object set.

[0115] If the amount of data in the candidate object library is less than the threshold, a rule-driven approach is used to correct the ranking of the coarsely screened objects in the candidate object set. Specifically, the rule-driven approach includes: Extract the attribute tag set of the seed object and the attribute tag set of each coarsely screened object. The attribute tag set includes the content category tag set and the word cloud tag set. The content category tag set refers to the set of content category tags that appear more frequently than a first preset frequency. The word cloud tag set refers to the set of word cloud tags that appear more frequently than a second preset frequency. The first preset frequency and the second preset frequency can be the same frequency or different frequencies.

[0116] Next, the similarity between the content category tags and word cloud tags of the seed object and each coarsely screened object is calculated. The specific calculation formula is as follows: in, Represents the seed object With coarse screening object Content category tag similarity between them Represents the seed object Content category tags In the corresponding content category tag set The proportion in Represents the object being coarsely screened. Content category tags In the corresponding content category tag set The proportion of.

[0117] in, Represents the seed object With coarse screening object Word cloud tag similarity between them Represents the seed object Word cloud tags In the proportion of the corresponding word cloud label set , Indicates the word cloud label of the rough screening object In the proportion of the corresponding word cloud label set .

[0118] After that, according to the content classification label similarity and the word cloud label similarity between the seed object and each rough screening object, the semantic similarity between the seed object and each rough screening object is corrected to obtain the comprehensive similarity between each rough screening object and the seed object, wherein the correction formula is as follows: Among them, Indicates the comprehensive similarity between the seed object And the rough screening object , , is a weight parameter, Control the influence degree of content classification label correction, Control the influence degree of word cloud label correction; , is a scaling parameter, used to adjust the sensitivity of the exponential function; Indicates the reference contribution of semantic similarity, Indicates the correction contribution of content classification label, Indicates the correction contribution of word cloud label.

[0119] If the data quantity in the candidate object library is greater than or equal to the threshold value, the model-driven scheme is used to correct the ranking of each rough screening object in the candidate object set. Among them, the model-driven scheme specifically includes: inputting the features of the seed object and the features of each rough screening object into the ranking model, and ranking each rough screening object by the ranking model to obtain a similar object recommendation list. Among them, the features include semantic features, behavior features and context features, the semantic features include semantic features describing the text, content classification labels, word cloud labels, etc.; the behavior features include fan quantity difference, interaction rate similarity, content publishing frequency, growth trend, etc.; the context features include target scene (such as product category promotion, market regional preference), timeliness factor, etc. The ranking model can use traditional LTR models such as RankSVM, LambdaMART, XGBoost-Ranker, etc., or deep learning models such as DSSM, MatchPyramid, BERT semantic matching model, etc.

[0120] ​Finally, according to the comprehensive similarity of each rough screening object and the seed object, the rough screening objects are sorted to obtain a similar object recommendation list. The similar object recommendation list not only contains basic information of the similar objects (such as the number of fans, the interaction rate, the region, the language, etc.), but also is attached with a similarity score and a core matching basis (such as similar description text and a set of coinciding attribute tags) for the downstream module or the user to interpret and screen.

[0121] (2) Multiple seed object similarity search, that is, the search target includes multiple seed objects, and the seed objects are positive seed objects: After the semantic similarity of each single preliminary screening object and each seed object is calculated in the manner of (1), the maximum or average value of the semantic similarity is taken as the similarity of the corresponding preliminary screening object and the search target. For example, the search target includes seed object A, seed object B and seed object C, the semantic similarity of preliminary screening object a and seed object A is cos1, the semantic similarity of preliminary screening object a and seed object B is cos2, and the semantic similarity of preliminary screening object a and seed object C is cos3. The maximum or average value of cos1, cos2 and cos3 is taken as the semantic similarity of the preliminary screening object and the search target.

[0122] Similarly, after the content classification tag similarity of each single rough screening object and each seed object is calculated in the manner of (1), the maximum or average value of the content classification tag similarity is taken as the content classification tag similarity of the corresponding rough screening object and the search target. After the word cloud tag similarity of each single rough screening object and each seed object is calculated in the manner of (1), the maximum or average value of the word cloud tag similarity is taken as the word cloud tag similarity of the corresponding rough screening object and the search target.

[0123] Then, according to the correction formula in (1), the semantic similarity of the seed object and each rough screening object is corrected according to the content classification tag similarity and the word cloud tag similarity of the seed object and each rough screening object, to obtain the comprehensive similarity of each rough screening object and the seed object. Finally, according to the comprehensive similarity of each rough screening object and the seed object, the rough screening objects are sorted to obtain a similar object recommendation list. The similar object recommendation list not only contains basic information of the similar objects (such as the number of fans, the interaction rate, the region, the language, etc.), but also is attached with a similarity score and a core matching basis (such as similar description text and a set of coinciding attribute tags). In addition, the similar object recommendation list emphasizes the overall consistency of the similar objects and the search target or the high similarity of the similar objects to a representative seed object in the search target, and is suitable for the recommendation requirement of finding similar objects with similar overall characteristics to a certain seed group.

[0124] (3) Positive and negative seed object similarity search, that is, the search target includes multiple seed objects, and the multiple seed objects include at least one positive seed object and at least one negative seed object: After the semantic similarity between the single preliminary screening object and each seed object is calculated in the manner of (1), the maximum value or the average value of the semantic similarity is taken as the similarity between the corresponding preliminary screening object and the search target.

[0125] Similarly, after the content classification label similarity between the single rough screening object and each seed object is calculated in the manner of (1), the maximum value or the average value of the content classification label similarity is taken as the content classification label similarity between the corresponding rough screening object and the search target. After the word cloud label similarity between the single rough screening object and each seed object is calculated in the manner of (1), the maximum value or the average value of the word cloud label similarity is taken as the word cloud label similarity between the corresponding rough screening object and the search target.

[0126] Then, according to the modified formula in (1), the semantic similarity between the positive seed object and each rough screening object is modified according to the content classification label similarity and the word cloud label similarity between the positive seed object and each rough screening object, to obtain the comprehensive similarity between each rough screening object and the positive seed object. The semantic similarity between the negative seed object and each rough screening object is modified according to the content classification label similarity and the word cloud label similarity between the negative seed object and each rough screening object, to obtain the comprehensive similarity between each rough screening object and the negative seed object.

[0127] Then, the comprehensive similarity between each rough screening object and the positive seed object, and the comprehensive similarity between each rough screening object and the negative seed object are combined by weighting according to the following formula, to obtain the comprehensive similarity between each rough screening object and the search target: wherein, the comprehensive similarity between each rough screening object and the search target is denoted as S (i, j), the comprehensive similarity between each rough screening object and the positive seed object is denoted as S (i, j+1), the comprehensive similarity between each rough screening object and the negative seed object is denoted as S (i, j-1), the weight coefficient is denoted as w, which is used to control the influence balance of the positive and negative constraints.

[0128] Finally, according to the comprehensive similarity of each rough screening object and the seed object, the rough screening objects are sorted to obtain a similar object recommendation list. The similar object recommendation list not only contains the basic information (such as the number of fans, the interaction rate, the region, the language, etc.) of the similar object, the similarity score and the core matching basis (such as the similar description text, the overlapping attribute tag set), but also is attached with the comprehensive similarity with the positive seed object and the comprehensive similarity with the negative seed object, to help the user understand the reasoning logic. This mode is suitable for more complex retrieval requirements, for example, while ensuring the consistency of the style of the retrieval target group, avoiding recommending objects that are competitors or do not meet the expected characteristics.

[0129] The object retrieval device provided by the present application is described below. The object retrieval device described below can be referred to in correspondence with the object retrieval method described above.

[0130] Based on any of the above embodiments, Figure 3 is a structural schematic diagram of the object retrieval device provided by the present application, as Figure 3 shown, the device comprises: The acquisition unit 310 is configured to acquire a retrieval target, and the retrieval target comprises at least one seed object. The extraction unit 320 is configured to extract semantic features of various seed objects from description texts of the various seed objects. The recall unit 330 is configured to recall a set of candidate objects similar in semantics to the retrieval target from a candidate object library according to the semantic features of the various seed objects. The sorting unit 340 is configured to sort the recalled set of candidate objects by using the attribute tags of the various seed objects, in a case where the data amount of the candidate object library is less than a threshold value. The determination unit 350 is configured to determine a similar object recommendation list corresponding to the retrieval target from the set of candidate objects according to the sorting result.

[0131] Figure 4 is a structural schematic diagram of the electronic device provided by the present application, as Figure 4 shown, the electronic device can comprise a processor 410, a communications interface 420, a memory 430 and a communications bus 440, wherein the processor 410, the communications interface 420 and the memory 430 complete mutual communication through the communications bus 440. The processor 410 can invoke a logical instruction in the memory 430 to execute the object retrieval method.

[0132] In addition, the logic instructions in the memory 430 described above can be implemented in the form of a software function unit and sold or used as an independent product, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0133] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program is executed by a processor, so that the computer can execute the object retrieval method provided by the above-mentioned methods.

[0134] In another aspect, the present application also provides a non-transitory computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the object retrieval method provided by the above-mentioned methods.

[0135] The device embodiments described above are only schematic, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, that is, they can be located in one place, or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0136] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary universal hardware platforms, and of course, can also be realized by hardware. Based on such understanding, the above technical solutions essentially or the parts that contribute to the prior art can be embodied in the form of a software product. The computer software product can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0137] It should be pointed out finally that the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit the same; and although the present application has been described in detail with reference to the foregoing embodiments, it should be appreciated by those skilled in the art that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features thereof can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An object retrieval method characterized by comprising: The method comprises the following steps: acquiring a search target, the search target comprising at least one seed object; extracting semantic features of the seed objects from description texts of the seed objects; recalling a candidate object set similar in semantics to the search target from a candidate object library according to the semantic features of the seed objects; sorting the recalled candidate object set according to attribute labels of the seed objects, in a case where a data volume of the candidate object library is less than a threshold value; determining a similar object recommendation list corresponding to the search target from the candidate object set according to a sorting result.

2. The object retrieval method according to claim 1, characterized in that, The step of recalling a candidate object set similar in semantics to the search target from a candidate object library according to the semantic features of the seed objects comprises the following steps: determining semantic similarity of each candidate object in the candidate object library to the search target according to similarity between the semantic features of each candidate object in the candidate object library and the semantic features of the seed objects respectively; adding a candidate object with a semantic similarity greater than a threshold value to the candidate object set as a rough screening object.

3. The object retrieval method according to claim 2, characterized in that, The step of determining semantic similarity of each candidate object in the candidate object library to the search target according to similarity between the semantic features of each candidate object in the candidate object library and the semantic features of the seed objects respectively comprises the following steps: in a case where the seed objects in the search target are all positive seed objects, taking a maximum value in the similarity between the semantic features of each candidate object and the semantic features of the seed objects as the semantic similarity of the corresponding candidate object to the search target, or taking an average value of the similarity between the semantic features of each candidate object and the semantic features of the seed objects as the semantic similarity of the corresponding candidate object to the search target; in a case where the seed objects in the search target comprise positive seed objects and negative seed objects, determining the semantic similarity of each rough screening object to the search target according to a difference between the semantic similarity of each rough screening object to the positive seed objects and the semantic similarity of each rough screening object to the negative seed objects.

4. The object retrieval method according to claim 2, characterized in that, The step of sorting the recalled candidate object set according to attribute labels of the seed objects comprises the following steps: determining label similarity of each rough screening object to the search target according to matching degrees of each rough screening object in the candidate object set to the attribute labels of the seed objects respectively; fusing the label similarity and the semantic similarity of each rough screening object to obtain comprehensive similarity of each rough screening object to the search target; sorting the recalled candidate object set according to the comprehensive similarity of each rough screening object.

5. The object retrieval method according to claim 4, characterized in that, The step of determining label similarity of each rough screening object to the search target according to matching degrees of each rough screening object in the candidate object set to the attribute labels of the seed objects respectively comprises the following steps: calculating intersection quantity and union quantity between attribute label sets of each rough screening object and attribute label sets of the seed objects respectively, the attribute label set being a set of attribute labels with a frequency greater than a preset frequency among all attribute labels of the object; The label matching degree between each coarse screening object and various sub-objects is determined based on the ratio of the number of intersections to the number of unions corresponding to each coarse screening object; Based on the tag matching degree of each coarse-screened object with various sub-objects, the tag similarity between each coarse-screened object and the search target is determined.

6. The object retrieval method according to claim 5, characterized in that, The step of determining the label similarity between each coarse-screened object and the search target based on the label matching degree between each coarse-screened object and various sub-objects includes: When all seed objects in the search target are positive seed objects, the maximum value of the label matching degree between each coarse-screened object and various sub-objects is taken as the label similarity between the corresponding coarse-screened object and the search target; or, the average value of the label matching degree between each coarse-screened object and various sub-objects is taken as the label similarity between the corresponding coarse-screened object and the search target. When the seed objects in the search target include positive seed objects and negative seed objects, the label similarity between each coarse-screened object and the search target is determined based on the difference between the label similarity between each coarse-screened object and the positive seed object and the label similarity between each coarse-screened object and the negative seed object.

7. The object retrieval method according to claim 4, characterized in that, The process of fusing the label similarity and semantic similarity of each coarsely screened object to obtain the comprehensive similarity of each coarsely screened object includes: Determine the semantic weight coefficient corresponding to the semantic similarity of each coarsely screened object, and the label weight coefficient corresponding to the label similarity of each coarsely screened object; Based on the semantic weight coefficient and the label weight coefficient of each coarse screening object, the semantic similarity of each coarse screening object is corrected to obtain the corrected semantic similarity of each coarse screening object. Based on the label weight coefficient of each coarse screening object, the label similarity of each coarse screening label is corrected to obtain the corrected label similarity of each coarse screening object; By integrating the corrected semantic similarity and corrected label similarity of each coarsely screened object, a comprehensive similarity of each coarsely screened object is obtained.

8. The object retrieval method according to any one of claims 1 to 7, characterized in that, The step of recalling a set of candidate objects that are semantically similar to the search target from the candidate object library further includes: If the amount of data in the candidate object library is greater than or equal to the threshold, the retrieval target, the attribute labels and semantic features of each coarsely screened object in the candidate object set are input into the ranking model to obtain the similar object recommendation list corresponding to the retrieval target output by the ranking model. The ranking model is trained based on the historical retrieval targets of each candidate object in the candidate object library, the semantic features and attribute tags of each candidate object, and the historical recommendation list corresponding to the historical retrieval targets.

9. An object retrieval apparatus characterized by comprising: include: An acquisition unit is used to acquire a retrieval target, wherein the retrieval target includes at least one seed object; The extraction unit is used to extract semantic features of various sub-objects from the descriptive text of various sub-objects; The recall unit is used to recall a set of candidate objects that are semantically similar to the retrieval target from the candidate object library based on the semantic features of various sub-objects; The sorting unit is used to sort the recalled set of candidate objects by using the attribute tags of various sub-objects when the amount of data in the candidate object library is less than a threshold. The determining unit is used to determine a list of similar objects corresponding to the search target from the candidate object set based on the ranking result.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the object retrieval method as described in any one of claims 1 to 8.

11. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the object retrieval method as described in any one of claims 1 to 8.

Images