CN114595370A

CN114595370A - Model training and sorting method and device, electronic equipment and storage medium

Info

Publication number: CN114595370A
Application number: CN202210143365.4A
Authority: CN
Inventors: 薛涛锋; 李悦; 陶然; 郭圣昱; 张凯; 杨一帆
Original assignee: Beijing Sankuai Online Technology Co Ltd
Current assignee: Beijing Sankuai Online Technology Co Ltd
Priority date: 2022-02-16
Filing date: 2022-02-16
Publication date: 2022-06-07

Abstract

The embodiment of the disclosure provides a model training and sequencing method and device, electronic equipment and a storage medium. The model training method comprises the following steps: acquiring sample data, wherein the sample data comprises sample searching information and sample keywords corresponding to the sample object; in a preset model to be trained, carrying out feature fusion on a sample keyword and sample search information to obtain a comprehensive semantic representation vector of the sample keyword, and acquiring a sample ordering parameter of a sample object based on the comprehensive semantic representation vector; and determining a trained model based on the sample sorting parameters as a sorting model. In the embodiment of the disclosure, the keyword information of the object is merged into the ranking model, and the keywords are mined aiming at the unstructured features of the object, so that the intention of the user can be better covered and depicted, and therefore, the ranking model merged into the keyword information can capture the self semantics of the keywords and the semantic correlation between the keywords and the search information of the user, and the accuracy of the ranking model is improved.

Description

Model training and sorting method and device, electronic equipment and storage medium

Technical Field

The present disclosure relates to the field of data processing technologies, and in particular, to a model training method, a model ranking method, an apparatus, an electronic device, and a storage medium.

Background

With the rapid development of the internet technology, users need to search the required information in the information ocean, just like a great sea fishing needle, and the search engine technology just solves the problem. The search engine is a retrieval technology which retrieves specified information by using a specific strategy according to user requirements and a certain algorithm and feeds the information back to the user.

In the search process, a plurality of objects are recalled first based on search information input by a user, and then the recalled objects are sorted. In the sorting step, the recalled objects are generally analyzed by using a sorting model to obtain a sorting parameter, and the recalled objects are sorted by using the sorting parameter.

In the prior art, the ranking model is usually modeled by a specific network structure, and is mainly used for analyzing the correlation between the search information input by the user and the characteristics of the object. However, since the characteristics of the object cannot accurately reflect the important characteristics of the object, the analysis accuracy of the ranking model in this way is poor, and the model effect is poor.

Disclosure of Invention

In view of the foregoing problems, embodiments of the present disclosure provide a model training method, a model training device, a model ranking method, an electronic device, and a storage medium, so as to improve accuracy of a ranking model.

According to a first aspect of embodiments of the present disclosure, there is provided a model training method, including: acquiring sample data; the sample data comprises sample searching information and sample keywords corresponding to the sample object; performing feature fusion on the sample keywords and the sample search information in a preset model to be trained to obtain a comprehensive semantic representation vector of the sample keywords, and acquiring sample ordering parameters of the sample object based on the comprehensive semantic representation vector; and determining a trained model based on the sample sorting parameters as a sorting model.

Optionally, performing feature fusion on the sample keyword and the sample search information, including: performing feature fusion on the current sample keyword and other sample keywords aiming at each sample keyword to obtain a fusion semantic representation vector of the current sample keyword; performing feature fusion on the current sample keyword and the sample search information aiming at each sample keyword to obtain the correlation degree between the current sample keyword and the sample search information; and calculating the comprehensive semantic representation vector of the sample keywords based on the fusion semantic representation vector and the correlation of each sample keyword.

Optionally, performing feature fusion on the sample keyword and the sample search information, including: inquiring the feature vector of each sample keyword and the feature vector of the sample search information from the corresponding relation between the preset words and the feature vectors; performing feature fusion on the feature vector of the current sample keyword and the feature vectors of other sample keywords aiming at each sample keyword to obtain a fusion semantic representation vector of the current sample keyword; performing feature fusion on the fusion semantic representation vector of the current sample keyword and the feature vector of the sample search information aiming at each sample keyword to obtain the correlation degree between the current sample keyword and the sample search information; and calculating a comprehensive semantic representation vector of the sample keywords based on the fusion semantic representation vector and the correlation of each sample keyword and the feature vector of the sample search information.

Optionally, performing feature fusion on the feature vector of the current sample keyword and the feature vectors of other sample keywords, including: and performing feature fusion on the feature vector of the current sample keyword and the feature vectors of other sample keywords through an attention mechanism to obtain a fusion semantic representation vector of the current sample keyword.

Optionally, the sample data further comprises a confidence level of the sample keyword; and performing feature fusion on the feature vector of the current sample keyword and the feature vectors of other sample keywords, wherein the feature fusion comprises the following steps: performing feature fusion on the feature vector of the current sample keyword and the feature vectors of other sample keywords through a self-attention mechanism to obtain a primary fusion semantic representation vector of the current sample keyword; sorting the sample keywords in a descending order according to the confidence, and acquiring position embedded vectors of the sample keywords based on a sorting result; and adding the preliminary fusion semantic representation vector of the current sample keyword and the position embedding vector to obtain a fusion semantic representation vector of the current sample keyword.

Optionally, the sample data further comprises a confidence level of the sample keyword; performing feature fusion on the fusion semantic representation vector of the current sample keyword and the feature vector of the sample search information, wherein the feature fusion comprises the following steps: calculating the weight of the degree of confidence of the current sample keyword and the relevance of the search information; calculating intermediate parameters of the current sample keywords based on the fusion semantic representation vector of the current sample keywords, the feature vector of the sample search information and the relevancy weight of the current sample keywords; and calculating the correlation degree between the current sample keyword and the sample search information based on the intermediate parameter and a preset temperature parameter.

Optionally, calculating a comprehensive semantic representation vector of the sample keywords based on the fusion semantic representation vector and the correlation of each sample keyword and the feature vector of the sample search information, including: calculating the sum of the products of the fusion semantic representation vector and the correlation degree of each sample keyword to obtain a preliminary semantic representation vector of the sample keyword; adding the preliminary semantic representation vector and the feature vector of the sample search information to obtain a preliminary comprehensive semantic representation vector of the sample keyword; and carrying out standardization processing on the preliminary comprehensive semantic representation vector to obtain a comprehensive semantic representation vector of the sample keyword.

Optionally, after calculating a sum of products of the fused semantic representation vector and the correlation of each sample keyword to obtain a preliminary semantic representation vector of the sample keyword, the method further includes: carrying out random discarding processing on the preliminary semantic representation vector; adding the preliminary semantic representation vector to a feature vector of the sample search information, comprising: and adding the primary semantic representation vector after the random discarding treatment and the feature vector of the sample searching information.

Optionally, the sample data further includes sample description information corresponding to the sample object; obtaining a sample ordering parameter of the sample object based on the comprehensive semantic representation vector, including: and performing feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information to obtain a sample ordering parameter of the sample object.

Optionally, performing feature fusion on the integrated semantic representation vector, the sample description information, and the sample search information, including: performing deep feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information to obtain a first sample ordering parameter of the sample object; performing shallow feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information to obtain a second sample ordering parameter of the sample object; calculating a sample ordering parameter for the sample object based on the first sample ordering parameter and the first sample ordering parameter.

Optionally, performing deep feature fusion on the integrated semantic representation vector, the sample description information, and the sample search information, including: acquiring a feature vector of the sample description information and a feature vector of the sample search information; generating a spliced feature vector based on the comprehensive semantic representation vector, the feature vector of the sample description information and the feature vector of the sample search information; and performing feature fusion processing on the spliced feature vector by using a preset deep fusion network to obtain the first sample sorting parameter.

Optionally, performing shallow feature fusion on the comprehensive semantic representation vector, the sample description information, and the sample search information, including: acquiring a feature vector of the sample description information and a feature vector of the sample search information; and performing feature fusion processing on the feature vector of the sample description information, the feature vector of the sample search information and the comprehensive semantic representation vector by using a preset shallow fusion network to obtain the second sample sorting parameter.

According to a second aspect of embodiments of the present disclosure, there is provided a sorting method, including: acquiring search information and keywords corresponding to the objects to be sorted; inputting the search information and the keywords into a pre-trained sequencing model to obtain sequencing parameters of the objects to be sequenced, which are output by the sequencing model; the ranking model is obtained by the model training method as described in any one of the above; and sequencing the objects to be sequenced based on the sequencing parameters.

According to a third aspect of embodiments of the present disclosure, there is provided a model training apparatus including: the first acquisition module is used for acquiring sample data; the sample data comprises sample searching information and sample keywords corresponding to the sample object; the training module is used for carrying out feature fusion on the sample keywords and the sample search information in a preset model to be trained to obtain comprehensive semantic representation vectors of the sample keywords and acquiring sample sequencing parameters of the sample objects based on the comprehensive semantic representation vectors; and the determining module is used for determining the trained model as the sequencing model based on the sample sequencing parameters.

Optionally, the training module comprises: the first fusion unit is used for performing feature fusion on the current sample keyword and other sample keywords aiming at each sample keyword to obtain a fusion semantic representation vector of the current sample keyword; the second fusion unit is used for carrying out feature fusion on the current sample keyword and the sample search information aiming at each sample keyword to obtain the correlation degree between the current sample keyword and the sample search information; and the first calculating unit is used for calculating the comprehensive semantic representation vector of the sample keywords based on the fusion semantic representation vector and the correlation of each sample keyword.

Optionally, the training module comprises: the query unit is used for querying the feature vector of each sample keyword and the feature vector of the sample search information from the corresponding relation between preset words and the feature vectors; the third fusion unit is used for performing feature fusion on the feature vector of the current sample keyword and the feature vectors of other sample keywords aiming at each sample keyword to obtain a fusion semantic representation vector of the current sample keyword; the fourth fusion unit is used for performing feature fusion on the fusion semantic representation vector of the current sample keyword and the feature vector of the sample search information aiming at each sample keyword to obtain the correlation degree between the current sample keyword and the sample search information; and the second calculating unit is used for calculating the comprehensive semantic representation vector of the sample keywords based on the fusion semantic representation vector and the correlation of each sample keyword and the feature vector of the sample search information.

Optionally, the third fusion unit is specifically configured to perform feature fusion on the feature vector of the current sample keyword and the feature vectors of other sample keywords through an attention mechanism, so as to obtain a fusion semantic representation vector of the current sample keyword.

Optionally, the sample data further comprises a confidence level of the sample keyword; the third fusion unit is specifically configured to perform feature fusion on the feature vector of the current sample keyword and the feature vectors of other sample keywords through a self-attention mechanism to obtain a preliminary fusion semantic representation vector of the current sample keyword; sorting the sample keywords in a descending order according to the confidence, and acquiring position embedded vectors of the sample keywords based on a sorting result; and adding the preliminary fusion semantic representation vector and the position embedding vector of the current sample keyword to obtain a fusion semantic representation vector of the current sample keyword.

Optionally, the sample data further comprises a confidence level of the sample keyword; the fourth fusion unit is specifically configured to calculate a correlation weight between the confidence of the current sample keyword and the search information; calculating intermediate parameters of the current sample keywords based on the fusion semantic representation vector of the current sample keywords, the feature vector of the sample search information and the relevancy weight of the current sample keywords; and calculating the correlation degree between the current sample keyword and the sample search information based on the intermediate parameter and a preset temperature parameter.

Optionally, the second calculating unit is specifically configured to calculate a sum of products of the fused semantic representation vector and the correlation of each sample keyword, so as to obtain a preliminary semantic representation vector of the sample keyword; adding the preliminary semantic representation vector and the feature vector of the sample search information to obtain a preliminary comprehensive semantic representation vector of the sample keyword; and carrying out standardization processing on the preliminary comprehensive semantic representation vector to obtain a comprehensive semantic representation vector of the sample keyword.

Optionally, the second calculating unit is further configured to perform random discarding processing on the preliminary semantic representation vector after calculating a sum of products of the fusion semantic representation vector and the relevancy of each sample keyword to obtain the preliminary semantic representation vector of the sample keyword; the second calculating unit is specifically configured to add the preliminary semantic representation vector after the random discard processing to the feature vector of the sample search information.

Optionally, the sample data further includes sample description information corresponding to the sample object; the training module is specifically configured to perform feature fusion on the comprehensive semantic representation vector, the sample description information, and the sample search information to obtain a sample ordering parameter of the sample object.

Optionally, the training module comprises: a fifth fusion unit, configured to perform deep feature fusion on the comprehensive semantic representation vector, the sample description information, and the sample search information to obtain a first sample ordering parameter of the sample object; a sixth fusion unit, configured to perform shallow feature fusion on the comprehensive semantic representation vector, the sample description information, and the sample search information to obtain a second sample ordering parameter of the sample object; a third calculating unit, configured to calculate a sample ordering parameter of the sample object based on the first sample ordering parameter and the first sample ordering parameter.

Optionally, the fifth fusing unit is specifically configured to obtain a feature vector of the sample description information and a feature vector of the sample search information; generating a spliced feature vector based on the comprehensive semantic representation vector, the feature vector of the sample description information and the feature vector of the sample search information; and performing feature fusion processing on the spliced feature vector by using a preset deep fusion network to obtain the first sample sorting parameter.

Optionally, the sixth fusing unit is specifically configured to obtain a feature vector of the sample description information and a feature vector of the sample search information; and performing feature fusion processing on the feature vector of the sample description information, the feature vector of the sample search information and the comprehensive semantic representation vector by using a preset shallow fusion network to obtain the second sample sorting parameter.

According to a fourth aspect of embodiments of the present disclosure, there is provided a sorting apparatus including: the second acquisition module is used for acquiring search information and keywords corresponding to the objects to be sorted; the prediction module is used for inputting the search information and the keywords into a pre-trained sequencing model to obtain sequencing parameters of the objects to be sequenced, which are output by the sequencing model; the ranking model is obtained by the model training method as described in any one of the above; and the sequencing module is used for sequencing the objects to be sequenced based on the sequencing parameters.

According to a fifth aspect of embodiments of the present disclosure, there is provided an electronic apparatus including: one or more processors; and one or more computer-readable storage media having instructions stored thereon; the instructions, when executed by the one or more processors, cause the processors to perform a model training method as described in any one of the above, or alternatively, a ranking method as described in any one of the above.

According to a sixth aspect of embodiments of the present disclosure, there is provided a computer readable storage medium having stored thereon a computer program which, when executed by a processor, causes the processor to perform a model training method as defined in any one of the above, or a ranking method as defined in any one of the above.

The embodiment of the disclosure provides a model training and sequencing method and device, electronic equipment and a storage medium. In the model training process, sample data is obtained, wherein the sample data comprises sample searching information and sample keywords corresponding to a sample object; in a preset model to be trained, performing feature fusion on the sample keywords and the sample search information to obtain a comprehensive semantic representation vector of the sample keywords, and acquiring sample ordering parameters of the sample object based on the comprehensive semantic representation vector; and determining a trained model based on the sample sorting parameters as a sorting model. Therefore, in the embodiment of the disclosure, the keyword information of the object is merged into the ranking model, the keywords are mined aiming at the unstructured features of the object, compared with the unstructured features of the object, the keywords can better refine and summarize the subject of the object, and can cover the wide intention of the user, so that the ranking model merged into the keyword information can capture the semantics of the keywords and the semantic correlation between the keywords and the search information of the user, and the accuracy of the ranking model is improved.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings needed to be used in the description of the embodiments of the present disclosure will be briefly introduced below, and it is obvious that the drawings in the following description are only some drawings of the embodiments of the present disclosure, and other drawings can be obtained according to these drawings by those skilled in the art without inventive exercise.

FIG. 1 is a flow chart of steps of a model training method of an embodiment of the present disclosure.

Fig. 2 is a schematic diagram of a keyword according to an embodiment of the disclosure.

FIG. 3 is a flow chart of steps of a sorting method of an embodiment of the present disclosure.

FIG. 4 is a flow chart of steps of another method of model training in accordance with an embodiment of the present disclosure.

FIG. 5 is a flow chart of steps of another ordering method of an embodiment of the present disclosure.

FIG. 6 is a schematic diagram of an overall process of an embodiment of the disclosure.

Fig. 7 is a schematic diagram of a semantic coding network according to an embodiment of the present disclosure.

FIG. 8 is a schematic diagram of a feature interaction network of an embodiment of the present disclosure.

Fig. 9 is a block diagram of a model training apparatus according to an embodiment of the present disclosure.

Fig. 10 is a block diagram of a sorting apparatus according to an embodiment of the present disclosure.

Detailed Description

Technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present disclosure, and it is apparent that the described embodiments are only a part of the embodiments of the present disclosure, and not all the embodiments of the present disclosure. All other embodiments, which can be derived by a person skilled in the art from the embodiments disclosed herein without making any creative effort, shall fall within the protection scope of the present disclosure.

The ranking model of the embodiment of the disclosure can be applied to ranking links in a search scene, ranking links in a recommendation scene, and the like. By integrating the keyword information of the object into the ranking model, compared with the ranking model which only analyzes the unstructured features of the object, the ranking model integrated with the keyword information can capture the semantics of the keyword and the semantic correlation between the keyword and the user search information, so that the accuracy of the ranking model is improved.

Referring to fig. 1, a flow chart of steps of a model training method of an embodiment of the present disclosure is shown.

As shown in fig. 1, the model training method may include the steps of:

step 101, sample data is obtained.

In this embodiment, the sample data may include sample search information and sample keywords corresponding to the sample object, and the sample data may further include actual ranking parameters of the sample object.

Sample objects may include, but are not limited to, content objects and the like. Content objects may include, but are not limited to: documents, web pages, pictures, video, audio, and the like.

Sample search information may include, but is not limited to, user-entered search information, and the like. The form of the search information may include, but is not limited to: text form, speech form, picture form, etc.

Sample keywords refer to words or phrases mined from information related to sample objects that can better refine and summarize the subject matter of the sample objects, each sample object may have at least one keyword. Referring to fig. 2, a schematic diagram of a keyword according to an embodiment of the present disclosure is shown. In fig. 2, 4 objects are shown, wherein the keywords of the first object include "old Shanghai stormy street" and "shopping good place", the keywords of the second object include "Shanghai bookshop" and "Ownship assault", the keywords of the third object include "Shanghai celebrity street" and "old foreign house", and the keywords of the fourth object include "Shanghai assault street" and "must-hit scenic spot".

The actual ordering parameter refers to an actual parameter relating to the ordering of the sample objects. Actual parameters may include, but are not limited to: click rate, conversion, exposure, etc.

And 102, performing feature fusion on the sample keywords and the sample search information in a preset model to be trained to obtain a comprehensive semantic representation vector of the sample keywords, and acquiring sample ordering parameters of the sample object based on the comprehensive semantic representation vector.

And inputting the sample search information and the sample keywords corresponding to the sample object into a preset model to be trained. In the model to be trained, the comprehensive semantic representation vector of the sample keywords is obtained by carrying out feature fusion on the sample keywords and the sample search information. The comprehensive semantic representation vector of the sample keywords can capture the search intention of a user and improve the search relevance, so that the sample sorting parameters obtained based on the comprehensive semantic representation vector of the sample keywords can more accurately represent the sorting characteristics of the sample objects.

And 103, determining a trained model based on the sample ranking parameters as a ranking model.

And the model to be trained acquires the sample ordering parameters of the sample objects based on the comprehensive semantic representation vector and outputs the sample ordering parameters of the sample objects. Whether training is complete may be determined based on the sample ordering parameters and the actual ordering parameters of the sample objects.

Alternatively, the loss function may be calculated based on the sample ordering parameters and the actual ordering parameters of the sample objects. The loss function is used to measure the degree of disagreement between the predicted value and the true value of the model. If the loss function is very small, the model is very close to the real distribution of the data, and the performance of the model is good; if the loss function is very large, which indicates that the difference between the real distribution of the model and the real distribution of the data is large, the performance of the model is poor. The task of training the model is to use an optimization method to find the model parameters for which the loss function minimizes. Thus, in case the loss function satisfies a preset condition (e.g. the loss function is smaller than a certain threshold), it may be determined that the training is complete. Alternatively, the loss function may include, but is not limited to: cross entropy loss functions, exponential loss functions, Dice loss functions, cross-over-loss functions, and the like.

In response to determining that the untraining is complete, the parameters of the model to be trained may be updated, and the training continues until the training is complete. In response to determining that the training is complete, the trained model is treated as a ranking model.

The objects to be ranked can be ranked based on the ranking model trained by the model training method shown in fig. 1.

Referring to fig. 3, a flow chart of steps of a sorting method of an embodiment of the present disclosure is shown.

As shown in fig. 3, the sorting method may include the steps of:

step 301, obtaining search information and keywords corresponding to the objects to be sorted.

The object to be sorted refers to an object to be sorted, for example, a plurality of objects recalled in the recall link may be used as the object to be sorted.

The search information corresponding to the object to be ranked may include, but is not limited to, search information input by a user, and the like. The form of the search information may include, but is not limited to: text form, speech form, picture form, etc. The search information corresponding to the plurality of objects to be ranked may be the same, for example, the search information corresponding to the plurality of objects to be ranked recalled based on the same search information is the same.

The keywords corresponding to the objects to be sorted refer to words or phrases which are mined from relevant information of the objects to be sorted and can better refine and summarize the subjects of the objects to be sorted, and each object to be sorted can have at least one keyword. For the specific process of mining the keywords, the relevant processing can be performed according to actual experience, and the embodiment is not discussed in detail here.

Step 302, inputting the search information and the keywords into a pre-trained ranking model to obtain ranking parameters of the objects to be ranked output by the ranking model.

Inputting search information and keywords corresponding to the objects to be sorted into a pre-trained sorting model, performing feature fusion on the search information and the keywords corresponding to the objects to be sorted in the sorting model to obtain a comprehensive semantic representation vector of the keywords, acquiring sorting parameters of the objects to be sorted based on the comprehensive semantic representation vector of the keywords, and outputting the sorting parameters of the objects to be sorted by the sorting model. The sorting parameters of the objects to be sorted may include, but are not limited to: click rate, conversion, exposure, etc.

Step 303, sorting the objects to be sorted based on the sorting parameters.

The objects to be sorted can be sorted according to a preset rule based on the sorting parameters of the objects to be sorted. For example, mathematical operations (such as weighting calculation, average calculation, etc.) are performed based on the sorting parameters of the objects to be sorted to obtain comprehensive sorting parameters of the objects to be sorted, and the objects to be sorted are sorted based on the comprehensive sorting parameters. For example, the larger the composite ranking parameter, the earlier the ranking order, and so on.

In the embodiment of the disclosure, the keyword information of the sample object is merged into the ranking model, the keywords are mined aiming at the unstructured features of the object, compared with the unstructured features of the object, the keywords can better refine and summarize the subject of the object, and can cover and depict the wide intention of the user, so that the ranking model merged into the keyword information can capture the semantics of the keywords and the semantic correlation between the keywords and the search information of the user, and the accuracy of the ranking model is improved.

Referring to FIG. 4, a flow chart of steps of another model training method of an embodiment of the present disclosure is shown.

As shown in fig. 4, the model training method may include the steps of:

step 401, sample data is obtained.

In this embodiment, the sample data may include sample search information, sample keywords, sample description information, and the like corresponding to the sample object, and the sample data may further include an actual ranking parameter of the sample object.

For the sample object, the sample search information, the sample keyword and the actual ranking parameter, reference may be made to the above-mentioned related description of step 101.

Sample description information may include, but is not limited to: the sample object's own features (such as document features of a document, picture features of a picture, etc.), context features, user features, and so forth. The self characteristics may include, but are not limited to: title, logo, author, etc. Contextual features may include, but are not limited to: time, place, etc. User characteristics may include, but are not limited to: age, gender, preferences, etc. of the user.

Step 402, in a preset model to be trained, performing feature fusion on the sample keywords and the sample search information to obtain a comprehensive semantic representation vector of the sample keywords, and performing feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information to obtain a sample ordering parameter of the sample object.

And inputting the sample searching information, the sample keywords and the sample description information corresponding to the sample object into a preset model to be trained. In the model to be trained, the comprehensive semantic representation vector of the sample keywords is obtained by carrying out feature fusion on the sample keywords and the sample search information. The comprehensive semantic representation vector of the sample keywords can capture the search intention of the user, and the search relevance is improved. The comprehensive semantic representation vector of the sample keywords, the sample description information and the sample search information are subjected to feature fusion to obtain the sample sorting parameters of the sample objects, so that sufficient feature interaction can be performed, and the search correlation and the result quality are improved.

And step 403, determining a trained model based on the sample sorting parameters as a sorting model.

For the specific process of step 403, reference may be made to the related description of step 103 above.

The objects to be ranked can be ranked based on the ranking model trained by the model training method shown in fig. 4.

Referring to FIG. 5, a flow chart of steps of another ordering method of an embodiment of the present disclosure is shown.

As shown in fig. 5, the sorting method may include the steps of:

step 501, obtaining search information, keywords and description information corresponding to the objects to be sorted.

For the objects to be sorted, the search information, the keywords, refer to the related description of step 301 above.

The description information of the object to be sorted may include, but is not limited to: the object to be ranked may be a document, a picture, a context, a user, or the like.

Step 502, inputting the search information, the keyword and the description information into a pre-trained ranking model to obtain ranking parameters of the object to be ranked output by the ranking model.

And inputting the search information, the keywords and the description information corresponding to the objects to be sorted into a pre-trained sorting model. In the sequencing model, the key words corresponding to the objects to be sequenced and the sample search information are subjected to feature fusion to obtain comprehensive semantic representation vectors of the key words, the comprehensive semantic representation vectors, the description information and the search information of the key words are subjected to feature fusion to obtain sequencing parameters of the objects to be sequenced, and the sequencing model outputs the sequencing parameters of the objects to be sequenced.

Step 503, sorting the objects to be sorted based on the sorting parameters.

For the specific process of step 503, reference may be made to the description related to step 303 above.

In the embodiment of the disclosure, keywords of the object are introduced, and a set of universal sequencing framework which integrates interactive relations among keyword semantic information, search information and object description information is provided. The framework can further perform sufficient feature interaction with important features in a scene by improving the expression capability of the keyword semantic vector, and improves the search relevance and the result quality.

In the following, a content search (specifically, a document search) in a search scenario will be described as an example. However, in practical applications, the method is not limited to this, and the sorting process for any object in any scene may be processed.

In a search scene, the keyword fused ordering method is to fuse keyword information of an object into an ordering model, so that the ordering model can capture the semantics of the keywords and the semantic relevance between the keywords and the search information of a user, and the search relevance and the result quality are improved.

The content search (which can be specifically a document search) is an important module for commenting search and bears an important mission for commenting on the ecological construction of the content. At present, the content understanding link produces high-coverage keywords (words or phrases), which are mined from text related to documents, and compared with unstructured text information of the documents, the keywords can better refine and summarize the subjects of the documents, and each document is marked with a plurality of keyword tags. How to fall down the keywords of the documents into the sequencing scene of content search, mine the correlation between the keywords and the user search information (also called as search words), capture the search intention of the user and influence the distribution of search data is a problem to be solved by the embodiment of the disclosure.

In the following, the description is made according to the usage of the keywords in the ranking and the relevance modeling method in the ranking model.

Under a search scene, the usage modes of the keyword labels in the sorting can be divided into the following two types:

explicit feature modeling

a. Single-dimensional characteristics of keywords: namely: the features of the keywords themselves are used as additional input features of the model, such as TF-IDF (Term Frequency-Inverse text Frequency) statistical features or pre-training vectors, and the like.

② model by implicit semantics

a. And (3) carrying out single-dimensional implicit semantic representation on the keywords: taking the keywords as a whole, mapping the keywords to a unique identifier, and randomly initializing word vectors; or the key words are regarded as the sequence of the words, word vectors are initialized randomly, the semantic representation vectors of the key words are obtained through aggregation by using a bag-of-words model or a sequence model, and finally the semantic representation vectors of the key words are learned end to end through the model.

b. Based on the pre-training model: the pre-training model is used as a feature extractor of the keyword label and then combined with the sequencing model to be finely adjusted in the sequencing task.

Under a search scenario, methods for modeling search relevance in a ranking stage can be divided into the following two categories:

digging correlation characteristics

Explicit cross-features between text of mined documents and user search information, such as: the editing distance between the title or body of the document and the search information, etc.

Model structural modeling correlation

In a search scenario, a relevance ranking method may model the relevance between document text and user search information through a particular network structure. For example:

a. introducing a double tower structure into the sequencing model, namely: and finally, the representation or dot product result of the text and the search information can be used as the input of the ranking model together with the representation vectors of other ranked features for deep fusion.

b. The relevance between text and search information is captured by stitching the text and search information of a document using a sequence characterization model, such as LSTM (Long Short-Term Memory) or Transformers.

c. Introducing a pre-training model, introducing an Encoder of BERT (Bidirectional Encoder representation based on a converter) into a sequencing model, encoding text or search information of a document, and finely adjusting in a downstream task of sequencing.

However, if the above-described method is simply adopted, there will be the following problems:

1. the above sorting method of fused keywords cannot effectively capture the deep semantic information of the keywords. For example: when the keyword itself is used as the input feature, if the method is used to mine the literal explicit feature such as the edit distance between the keyword itself and the search information, when the keyword and the search information are not in agreement literally, but there are close meaning, similarity, upper and lower relation semantically, and the like, such explicit relation feature cannot capture the semantic correlation between the keyword and the search information. When latent semantic modeling is performed on a keyword, if a random initialization and end-to-end learning mode is adopted, the sequencing model may be difficult to learn the latent semantic representation of the keyword, because the user behavior in a training sample is very sparse, that is: the learned token vectors have poor expressive power.

2. The keywords are noisy, and if the keywords are fused separately, the noise of the keywords can be ignored, so that the robustness of the model is poor. Typically, for a document, there is a confidence score for each keyword to measure how confident it represents the document. Some keywords have low confidence, and direct introduction brings noise to the model. Even if the confidence level of some keywords is high, under certain search information, the keywords and the search information may be irrelevant, the keywords are directly introduced, and the semantic relevance of the search information and the document captured by the model is also noise.

3. The method for modeling the ranking relevance is mainly based on the relevance between the unstructured text of the document and the search information, and lacks a search ranking framework which can model the structured keywords of the document and discover the relationship between the search intention of the user and the document theme. The keywords of the document are the extraction and summarization of the document theme, and can cover and depict the wide search intention of the user in the search scene. If the method can be fully utilized in the ranking model, the method is important for improving the relevance of the search results.

In order to better fuse the keywords into the ranking model and thus improve the relevance and quality of the search, the present embodiment performs the following two optimizations for the above manner:

1. the embodiment provides a deep semantic coding network fusing keywords and user search information, which can effectively mine the semantic information of the keywords, automatically filter noise information and improve the accuracy and robustness of semantic representation.

2. The method introduces keywords, and provides a set of general sequencing flow and framework based on pre-training semantic vectors, fusing keyword semantic information, modeling labels and interactive relations among important features of search scenes. The frame improves the expression capability of the keyword semantic vector through the pre-training vector, further performs sufficient feature interaction with important features of context, documents, users and the like in a search scene, and improves the search relevance and the result quality.

Referring to fig. 6, a schematic diagram of an overall process of an embodiment of the present disclosure is shown. FIG. 6 provides a general ranking framework based on keywords and pre-training vectors, which can effectively integrate keywords into a ranking model to extract the interaction between semantic vectors and modeling features. As shown in FIG. 6, the overall process may include two parts of the pre-processing flow and the ranking model of the keywords and the pre-training vectors. Which are described separately below.

Firstly, preprocessing process of keywords and pre-training vectors.

Firstly, mining key words and training a pre-training model from massive corpora; then, constructing a keyword dictionary based on the online exposure log data and the mined keywords; next, a pre-training model is used to infer and extract pre-training vectors for the keywords. The specific process is as follows:

(1) and training a pre-training model. Training is carried out based on the mass corpora to obtain a pre-training model. The pre-training model can comprise a BERT model and the like, wherein the BERT model is a vertical domain pre-training model based on a Transformers structure and is excellent in tasks such as single sentence distribution, sequence annotation, inter-sentence relation and the like.

(2) And (5) mining the keywords. And (3) mining key words of the massive corpora by using unsupervised methods such as dependency syntax analysis, TF-IDF, text clustering and the like to obtain a large number of key words, wherein the mined key words can well measure the theme of the content. After the keywords are obtained by mining, the confidence of the keywords can also be obtained.

(3) And constructing a keyword dictionary. And searching exposure log data based on the online content for statistics, low-frequency filtering and the like, filtering low-frequency words from keywords related to the exposed content and user search information, and counting and constructing to obtain a keyword dictionary (comprising the keywords and the search information).

(4) And generating a keyword pre-training vector. By using the pre-training model, a pre-training vector (i.e., a feature vector) of each word in the keyword dictionary is extracted to obtain a keyword pre-training vector table (i.e., a corresponding relationship between the word and the feature vector). The pre-training vector contains rich semantic information, and can well depict semantic relevance among different keywords or search information. The vector is the average pooling result of token sequence characterization vectors in the word. Compared with [ CLS ] token vectors, the vector obtained by average pooling has more complete semantic features.

And II, sequencing the models.

The ranking model may include a semantic coding network and a feature interaction network. On one hand, extracting comprehensive semantic representation vectors of the keywords through a semantic coding network; and on the other hand, feature fusion is carried out through a feature interaction network, and finally a predicted value of the model is output for offline training and online reasoning. The specific flow details are as follows:

(1) and carrying out keyword semantic coding on the basis of the keyword pre-training vector table to obtain a comprehensive semantic representation vector of the keyword.

And performing feature fusion on the keywords and the search information through a semantic coding network to obtain a comprehensive semantic representation vector of the keywords.

Referring to fig. 7, a schematic diagram of a semantic coding network of an embodiment of the present disclosure is shown. As shown in fig. 7, the semantic coding network may include an Input Layer (Input Layer), an Embedding Layer (Embedding Layer), a Self-attention Layer (Self-attention Layer), a Confidence-Aware aggregation Layer (Confidence-Aware aggregation), and an Output Layer (Output Layer).

An input layer: the input of the input layer is keywords and search information. As in FIG. 7, the keywords include keywords₁，keyword₂，…，keyword_nAnd the search information is query.

Embedding layer: the embedding layer is used for inquiring the feature vector of each keyword and the feature vector of the search information based on the corresponding relation between the words and the feature vectors. As in FIG. 7, keyword keywords₁The feature vector of

Keyword word₂The feature vector of

… keyword_nIs a feature vector of

The feature vector of the search information query is e_q。

Based on the keyword pre-training vector table (i.e. the corresponding relation between the words and the feature vectors), the parameters of the embedding layer of the semantic coding network are initialized, and all the related features share the embedding layer. Namely: the parameter space is shared by the keyword features and the user search information features, so that the complexity can be obviously reduced, and overfitting can be prevented. Compared with random initialization, initialization by using the pre-trained feature vectors can ensure that the model can effectively capture and focus on the relevant features of the text level from the beginning, thereby influencing the sequencing learning process. In particular, using pre-trained feature vectors

To initialize a characterization vector parameter matrix E, which is learnable and is fine-tuned during training of the ranking model.

Self-attention layer: for discovering relationships between keywords.

There may be multiple keywords in a document, and there is a collaborative relationship between different keywords, for example: the method comprises the steps of "basketball", "training", and seeing any keyword alone, the topic of the document is difficult to completely depict, the two keywords are combined to be capable of representing the topic of the document, and the fact that under what search information, the document result is most suitable for being displayed to a user can be deduced. Clearly, when a user searches for "basketball training," this document is highly relevant and meets the user's needs. Thus, relationships between keywords may be mined to help the model better capture the relationship of each keyword to the user's search information.

Taking offline training as an example, the self-attention layer performs feature fusion on the current sample keyword and other sample keywords aiming at each sample keyword to obtain a fusion semantic representation vector of the current sample keyword.

Optionally, the self-attention layer performs feature fusion on the feature vector of the current sample keyword and the feature vectors of other sample keywords through a self-attention mechanism for each sample keyword to obtain a fusion semantic representation vector of the current sample keyword.

And mining the relation between the keywords through an attention mechanism to obtain a fused keyword semantic representation vector, so that each keyword can be fused into semantic information of other keywords.

The formula is expressed as formula one:

in formula one, Self-Attention mechanism calculation process is represented by Self-Attention-mechanism () of Self-Attention-mechanism calculation process,

a feature vector representing a keyword of the sample,

a fused semantic representation vector representing a sample keyword. This manner corresponds to the case shown in fig. 7.

Optionally, the self-attention layer performs feature fusion on the feature vector of the current sample keyword and feature vectors of other sample keywords through a self-attention mechanism to obtain a preliminary fusion semantic representation vector of the current sample keyword; sorting the sample keywords in a descending order according to the confidence degrees of the sample keywords, and acquiring position embedded vectors of the sample keywords based on a sorting result; and adding the preliminary fusion semantic representation vector of the current sample keyword and the position embedding vector to obtain a fusion semantic representation vector of the current sample keyword.

For a certain document i, the confidence of each keyword k is different. The keywords are firstly arranged in a descending order according to the confidence degree to form a keyword sequence { k₁,k₂,...,k_nAnd then forming a keyword initial characterization sequence through table lookup, namely:

and then, mining the relation between the keywords through a self-attention mechanism to obtain a fused keyword semantic representation sequence, so that each keyword can be fused into semantic information of other keywords.

The formula is expressed as formula two and formula three:

in formula two, Self-Attention mechanism calculation process is represented by Self-Attention-mechanism () of Self-Attention-mechanism calculation process,

a feature vector representing a keyword of the sample,

a preliminary fused semantic representation vector representing sample keywords. In the third formula, p_iSample keyword k representing confidence rank ith_iPosition embedding vector of (1), p_iThe confidence bias for differently ranked sample keywords can be captured from the global level,

sample keyword k representing confidence rank ith_iThe preliminary fused semantic representation vector of (a) is,

sample keyword k representing confidence rank ith_iThe fused semantic representation vector of (2).

Confidence-aware convergence layer: the method is used for automatically selecting effective keywords related to user search information and automatically filtering noise information.

The user's search information may be related to a part of the keywords, and not related to a part of the keywords, so that it is necessary to select the keywords most related to the search information.

Taking offline training as an example, the convergence layer for confidence perception performs feature fusion on current sample keywords and sample search information for each sample keyword to obtain the correlation degree between the current sample keywords and the sample search information, and calculates a preliminary comprehensive semantic representation vector of the sample keywords based on the fusion semantic representation vector and the correlation degree of each sample keyword.

Optionally, the confidence perception convergence layer performs feature fusion on the fusion semantic representation vector of the current sample keyword and the feature vector of the sample search information for each sample keyword to obtain the correlation between the current sample keyword and the sample search information; and calculating a preliminary comprehensive semantic representation vector of the sample keywords based on the fusion semantic representation vector and the correlation of each sample keyword and the characteristic vector of the sample search information.

Optionally, the process of performing feature fusion on the fusion semantic representation vector of the current sample keyword and the feature vector of the sample search information to obtain the correlation between the current sample keyword and the sample search information may include: calculating the weight of the degree of confidence of the current sample keyword and the relevance of the search information; calculating intermediate parameters of the current sample keywords based on the fusion semantic representation vector of the current sample keywords, the feature vector of the sample search information and the relevancy weight of the current sample keywords; and calculating the correlation degree between the current sample key word and the sample searching information based on the intermediate parameter and a preset temperature parameter.

As shown in fig. 7, in the confidence-aware convergence layer, through an Attention Network (Attention Network), a relevance weight between confidence-aware, user search information and a fused semantic representation vector of a keyword is calculated by using an Attention mechanism, and useful information is adaptively selected from the fused semantic representation vector of the keyword.

Ith sample keyword k_iThe formula of the confidence of (2) and the relevance weight of the search information is expressed as formula four:

in the fourth formula, softmax () represents the calculation process of the softmax function,

represents the ith sample keyword k_iThe degree of confidence of (a) is,

represents the ith sample keyword k_iThe correlation weight of (2).

If it corresponds to formula one, the ith sample keyword k_iThe fused semantic representation vector of

The ith sample keyword k_iIs expressed as formula five:

in the formula five, MLP represents a multi-layer feedforward neural network, | | | | represents a splicing operation,

represents the ith sample keyword k_iFused semantic representation vector of e_qFeature vector, s (q, k), representing sample search information_i) Represents the ith sample keyword k_iThe intermediate parameter(s).

If it corresponds to the above formula two and formula three, the ith sample keyword k_iThe fused semantic representation vector of

The ith sample keyword k_iIs expressed as formula six:

in the formula six, the first step is carried out,

represents the ith sample keyword k_iThe other parameters refer to the related description of the formula five.

Ith sample keyword k_iThe formula of the correlation with the sample search information is represented as formula seven:

a_A(q，k_i)＝softmax(s(q，k_i) /τ) formula seven

In formula seven, softmax () represents the calculation process of the softmax function, a_A(q，k_i) Represents the ith sample keyword k_iAnd a correlation degree with the sample search information, and τ represents a temperature parameter. The temperature parameter can enlarge the difference degree of importance of different keywords, namely: the more the content subject can be reflected, the more the keywords related to the user search information are likely to obtain larger weight, so that the contribution degree to the comprehensive semantic representation vector is larger, the noise keywords unrelated to the search information have small contribution degree, and the interference of the noise keywords can be obviously reduced. In a word, the importance of keywords with different confidence levels can be effectively distinguished by combining the attention mechanism of temperature parameters and confidence perception, and the robustness of the model is improved.

Optionally, the process of calculating a preliminary comprehensive semantic representation vector of the sample keywords based on the fusion semantic representation vector and the relevancy of each sample keyword and the feature vector of the sample search information may include: calculating the sum of the products of the fusion semantic representation vector and the correlation degree of each sample keyword to obtain a preliminary semantic representation vector of the sample keyword; and adding the preliminary semantic representation vector and the feature vector of the sample search information to obtain a preliminary comprehensive semantic representation vector of the sample keyword.

The formula of the preliminary semantic representation vector of the sample keyword is represented as formula eight:

The formula of the preliminary semantic representation vector of the sample keyword is expressed as formula nine:

formula eight and formula nine, h_KThe preliminary semantic representation vector, sigma, represents the sum of the sample keywords.

The formula of the preliminary comprehensive semantic representation vector of the sample keywords is expressed as formula ten:

in the formula ten, e_qA feature vector representing the sample search information,

a preliminary synthetic semantic representation vector representing sample keywords.

Optionally, since there may be some inaccuracy in the keyword, forcing the keyword into the ranking model may bring noise to the model, and in order to ensure robustness, part of the neurons in the preliminary semantic representation vector may be randomly discarded through a Dropout mechanism.

Therefore, after calculating the sum of the products of the fused semantic representation vector and the correlation of each sample keyword to obtain the preliminary semantic representation vector of the sample keyword, the method further comprises: and carrying out random discarding treatment on the preliminary semantic representation vector.

The formula of the primary semantic representation vector after the random discarding process is expressed as formula eleven:

h′_K＝dropout(h_K) Formula eleven

In formula eleven, dropout () represents the process of the discard mechanism, h'_KRepresents the preliminary semantic representation vector after the random discard process, h_KA preliminary semantic representation vector representing sample keywords.

Correspondingly, adding the primary semantic representation vector after random discarding processing and the feature vector of the sample search information to obtain a primary comprehensive semantic representation vector of the sample keyword.

The formula of the preliminary comprehensive semantic representation vector of the sample keywords is expressed as the formula twelve:

in the formula twelve, e_qA feature vector representing the sample search information,

An output layer: and the comprehensive semantic representation vector is used for outputting the comprehensive semantic representation vector of the key word based on the preliminary comprehensive semantic representation vector of the key word.

Taking offline training as an example, the preliminary comprehensive semantic representation vector of the sample keyword is subjected to standardization processing on an output layer to obtain the comprehensive semantic representation vector of the sample keyword.

The formula of the comprehensive semantic representation vector of the sample keyword is expressed as formula thirteen:

in equation thirteen, Layer norm () represents the Normalization process (Layer Normalization),

a synthetic semantic representation vector representing the sample keywords.

(2) And carrying out deep layer feature fusion and shallow layer feature fusion on the comprehensive semantic representation vector based on the keywords to obtain a predicted value.

And inputting the comprehensive semantic representation vector of the keyword into a feature interaction network for feature fusion, and acquiring the sequencing parameters of the objects based on the comprehensive semantic representation vector of the keyword through the feature interaction network. Specifically, feature fusion is performed on the comprehensive semantic representation vector, the description information and the search information of the keyword to obtain the sequencing parameters of the object.

Referring to FIG. 8, a schematic diagram of a feature interaction network of an embodiment of the present disclosure is shown. Fig. 8 illustrates a Multi-target model, which mainly includes a deep-layer fusion Network (MMOE (Multi-gate Mixture-of-Experts) Network) and a shallow-layer fusion Network (FM (Factor Machine) Network). The underlying input features include object description information, search information (query), and a comprehensive semantic representation vector of the Keyword obtained through the aforementioned semantic coding network (Keyword Encoder). As shown in fig. 8, the description information of the object may include, but is not limited to, a document Feature (Item Feature), a Context Feature (Context Feature), and the like.

Taking offline training as an example, optionally, the process of performing feature fusion on the comprehensive semantic representation vector, the sample description information, and the sample search information may include: performing deep feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information to obtain a first sample ordering parameter of the sample object; performing shallow feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information to obtain a second sample ordering parameter of the sample object; calculating a sample ordering parameter for the sample object based on the first sample ordering parameter and the first sample ordering parameter.

Optionally, performing deep feature fusion on the integrated semantic representation vector, the sample description information, and the sample search information, including: acquiring a feature vector of the sample description information and a feature vector of the sample search information; generating a spliced feature vector based on the comprehensive semantic representation vector, the feature vector of the sample description information and the feature vector of the sample search information; and performing feature fusion processing on the spliced feature vector by using a preset deep fusion network to obtain a first sample sorting parameter.

As shown in fig. 8, a feature vector is extracted by means of an embedded matrix or sequence modeling for each feature of the sample description information, and a feature vector of the sample search information is obtained. And (5) splicing the feature vectors of the same type (Concat), as shown in fig. 8, splicing the feature vectors of the document features, and splicing the feature vectors of the context features and the feature vectors of the sample search information. And then splicing the spliced vector and the comprehensive semantic representation vector of the sample keyword to generate a spliced feature vector. The spliced feature vectors are preliminarily fused through a Fusion Network and then serve as the input of a deep Fusion Network (MMoE Network), the deep Fusion Network extracts high-order hidden vectors, semantic information and other information contained in keywords can be fully fused, and a predicted value specific to a task, namely a first sample sorting parameter, is output.

As shown in fig. 8, a feature vector is extracted by means of embedded matrix or sequence modeling for each feature of the sample description information, and a feature vector of the sample search information is obtained. And the characteristic vector of the sample description information, the characteristic vector of the sample search information and the comprehensive semantic representation vector of the sample keywords are used as the input of a shallow fusion Network (FM Network), and the shallow fusion Network performs sufficient feature interaction on the comprehensive semantic representation vector of the keywords and the characteristic vectors of some memorability sparse features and outputs a predicted value shared by tasks, namely a second sample sequencing parameter.

Optionally, the first sample sorting parameter and the first sample sorting parameter are added correspondingly to obtain a sample sorting parameter of each task corresponding to the sample object. Sample ordering parameters may include, but are not limited to, click-through rate, conversion rate, exposure, user interaction behavior parameters, and the like.

In the embodiment of the disclosure, a keyword semantic coding network with confidence perception is provided, which can effectively extract semantic information in keywords, automatically filter noise information and improve robustness of semantic representation; meanwhile, a framework for fusing the comprehensive semantic representation vectors of the keywords into the ranking model is provided, so that sufficient feature interaction can be performed with the help of the expression capability of the pre-training vectors and the keywords and the features such as important context, documents, users and the like in a search scene, and the relevance and the quality of the search results are improved.

By introducing the keyword information into the sequencing model, initializing parameters based on the pre-training vector, and then finely adjusting in the sequencing model, the sequencing model effectively captures semantic relevance information, and the sequencing relevance is improved. Through an attention mechanism based on confidence perception, the importance of different keywords can be well distinguished, noise is reduced, and the robustness of semantic representation can be further ensured through temperature parameters and a Dropout mechanism.

The keywords are very important structural information of the content community and the platform, can be used for depicting the subject of the content and the explicit interest preference of the user, and are widely applied to content search and recommendation scenes. Wherein, the search scene can be used for capturing the relevance of the search words and the keywords of the user; the recommendation scene can be used for constructing a keyword sequence of the content clicked by the user, describing long and short term interests of the user and constructing the user portrait label, and then carrying out generalization and accurate capture on the user interests based on the user portrait label and the keyword of the content, so that the recommendation performance is improved. The method can be adopted in a content platform.

Referring to fig. 9, a block diagram of a model training apparatus according to an embodiment of the present disclosure is shown.

As shown in fig. 9, the model training apparatus may include the following modules:

a first obtaining module 901, configured to obtain sample data; the sample data comprises sample searching information and sample keywords corresponding to the sample object;

a training module 902, configured to perform feature fusion on the sample keywords and the sample search information in a preset model to be trained, to obtain a comprehensive semantic representation vector of the sample keywords, and obtain sample ordering parameters of the sample object based on the comprehensive semantic representation vector;

a determining module 903, configured to determine a trained model based on the sample ordering parameter as an ordering model.

Optionally, the training module 902 comprises: the first fusion unit is used for performing feature fusion on the current sample keyword and other sample keywords aiming at each sample keyword to obtain a fusion semantic representation vector of the current sample keyword; the second fusion unit is used for carrying out feature fusion on the current sample keyword and the sample search information aiming at each sample keyword to obtain the correlation degree between the current sample keyword and the sample search information; and the first calculating unit is used for calculating the comprehensive semantic representation vector of the sample keywords based on the fusion semantic representation vector and the correlation of each sample keyword.

Optionally, the training module 902 comprises: the query unit is used for querying the feature vector of each sample keyword and the feature vector of the sample search information from the corresponding relation between preset words and the feature vectors; the third fusion unit is used for performing feature fusion on the feature vector of the current sample keyword and the feature vectors of other sample keywords aiming at each sample keyword to obtain a fusion semantic representation vector of the current sample keyword; the fourth fusion unit is used for performing feature fusion on the fusion semantic representation vector of the current sample keyword and the feature vector of the sample search information aiming at each sample keyword to obtain the correlation degree between the current sample keyword and the sample search information; and the second calculating unit is used for calculating the comprehensive semantic representation vector of the sample keywords based on the fusion semantic representation vector and the correlation of each sample keyword and the feature vector of the sample search information.

Optionally, the third fusing unit is specifically configured to perform feature fusion on the feature vector of the current sample keyword and feature vectors of other sample keywords through a self-attention mechanism, so as to obtain a fused semantic feature vector of the current sample keyword.

Optionally, the sample data further comprises a confidence level of the sample keyword; the third fusion unit is specifically configured to perform feature fusion on the feature vector of the current sample keyword and the feature vectors of other sample keywords through a self-attention mechanism to obtain a preliminary fusion semantic representation vector of the current sample keyword; sorting the sample keywords in a descending order according to the confidence, and acquiring position embedded vectors of the sample keywords based on a sorting result; and adding the preliminary fusion semantic representation vector of the current sample keyword and the position embedding vector to obtain a fusion semantic representation vector of the current sample keyword.

Optionally, the sample data further includes sample description information corresponding to the sample object; the training module 902 is specifically configured to perform feature fusion on the comprehensive semantic representation vector, the sample description information, and the sample search information to obtain a sample ordering parameter of the sample object.

Optionally, the training module 902 comprises: a fifth fusion unit, configured to perform deep feature fusion on the comprehensive semantic representation vector, the sample description information, and the sample search information to obtain a first sample ordering parameter of the sample object; a sixth fusion unit, configured to perform shallow feature fusion on the comprehensive semantic representation vector, the sample description information, and the sample search information to obtain a second sample ordering parameter of the sample object; a third calculating unit, configured to calculate a sample ordering parameter of the sample object based on the first sample ordering parameter and the first sample ordering parameter.

Optionally, the sixth fusing unit is specifically configured to obtain a feature vector of the sample description information and a feature vector of the sample search information; and performing feature fusion processing on the feature vector of the sample description information, the feature vector of the sample search information and the comprehensive semantic representation vector by using a preset shallow fusion network to obtain a second sample sorting parameter.

Referring to fig. 10, a block diagram of a sorting apparatus according to an embodiment of the present disclosure is shown.

As shown in fig. 10, the sorting apparatus may include the following modules:

a second obtaining module 1001, configured to obtain search information and a keyword corresponding to an object to be ranked;

the prediction module 1002 is configured to input the search information and the keyword into a pre-trained ranking model, so as to obtain a ranking parameter of the object to be ranked output by the ranking model; the ranking model is obtained by the model training method as described in any of the above embodiments.

And the sorting module 1003 is configured to sort the objects to be sorted based on the sorting parameter.

For the device embodiment, since it is basically similar to the method embodiment, the description is simple, and for the relevant points, refer to the partial description of the method embodiment.

In an embodiment of the present disclosure, an electronic device is also provided. The electronic device may include one or more processors, and one or more computer-readable storage media having instructions, such as an application program, stored thereon. The instructions, when executed by one or more processors, cause the processors to perform a model training method as in any of the embodiments above, or alternatively, a ranking method as in any of the embodiments above.

In an embodiment of the present disclosure, there is also provided a non-transitory computer readable storage medium having stored thereon a computer program executable by a processor of an electronic device, the computer program, when executed by the processor, causing the processor to perform a model training method as described in any of the embodiments above, or to perform a ranking method as described in any of the embodiments above.

The aforementioned processor may be a general-purpose processor, and may include but is not limited to: a Central Processing Unit (CPU), a Network Processor (NP), a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other Programmable logic device, a discrete Gate or transistor logic device, a discrete hardware component, and so on.

The above-mentioned computer readable storage media may include, but are not limited to: read Only Memory (ROM), Random Access Memory (RAM), Compact Disc Read Only Memory (CD-ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), hard disk, floppy disk, flash Memory, and the like.

The algorithms and displays presented herein are not inherently related to any particular computer, virtual machine, or other apparatus. Various general purpose systems may also be used with the teachings herein. The required structure for constructing such a system will be apparent from the description above. In addition, embodiments of the present disclosure are not directed to any particular programming language. It is appreciated that a variety of programming languages may be used to implement the teachings of the embodiments of the present disclosure as described herein, and any descriptions of specific languages are provided above to disclose the best modes of the embodiments of the present disclosure.

In the description provided herein, numerous specific details are set forth. However, it is understood that embodiments of the disclosure may be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.

Similarly, it should be appreciated that in the foregoing description of exemplary embodiments of the disclosure, various features of the embodiments of the disclosure are sometimes grouped together in a single embodiment, figure, or description thereof for the purpose of streamlining the disclosure and aiding in the understanding of one or more of the various inventive aspects. However, the disclosed method should not be interpreted as reflecting an intention that: that is, claimed embodiments of the disclosure require more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive aspects lie in less than all features of a single foregoing disclosed embodiment. Thus, the claims following the detailed description are hereby expressly incorporated into this detailed description, with each claim standing on its own as a separate embodiment of an embodiment of this disclosure.

Those skilled in the art will appreciate that the modules in the devices in an embodiment may be adaptively changed and arranged in one or more devices different from the embodiment. The modules or units or components of the embodiments may be combined into one module or unit or component, and furthermore they may be divided into a plurality of sub-modules or sub-units or sub-components. All of the features disclosed in this specification (including any accompanying claims, abstract and drawings), and all of the processes or elements of any method or apparatus so disclosed, may be combined in any combination, except combinations where at least some of such features and/or processes or elements are mutually exclusive. Each feature disclosed in this specification (including any accompanying claims, abstract and drawings) may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise.

The various component embodiments of the disclosure may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. It will be understood by those skilled in the art that a microprocessor or Digital Signal Processor (DSP) may be used in practice to implement some or all of the functions of some or all of the components in a motion picture generating device according to an embodiment of the present disclosure. Embodiments of the present disclosure may also be implemented as an apparatus or device program for performing a portion or all of the methods described herein. Such programs implementing embodiments of the present disclosure may be stored on a computer readable medium or may be in the form of one or more signals. Such a signal may be downloaded from an internet website or provided on a carrier signal or in any other form.

It should be noted that the above-mentioned embodiments illustrate rather than limit embodiments of the disclosure, and that those skilled in the art will be able to design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. Embodiments of the disclosure may be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by one and the same item of hardware. The usage of the words first, second and third, etcetera do not indicate any ordering. These words may be interpreted as names.

It is clear to those skilled in the art that, for convenience and brevity of description, the specific working processes of the above-described systems, apparatuses and units may refer to the corresponding processes in the foregoing method embodiments, and are not described herein again.

The above description is only a specific implementation of the embodiments of the present disclosure, but the scope of the embodiments of the present disclosure is not limited thereto, and any person skilled in the art can easily conceive of changes or substitutions within the technical scope of the embodiments of the present disclosure, and all the changes or substitutions should be covered by the scope of the embodiments of the present disclosure.

Claims

1. A method of model training, comprising:

acquiring sample data; the sample data comprises sample searching information and sample keywords corresponding to the sample object;

performing feature fusion on the sample keywords and the sample search information in a preset model to be trained to obtain a comprehensive semantic representation vector of the sample keywords, and acquiring sample ordering parameters of the sample object based on the comprehensive semantic representation vector;

and determining a trained model based on the sample sorting parameters as a sorting model.

2. The method of claim 1, wherein feature fusing the sample keyword and the sample search information comprises:

performing feature fusion on the current sample keyword and other sample keywords aiming at each sample keyword to obtain a fusion semantic representation vector of the current sample keyword;

performing feature fusion on the current sample keyword and the sample search information aiming at each sample keyword to obtain the correlation degree between the current sample keyword and the sample search information;

and calculating the comprehensive semantic representation vector of the sample keywords based on the fusion semantic representation vector and the correlation of each sample keyword.

3. The method of claim 1, wherein feature fusing the sample keyword and the sample search information comprises:

inquiring the feature vector of each sample keyword and the feature vector of the sample search information from the corresponding relation between the preset word and the feature vector;

performing feature fusion on the feature vector of the current sample keyword and the feature vectors of other sample keywords aiming at each sample keyword to obtain a fusion semantic representation vector of the current sample keyword;

performing feature fusion on the fusion semantic representation vector of the current sample keyword and the feature vector of the sample search information aiming at each sample keyword to obtain the correlation degree between the current sample keyword and the sample search information;

and calculating a comprehensive semantic representation vector of the sample keywords based on the fusion semantic representation vector and the correlation of each sample keyword and the feature vector of the sample search information.

4. The method of claim 3, wherein feature fusing the feature vector of the current sample keyword with the feature vectors of other sample keywords comprises:

and performing feature fusion on the feature vector of the current sample keyword and the feature vectors of other sample keywords through an attention mechanism to obtain a fusion semantic representation vector of the current sample keyword.

5. The method of claim 3, wherein the sample data further comprises a confidence level for the sample keyword; and performing feature fusion on the feature vector of the current sample keyword and the feature vectors of other sample keywords, wherein the feature fusion comprises the following steps:

performing feature fusion on the feature vector of the current sample keyword and the feature vectors of other sample keywords through a self-attention mechanism to obtain a primary fusion semantic representation vector of the current sample keyword;

sorting the sample keywords in a descending order according to the confidence, and acquiring position embedded vectors of the sample keywords based on a sorting result;

and adding the preliminary fusion semantic representation vector of the current sample keyword and the position embedding vector to obtain a fusion semantic representation vector of the current sample keyword.

6. The method of claim 3, wherein the sample data further comprises a confidence level for the sample keyword; performing feature fusion on the fusion semantic representation vector of the current sample keyword and the feature vector of the sample search information, including:

calculating the weight of the degree of confidence of the current sample keyword and the relevance of the search information;

calculating intermediate parameters of the current sample keywords based on the fusion semantic representation vector of the current sample keywords, the feature vector of the sample search information and the relevancy weight of the current sample keywords;

and calculating the correlation degree between the current sample keyword and the sample search information based on the intermediate parameter and a preset temperature parameter.

7. The method of claim 3, wherein computing a composite semantic representation vector for each sample keyword based on the fused semantic representation vector and relevance for the sample keyword and the feature vector for the sample search information comprises:

calculating the sum of the products of the fusion semantic representation vector and the correlation degree of each sample keyword to obtain a preliminary semantic representation vector of the sample keyword;

adding the preliminary semantic representation vector and the feature vector of the sample search information to obtain a preliminary comprehensive semantic representation vector of the sample keyword;

and carrying out standardization processing on the preliminary comprehensive semantic representation vector to obtain a comprehensive semantic representation vector of the sample keyword.

8. The method of claim 7,

after the sum of the products of the fusion semantic representation vector and the correlation of each sample keyword is calculated to obtain a preliminary semantic representation vector of the sample keyword, the method further comprises the following steps: carrying out random discarding processing on the preliminary semantic representation vector;

adding the preliminary semantic representation vector to a feature vector of the sample search information, comprising: and adding the primary semantic representation vector after the random discarding treatment and the feature vector of the sample searching information.

9. The method of claim 1, wherein the sample data further comprises sample description information corresponding to the sample object; obtaining a sample ordering parameter of the sample object based on the comprehensive semantic representation vector, including:

and performing feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information to obtain a sample ordering parameter of the sample object.

10. The method of claim 9, wherein feature fusing the integrated semantic representation vector, the sample description information, and the sample search information comprises:

performing deep feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information to obtain a first sample ordering parameter of the sample object;

performing shallow feature fusion on the comprehensive semantic representation vector, the sample description information and the sample search information to obtain a second sample ordering parameter of the sample object;

calculating a sample ordering parameter for the sample object based on the first sample ordering parameter and the first sample ordering parameter.

11. The method of claim 10, wherein performing deep feature fusion on the integrated semantic representation vector, the sample description information, and the sample search information comprises:

acquiring a feature vector of the sample description information and a feature vector of the sample search information;

generating a spliced feature vector based on the comprehensive semantic representation vector, the feature vector of the sample description information and the feature vector of the sample search information;

and performing feature fusion processing on the spliced feature vector by using a preset deep fusion network to obtain the first sample sorting parameter.

12. The method of claim 10, wherein performing shallow feature fusion on the integrated semantic representation vector, the sample description information, and the sample search information comprises:

and performing feature fusion processing on the feature vector of the sample description information, the feature vector of the sample search information and the comprehensive semantic representation vector by using a preset shallow fusion network to obtain the second sample sorting parameter.

13. A method of sorting, comprising:

acquiring search information and keywords corresponding to objects to be sorted;

inputting the search information and the keywords into a pre-trained sequencing model to obtain sequencing parameters of the objects to be sequenced, which are output by the sequencing model; the ranking model is obtained by a model training method according to any one of claims 1 to 12;

and sequencing the objects to be sequenced based on the sequencing parameters.

14. A model training apparatus, comprising:

the first acquisition module is used for acquiring sample data; the sample data comprises sample searching information and sample keywords corresponding to the sample object;

the training module is used for performing feature fusion on the sample keywords and the sample search information in a preset model to be trained to obtain a comprehensive semantic representation vector of the sample keywords, and acquiring sample ordering parameters of the sample object based on the comprehensive semantic representation vector;

and the determining module is used for determining the trained model as the sequencing model based on the sample sequencing parameters.

15. A sequencing apparatus, comprising:

the second acquisition module is used for acquiring search information and keywords corresponding to the objects to be sorted;

the prediction module is used for inputting the search information and the keywords into a pre-trained sequencing model to obtain sequencing parameters of the objects to be sequenced, which are output by the sequencing model; the ranking model is obtained by a model training method according to any one of claims 1 to 12;

and the sequencing module is used for sequencing the objects to be sequenced based on the sequencing parameters.

16. An electronic device, comprising:

one or more processors; and

one or more computer-readable storage media having instructions stored thereon;

the instructions, when executed by the one or more processors, cause the processors to perform a model training method as claimed in any one of claims 1 to 12, or to perform a ranking method as claimed in claim 13.

17. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, causes the processor to carry out a model training method as claimed in any one of claims 1 to 12, or a ranking method as claimed in claim 13.