CN113343132A

CN113343132A - Model training method, information display method and device

Info

Publication number: CN113343132A
Application number: CN202110733462.4A
Authority: CN
Inventors: 钟啸林; 刘影
Original assignee: Beijing Sankuai Online Technology Co Ltd
Current assignee: Beijing Sankuai Online Technology Co Ltd
Priority date: 2021-06-30
Filing date: 2021-06-30
Publication date: 2021-09-03
Anticipated expiration: 2041-06-30
Also published as: CN113343132B

Abstract

The specification discloses a model training method, an information display method and an information display device, and a training sample is obtained. Secondly, determining a coding result of the search statement under the compensation action of the first relevant information as a first coding result through a sequencing model to be trained, and determining a coding result of the search statement under the compensation action of the second relevant information corresponding to the search result as a second coding result corresponding to the search result. And finally, obtaining a sequencing result aiming at each search result according to the first coding result and the second coding result corresponding to each search result, and training a sequencing model according to the sequencing result and the label information corresponding to each search result. The method can combine all the characteristics corresponding to the search sentences with one another to carry out coding, and combine all the characteristics corresponding to the search results with one another to carry out coding, thereby further improving the influence of the correlation between the search sentences and the search results on the sequencing results and improving the information browsing experience of the user.

Description

Model training method, information display method and device

Technical Field

The specification relates to the technical field of computers, in particular to a model training method, an information display method and an information display device.

Background

With the continuous development of electronic technology and network technology, a large amount of information exists in the internet, search results displayed to a user can be determined according to search sentences of the user, and the search results displayed on a page are limited, so that the search results need to be ranked, and search results interested by the user are preferentially displayed.

In practical application, the method for ranking search results is often ranked according to the similarity between the search statement input by the user and the search results. In the method for sequencing the search results, the accuracy of the determined correlation between the search terms and the search results is not high, and characteristics that the search terms and the search results are not correlated may occur, so that the output sequencing results cannot meet the actual requirements of the user, and the information browsing experience of the user is poor.

Therefore, how to improve the rationality of the ranking model for ranking the search results is an urgent problem to be solved.

Disclosure of Invention

The present specification provides a model training method, an information displaying method and an information displaying apparatus, so as to partially solve the above problems in the prior art.

The technical scheme adopted by the specification is as follows:

the present specification provides a method of model training, comprising:

acquiring a training sample, wherein the training sample comprises a search statement, search results corresponding to the search statement, first related information and second related information corresponding to the search results, the first related information represents user characteristics of a user who sends the search statement, and the second related information corresponding to the search results represents object characteristics of an object to which the search results belong for each search result;

determining a coding result of the search statement under the compensation action of the first relevant information as a first coding result through a sequencing model to be trained, and determining a coding result of the search statement under the compensation action of second relevant information corresponding to the search result as a second coding result corresponding to the search result;

and obtaining a sequencing result aiming at each search result according to the first coding result and a second coding result corresponding to each search result, and training the sequencing model according to the sequencing result and the label information corresponding to each search result.

Optionally, the ranking model includes an overall correlation submodel, and the overall correlation submodel includes a feature layer and an encoding layer;

determining, by a ranking model to be trained, a coding result of the search statement under a compensation effect of the first relevant information as a first coding result, and a coding result of the search result under a compensation effect of second relevant information corresponding to the search result as a second coding result corresponding to the search result, specifically including:

inputting the search statement and the first related information into the feature layer to obtain a feature vector corresponding to the search statement and a feature vector corresponding to the first related information, and inputting the feature vector corresponding to the search statement and the feature vector corresponding to the first related information into the coding layer for coding to obtain the first coding result;

and inputting the search result and the second related information corresponding to the search result into the feature layer to obtain a feature vector corresponding to the search result and a feature vector of the second related information corresponding to the search result, and inputting the feature vector corresponding to the search result and the feature vector of the second related information corresponding to the search result into the coding layer for coding to obtain a second coding result corresponding to the search result.

Optionally, the ranking model further comprises a feature association submodel;

before obtaining a ranking result for each search result according to the first encoding result and the second encoding result corresponding to each search result, the method further includes:

for each search result, inputting the search statement, the search result, the first related information and second related information corresponding to the search result into the feature association submodel as target information, so as to determine auxiliary correlation between the search statement and the search result on the premise of determining the correlation between at least part of feature dimensions involved in the target information;

obtaining a ranking result for each search result according to the first encoding result and a second encoding result corresponding to each search result, which specifically includes:

and obtaining a sequencing result aiming at each search result according to the first coding result, the second coding result corresponding to each search result and the auxiliary correlation.

Optionally, the feature association submodel includes: a full-scale feature correlation model;

on the premise of determining the relevance between at least part of feature dimensions involved in the target information, determining the auxiliary relevance between the search statement and the search result, specifically comprising:

determining each feature dimension combination, wherein each feature dimension combination comprises at least one feature dimension related to the target information;

for each characteristic dimension combination, determining the relevance between the characteristic dimension combination and each other characteristic dimension combination through the full-quantity characteristic relevance model;

and determining the auxiliary correlation between the search result and the search statement according to the determined correlation between the feature dimension combinations.

Optionally, the feature association submodel includes: setting a characteristic correlation model;

determining feature information of each set feature dimension from the target information, wherein each set feature dimension is a part of feature dimensions in all feature dimensions related to the target information;

inputting the feature information into the set feature association model to determine, for each set feature dimension, an association between the set feature dimension and each of the other set feature dimensions;

and determining the auxiliary correlation between the search result and the search statement according to the determined correlation between the set characteristic dimensions.

Optionally, obtaining a ranking result for each search result according to the first encoding result, the second encoding result corresponding to each search result, and the auxiliary correlation, specifically includes:

for each search result, determining the correlation between the search result and the search sentence as comprehensive correlation under the condition of integrally coding the first related information and the search sentence and integrally coding the second related information corresponding to the search result and the search result according to the first coding result and the second coding result corresponding to the search result;

determining a ranking score for the search result according to the composite relevance and the auxiliary relevance;

and determining the ranking result according to the ranking score corresponding to each search result.

Optionally, training the ranking model according to the ranking results and the label information corresponding to each search result, specifically including:

determining a ranking result evaluation score of the ranking result as an original ranking evaluation score;

adjusting the sorting result according to the clicked search result determined based on the tag information to obtain an adjusted sorting result, and determining a sorting result evaluation score corresponding to the adjusted sorting result as an optimal sorting evaluation score;

and training the ranking model by taking the deviation between the minimized original ranking evaluation score and the optimal ranking evaluation score as an optimization target.

The present specification provides a method of information presentation, comprising:

receiving a search sentence input by a user;

determining each search result corresponding to the search statement;

inputting the search results into a pre-trained sequencing model to obtain sequencing results corresponding to the search sentences, wherein the sequencing model is obtained by training through the model training method;

and displaying information to the user according to the sequencing result.

The present specification provides an apparatus for model training, comprising:

an obtaining module, configured to obtain a training sample, where the training sample includes a search statement, search results corresponding to the search statement, first related information, and second related information corresponding to the search results, where the first related information represents a user characteristic of a user who sends the search statement, and for each search result, the second related information corresponding to the search result represents an object characteristic of an object to which the search result belongs;

the coding module is used for determining a coding result of the search statement under the compensation action of the first relevant information as a first coding result through a sequencing model to be trained, and determining a coding result of the search statement under the compensation action of second relevant information corresponding to the search result as a second coding result corresponding to the search result;

and the training module is used for obtaining a sequencing result aiming at each search result according to the first coding result and the second coding result corresponding to each search result, and training the sequencing model according to the sequencing result and the label information corresponding to each search result.

This specification provides an apparatus for information presentation, comprising:

the receiving module is used for receiving a search statement input by a user;

the determining module is used for determining each search result corresponding to the search statement;

the ranking module is used for inputting the search results into a pre-trained ranking model to obtain ranking results corresponding to the search sentences, and the ranking model is obtained by training through the model training method;

and the display module is used for displaying the information to the user according to the sequencing result.

The present specification provides a computer-readable storage medium storing a computer program which, when executed by a processor, implements the above-described method of model training and method of information presentation.

The present specification provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, the processor implements the above model training method and information presentation method.

The technical scheme adopted by the specification can achieve the following beneficial effects:

in the model training method and the information display method provided in this specification, a training sample is obtained, where the training sample includes a search statement, search results corresponding to the search statement, first related information, and second related information corresponding to the search results, the first related information represents a user characteristic of a user who sends the search statement, and for each search result, the second related information corresponding to the search result represents an object characteristic of an object to which the search result belongs. Secondly, determining a coding result of the search statement under the compensation action of the first relevant information as a first coding result through a sequencing model to be trained, and determining a coding result of the search statement under the compensation action of the second relevant information corresponding to the search result as a second coding result corresponding to the search result. And finally, obtaining a sequencing result aiming at each search result according to the first coding result and the second coding result corresponding to each search result, and training a sequencing model according to the sequencing result and the label information corresponding to each search result.

It can be seen from the above method that the method can determine the ranking result for each search result according to the coding result of the search statement under the compensation action of the first related information and the coding result of the search result under the compensation action of the second related information corresponding to the search result. That is to say, all the features corresponding to the search sentences are combined and encoded integrally, and all the features corresponding to the search results are combined and encoded integrally, so that the influence of the correlation between the search sentences and each search result on the ranking results is further improved, the condition that the search results with low correlation between the search sentences are ranked in the front is avoided as much as possible, the ranking results corresponding to the search sentences are more reasonable, and the search results with high correlation are displayed to the user in the page, so that the information browsing experience of the user is improved.

Drawings

The accompanying drawings, which are included to provide a further understanding of the specification and are incorporated in and constitute a part of this specification, illustrate embodiments of the specification and together with the description serve to explain the specification and not to limit the specification in a non-limiting sense. In the drawings:

FIG. 1 is a schematic flow chart of a method of model training in the present specification;

FIG. 2 is a diagram illustrating a model structure of a ranking model provided in an embodiment of the present disclosure;

FIG. 3 is a diagram illustrating a model structure of a ranking model provided in an embodiment of the present disclosure;

FIG. 4 is a flow chart of a method for displaying information in the present specification;

FIG. 5 is a schematic diagram of an apparatus for model training provided herein;

FIG. 6 is a schematic view of an information presentation device provided herein;

fig. 7 is a schematic diagram of an electronic device corresponding to fig. 1 provided in the present specification.

Detailed Description

In order to make the objects, technical solutions and advantages of the present disclosure more clear, the technical solutions of the present disclosure will be clearly and completely described below with reference to the specific embodiments of the present disclosure and the accompanying drawings. It is to be understood that the embodiments described are only a few embodiments of the present disclosure, and not all embodiments. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments in the present specification without any creative effort belong to the protection scope of the present specification.

The technical solutions provided by the embodiments of the present description are described in detail below with reference to the accompanying drawings.

Fig. 1 is a schematic flow chart of a method for model training in this specification, which includes the following steps:

s100: the method comprises the steps of obtaining a training sample, wherein the training sample comprises a search statement, each search result corresponding to the search statement, first related information and second related information corresponding to each search result, the first related information represents user characteristics of a user sending the search statement, and the second related information corresponding to each search result represents object characteristics of an object to which the search result belongs.

In the embodiment of the present specification, the execution subject for training the ranking model may be a server, or may be an electronic device such as a desktop computer, and for convenience of description, the method for training the ranking model provided in the present specification will be described below with only the server as the execution subject.

In this embodiment, the server may obtain a training sample, where the training sample includes a search statement, search results corresponding to the search statement, first related information, and second related information corresponding to the search results, and the first related information represents a user characteristic of a user who sends the search statement. And aiming at each search result, the second related information corresponding to the search result represents the object characteristics of the object to which the search result belongs.

The search sentence can refer to a search sentence or a search keyword and the like input by the user according to actual needs. The first related information is used for reflecting the corresponding characteristics of the user when the user sends the search statement, such as the geographical position of the user when the user sends the search statement, the personal preference of the user, the business category corresponding to the information clicked historically by the user, and the like. Of course, the first related information may also include the corresponding self-attribute of the search statement, such as the corresponding category in the search statement (the search statement is a chicken bouillon, and the category is a dish).

The second related information is used to reflect the personalized characteristics of the object to which the search result belongs, for example, if the object to which the search result belongs is a business, the second related information may be used to indicate the area where the business is located, the business name of the business, and the like. The second related information may further include a category corresponding to the search result (for example, when the search statement is a palace chicken dice, the searched search result may include two search results, namely a dish of the palace chicken dice and a related comment of the palace chicken dice, and the two search results belong to different categories), a descriptive text corresponding to the search result (for example, if the search result is a dish a, the descriptive text corresponding to the search result may include a taste, a preparation material, and the like of the dish a), and the like.

The training sample further includes label information corresponding to each search result, and the label information mentioned here may be used to represent an actual browsing situation of the user corresponding to each search result in history, such as whether the user clicks on the search result corresponding to the search statement. Of course, the label information may be manually set through practical experience.

S102: and determining a coding result of the search statement under the compensation action of the first relevant information as a first coding result through a sequencing model to be trained, and determining a coding result of the search statement under the compensation action of the second relevant information corresponding to the search result as a second coding result corresponding to the search result.

In this embodiment, the server may determine, through the ranking model to be trained, an encoding result of the search statement under the compensation effect of the first relevant information as a first encoding result, and an encoding result of the search result under the compensation effect of the second relevant information corresponding to the search result as a second encoding result corresponding to the search result.

In practical application, the server sorts the search results according to the similarity between the text information corresponding to the search sentence input by the user and each text information corresponding to the search result, and the accuracy of the correlation between the search sentence and the search result determined in this way is low, which may cause that the output sorted result cannot meet the actual requirements of the user. Therefore, in the process of model training, the server can determine the similarity between the characteristics according to the similarity between the text information corresponding to the search sentence and the text information corresponding to the search result, the user characteristics of the user corresponding to the search sentence and the object characteristics of the object to which the search result belongs, and further determine the correlation between the search sentence and the search result so as to improve the accuracy of the correlation between the search sentence and the search result.

In the embodiment of the present specification, the order model includes an overall correlation submodel, and the overall correlation submodel includes a feature layer and an encoding layer. The server may input the search statement and the first related information to the feature layer to obtain a feature vector corresponding to the search statement and a feature vector corresponding to the first related information, and input the feature vector corresponding to the search statement and the feature vector corresponding to the first related information to the coding layer for coding to obtain a first coding result. And inputting the search result and the second related information corresponding to the search result into a feature layer to obtain a feature vector corresponding to the search result and a feature vector of the second related information corresponding to the search result, and inputting the feature vector corresponding to the search result and the feature vector of the second related information corresponding to the search result into an encoding layer for encoding to obtain a second encoding result corresponding to the search result.

In this embodiment, the server may extract, from the first related information, a user feature of a user corresponding to the search term in the overall related sub-model, and extract, from the second related information corresponding to the search result, an object feature of an object to which the search result belongs in the overall related sub-model.

Secondly, inputting the search statement and the user characteristics of the user corresponding to the search statement in the overall relevant submodel into a characteristic layer of the overall relevant submodel to obtain a characteristic vector corresponding to the search statement and a characteristic vector of the user characteristics of the user corresponding to the search statement. And inputting the feature vector corresponding to the search statement and the feature vector of the user feature corresponding to the search statement into the coding layer of the overall relevant submodel to obtain a first coding result obtained by combining the search statement and the user feature of the user corresponding to the search statement and carrying out overall coding.

And inputting the search result and the object characteristics of the object to which the search result belongs into a characteristic layer of the overall related submodel to obtain a characteristic vector corresponding to the search result and a characteristic vector of the object characteristics of the object to which the search result belongs. And inputting the feature vector corresponding to the search result and the feature vector of the object feature of the object to which the search result belongs into an encoding layer of an overall related submodel to obtain a second encoding result corresponding to the search result after the search result is combined with the object feature of the object to which the search result belongs and the overall encoding is performed.

And finally, determining the correlation between the search statement and the search result as comprehensive correlation according to the first coding result and the second coding result corresponding to the search result, wherein if the similarity between the first coding result and the second coding result corresponding to the search result is higher, the comprehensive correlation is higher. Because the first coding result and the second coding result corresponding to the search result are obtained by further fusing a plurality of features extracted from the first related information used for representing the features of the user and the second related information used for representing the object features corresponding to the search result into a coding process on the basis of the search result and the search sentence text, the relevance between the search sentence determined in the mode and the search result can be effectively embodied, and the accuracy of the relevance between the determined search sentence and the search result is ensured.

In this embodiment of the present specification, the server may determine, through a sub-Neural network included in the overall correlation sub-model in the ranking model, a feature vector of a user feature of a user corresponding to the search statement and a feature vector of an object feature of an object to which the search result belongs, where the sub-Neural network may be a Convolutional Neural Network (CNN) or a Recurrent Neural Network (RNN). The manner in which the server determines the encoding result may be various. For example, the server may add and sum the feature vector corresponding to the search term and the feature vector of the user feature of the user corresponding to the search term, and perform pooling operation through the pooling layer to obtain the first encoding result. Similarly, the server may add and sum the feature vector corresponding to the search result and the feature vector of the object feature of the object to which the search result belongs, and perform pooling operation through the pooling layer to obtain the second encoding result corresponding to the search result.

For another example, the server may splice the feature vector corresponding to the search statement and the feature vector of the user feature of the user corresponding to the search statement, and unify the dimensions of the feature vectors through the full connection layer to obtain the first encoding result. Similarly, the server may splice the feature vector corresponding to the search result with the feature vector of the object feature of the object to which the search result belongs, and unify the dimensions of the feature vectors through the full connection layer to obtain the second encoding result corresponding to the search result. The mode of determining the overall characteristic information is not limited in the present specification.

In embodiments of the present description, the manner in which the server determines the integrated correlations may be varied. For example, the server may calculate, by using a Cosine similarity (Cosine similarity) operation manner, a first encoding result corresponding to the search term and a second encoding result corresponding to the search result to obtain an included angle Cosine value between the search term and the search result, where the greater the included angle Cosine value, the greater the comprehensive correlation between the search term and the search result.

For another example, the server may calculate, by using a Hadamard product (Hadamard product) operation manner, a first coding result corresponding to the search statement and a second coding result corresponding to the search result to obtain a Hadamard product between the search statement and the search result, where the greater the Hadamard product is, the greater the comprehensive correlation between the search statement and the search result is. The manner in which the overall correlation is determined is not limited in this specification.

By the method, the server can enable the correlation between the search result displayed to the user and the search sentence input by the user to be strong by improving the comprehensive correlation between the search sentence and the search result, avoid the situation that the search result is inconsistent with the search sentence, and enable the sequencing result corresponding to the search sentence to be determined to meet the actual requirement of the user.

In this embodiment of the present specification, the ranking model further includes a feature association submodel, and the server may input, for each search result, the search statement, the search result, the first relevant information, and the second relevant information corresponding to the search result as target information into the feature association submodel, to determine an auxiliary relevance between the search statement and the search result on the premise of determining a relevance between at least some feature dimensions involved in the target information, and obtain a ranking result for each search result according to the first encoding result, the second encoding result corresponding to each search result, and the auxiliary relevance. As shown in fig. 2.

Fig. 2 is a schematic diagram of a model structure of an ordering model provided in an embodiment of the present specification.

In fig. 2, the server may input, for each search result, a search statement, the search result, first relevant information, and second relevant information corresponding to the search result as target information into the ranking model, determine a comprehensive relevance between the search statement and the search result through the overall relevance submodel, and determine a corresponding auxiliary relevance between the search statement and the search result through the feature relevance submodel. And the server determines the sequencing result aiming at each search result according to the comprehensive relevance and the auxiliary relevance. There are a variety of ways to determine the ranking results of the search results mentioned herein. For example, the server may directly add the comprehensive correlation and the auxiliary correlation for summation, determine a summation result corresponding to each search result, and obtain a ranking result for each search result according to the summation result corresponding to each search result.

For another example, the server may input the comprehensive relevance and the auxiliary relevance to a preset weighting layer, and obtain a more accurate ranking result of each search result by adjusting the weights corresponding to the comprehensive relevance and the auxiliary relevance in advance. The method for determining the ranking result of each search result is not limited in this specification.

In the embodiment of the present specification, the feature association submodel includes: and the server can determine each characteristic dimension combination, and each characteristic dimension combination comprises at least one characteristic dimension related in the target information. The feature dimension referred to herein may be a user feature of a user of a search statement or an object feature of an object to which a search result is attributed. And determining the relevance between the characteristic dimension combination and each other characteristic dimension combination through a full-quantity characteristic relevance model aiming at each characteristic dimension combination, and determining the auxiliary relevance between the search result and the search statement according to the determined relevance between the characteristic dimension combinations.

The server can input each feature dimension corresponding to the search statement and each feature dimension corresponding to the search result into the full-scale feature association model, and determine the auxiliary correlation between the search statement and the search result. And if the high-order feature combinations determined by the full-quantity feature association model are more among the feature dimensions corresponding to the search sentence and the feature dimensions corresponding to the search result, the auxiliary relevance between the search sentence and the search result is higher.

The high-order feature combination mentioned here may refer to determining a feature combination with high relevance between feature dimension combinations through a full-scale feature association model in a training process. That is to say, the high-order feature combination can be understood as a high-order implicit feature association capability learned by the full-scale feature association model through association between one sentence and one sentence, and the association between different feature dimension combinations is explored through calculation between a plurality of feature dimensions and a plurality of feature dimensions.

In practical application, the server can determine the number of high-order feature combinations between the search statement and the search result through the full-scale feature association model, and accordingly determine the auxiliary relevance between the search statement and the search result.

Specifically, the server may determine the higher-order feature combination between the search statement and the search result in various ways. For example, the server may input a search statement and the search result into a full-scale feature association model, and determine each feature dimension combination. Secondly, aiming at each determined feature dimension combination, multiplying the feature dimension combination with each other feature dimension combination, determining the relevance between the feature dimension combinations, and training a ranking model by taking the deviation between a ranking result corresponding to a minimized search statement and an optimal ranking result corresponding to the search statement determined based on label information as an optimization target so as to obtain an accurate high-order feature combination between the feature dimension combinations.

For another example, the server may determine the relevance between the feature dimension combinations through a Deep Neural Network (DNN) included in the full-scale feature association model, so as to obtain accurate high-order feature combinations between the feature dimension combinations. By the method, the server can determine the relevance among the feature dimension combinations, so that the sequencing result of the sequencing model is more diversified.

It should be noted that, because the full-scale feature association model can determine the association between feature dimension combinations in the training process, if the number of samples corresponding to one feature dimension is very small, interference may be caused to the training of the full-scale feature association model, and the training efficiency of the full-scale feature association model is reduced. Therefore, the server can preprocess the training sample before selecting the feature dimension combination, and remove the feature dimension in the training sample, so as to improve the training efficiency of the full-scale feature association model.

In practical application, the full-quantity feature association model in the ranking model determines that the high-order feature combinations among the feature dimension combinations may be wrong, so that the accuracy of the obtained ranking result corresponding to the search statement is low. In order to improve the accuracy of the ranking result corresponding to the search statement, the server can improve the accuracy of the ranking result corresponding to the search statement through a preset set relationship between the feature dimension combination corresponding to the search statement and the feature dimension combination corresponding to the search result.

In the embodiment of the present specification, the feature association submodel includes: the method comprises the steps that a characteristic association model is set, a server can determine characteristic information of each set characteristic dimension from target information, each set characteristic dimension is part of all characteristic dimensions related to the target information, the characteristic information is input into the set characteristic association model, association between the set characteristic dimension and each other set characteristic dimension is determined according to each set characteristic dimension, and auxiliary association between a search result and a search statement is determined according to the determined association between the set characteristic dimensions. The set feature dimension mentioned here may be each feature dimension having a set relationship determined by human experience, that is, the set relationship between a single feature and a single feature set by human is used to ensure that the relevance between the search statement and the search result is not too low.

The server may input each set feature dimension corresponding to the search term and each set feature dimension corresponding to the search result into the set feature association model, and determine an auxiliary correlation between the search term and the search result. The method comprises the steps of determining set characteristic dimensions with set relations between set characteristic dimensions corresponding to a search statement and set characteristic dimensions corresponding to a search result according to the predetermined set relations between characteristic dimension combinations, wherein if the set characteristic dimensions with the set relations between the set characteristic dimensions corresponding to the search statement and the set characteristic dimensions corresponding to the search result are determined, the more the set characteristic dimensions with the set relations determined by a set characteristic association model are, the higher the auxiliary relevance between the search statement and the search result is.

By the method, the server can set the important characteristic dimension determined empirically between the search sentence and the search result, and train the ranking model based on the set characteristic dimension with the set relationship in the training process, so that the convergence speed of the ranking model is increased, and the training efficiency of the ranking model is improved.

It should be noted that the server may determine the auxiliary relevance between the search result and the search sentence together by the relevance between each feature dimension combination determined by the full-scale feature association model and the relevance between each set feature dimension determined by the set feature association model. As shown in fig. 3.

Fig. 3 is a schematic diagram of a model structure of a ranking model provided in an embodiment of the present specification.

As can be seen from fig. 3, the feature association submodel may include a full-scale feature association model and a set feature association model, and the full-scale feature association model and the set feature association model jointly determine an auxiliary correlation corresponding to the feature association submodel, and in the above method, the comprehensive correlation output by the overall correlation submodel is taken as a main body in the ranking model, and the auxiliary correlation is used to further improve the accuracy and diversity of the ranking result. That is, the comprehensive correlation between the search sentence and the search result has a strong influence on the ranking result obtained by the ranking model.

In this embodiment, the feature association submodel may have various forms, for example, a Factorization-based predictive Neural Network for CTR Prediction, Deep memory Network (Wide & Deep Learning for Recommander Systems), Deep Cross Network (Deep & Cross Network, DCN), and the like, and this specification does not specifically limit the feature association submodel.

In this embodiment, the server may input the feature dimension corresponding to the search statement and the feature dimension corresponding to the search result into the feature association submodel, and the feature association submodel cannot distinguish whether the feature dimension belongs to the search statement or the search result, so that the association between the feature dimensions determined by the feature association submodel may perform invalid calculation on each feature dimension corresponding to the search statement or each feature dimension corresponding to the search result, which may cause inaccuracy of the auxiliary correlation output by the feature association submodel, and the comprehensive correlation output by the overall correlation submodel is the first encoded result corresponding to the search statement and the second encoded result corresponding to the search result to perform calculation, thereby avoiding mutual calculation between the feature dimensions related to the search statement itself, and the mutual calculation is carried out among all characteristic dimensions related to the search result, so that the training efficiency of the sequencing model is improved, and the reasonability of the sequencing result determined by the sequencing model is further improved.

S104: and obtaining a sequencing result aiming at each search result according to the first coding result and a second coding result corresponding to each search result, and training the sequencing model according to the sequencing result and the label information corresponding to each search result.

In this embodiment, the server may obtain a ranking result for each search result according to the first encoding result and the second encoding result corresponding to each search result, and train the ranking model according to the ranking result and the tag information corresponding to each search result. For each search result, the higher the comprehensive relevance between the search sentence and the search result is, the higher the ranking of the search result is, the higher the auxiliary relevance between the search sentence and the search result is, and the higher the ranking of the search result is.

In this specification embodiment, the server may determine, for each search result, a correlation between the search result and the search sentence as an integrated correlation in a case where the first related information is integrally encoded with the search sentence and the second related information corresponding to the search result is integrally encoded with the search result, according to the first encoding result and the second encoding result corresponding to the search result. Second, a ranking score for the search result is determined based on the composite relevance and the auxiliary relevance. And finally, determining a sorting result according to the sorting score corresponding to each search result.

In an embodiment of the present specification, the server may determine a ranking result evaluation score of the ranking result as an original ranking evaluation score. The original ranking evaluation score mentioned here can be used to measure the overall quality of the ranking result corresponding to the search sentence. And adjusting the sorting result according to the clicked search result determined based on the tag information to obtain an adjusted sorting result, and determining a sorting result evaluation score corresponding to the adjusted sorting result as an optimal sorting evaluation score. And training a sequencing model by taking the deviation between the minimized original sequencing evaluation score and the optimal sequencing evaluation score as an optimization target. Through multiple rounds of iterative training, the deviation can be continuously reduced and converged in a numerical range, and then the training of the sequencing model is completed.

Specifically, the ranking result evaluation score may be determined according to a ranking position where the search result is located, the closer the ranking position where the search result is located is, the higher the evaluation score corresponding to the search result is, and the evaluation scores corresponding to the search results corresponding to the search sentences are summed to obtain the ranking result evaluation score corresponding to the ranking result.

The above-mentioned optimal sorting result may be that, according to tag information including the click condition of each search result, according to comprehensive relevance and auxiliary relevance, a sorting score for each search result is determined, and sorting is performed in order from high to low, and then the search result clicked by the user in the tag information is adjusted to the front position in the sorting result, so as to determine the adjusted sorting result corresponding to the search statement. And the server determines the optimal ranking evaluation score according to the adjusted ranking result. Of course, the server may also determine the optimal ranking evaluation score as the optimal ranking result according to a ranking result determined in advance by a human.

In the process, the relevance between the search statement and the search result is determined not to be too low according to the set relationship between the features determined by the set feature relevance model through human experience. Through a full-quantity feature association model, the association between feature dimension combinations (the feature dimension combinations comprise a plurality of feature dimensions) can be learned through a training mode, namely, the implicit association between a search statement and each search result is found, so that the sequencing results corresponding to the search statement are more diversified. The method has a certain degree of relevance and diversification between the search sentences and the search results determined by setting the feature association model and the full-scale feature association model. However, the determined degree of match between the search sentence and the search result may not be sufficiently accurate. Therefore, the server can determine the comprehensive correlation between the search sentences and the search results through the overall correlation sub-model, so that the influence of the correlation between the search sentences and the search results on the ranking results is further improved, the condition that the search results with low correlation between the search sentences are ranked in the front is avoided as much as possible, the ranking results corresponding to the search sentences are more reasonable, and the information browsing experience of the user is further improved.

After training of the ranking model is completed, the embodiment of the present specification may display information to a user through the ranking model, and a specific process is shown in fig. 4.

Fig. 4 is a flowchart illustrating a method for displaying information in the present specification.

S400: a search sentence input by a user is received.

S402: and determining each search result corresponding to the search statement.

S404: and inputting the search results into a pre-trained sequencing model to obtain sequencing results corresponding to the search sentences, wherein the sequencing model is obtained by training through the model training method.

S406: and displaying information to the user according to the sequencing result.

In this embodiment, the server may receive a search statement input by a user, and determine each search result corresponding to the search statement. And determining the first related information, each search result and second related information corresponding to each search result according to the search statement. Secondly, the server can input the search statement, the first related information, each search result and the second related information corresponding to each search result into a pre-trained ranking model to obtain a ranking result corresponding to the search statement, wherein the ranking model is obtained by training through the model training method. And finally, the server can display information to the user according to the sequencing result.

In this embodiment, the server may input the search term and each search result into the ranking model, determine the feature vector corresponding to the search term and each search result, and determine the ranking result corresponding to the search term according to the feature vector corresponding to the search term and each search result. The manner of determining the ranking result corresponding to the search sentence is basically the same as the manner mentioned in the above model training process, and is not described in detail here.

As can be seen from the above, first, the server determines the comprehensive relevance and the auxiliary relevance of the search term and each search result according to the ranking model. And then, the server determines the sequencing result corresponding to the search statement according to the comprehensive correlation and the auxiliary correlation of the search statement and each search result. Therefore, the comprehensive relevance can ensure that the relevance between the search result with the top ranking position in the ranking result and the search sentence is high, the accuracy of the relevance between the search result and the search sentence in the ranking result is ensured by the auxiliary relevance, and the diversity of the ranking result is ensured, thereby further improving the information browsing experience of the user.

Based on the same idea, the present specification further provides a corresponding model training apparatus and an information displaying apparatus, as shown in fig. 5, for the method for model training and the method for information displaying provided in one or more embodiments of the present specification.

Fig. 5 is a schematic diagram of an apparatus for model training provided in the present specification, including:

an obtaining module 500, configured to obtain a training sample, where the training sample includes a search statement, search results corresponding to the search statement, first related information, and second related information corresponding to the search results, where the first related information represents a user characteristic of a user who sends the search statement, and for each search result, the second related information corresponding to the search result represents an object characteristic of an object to which the search result belongs;

the encoding module 502 is configured to determine, through a ranking model to be trained, an encoding result of the search statement under a compensation effect of the first relevant information as a first encoding result, and an encoding result of the search result under a compensation effect of second relevant information corresponding to the search result as a second encoding result corresponding to the search result;

a training module 504, configured to obtain a ranking result for each search result according to the first encoding result and the second encoding result corresponding to each search result, and train the ranking model according to the ranking result and the tag information corresponding to each search result.

Optionally, the encoding module 502 is specifically configured to, where the sorting model includes an overall related submodel, where the overall related submodel includes a feature layer and an encoding layer, input the search statement and the first related information into the feature layer to obtain a feature vector corresponding to the search statement and a feature vector corresponding to the first related information, input the feature vector corresponding to the search statement and the feature vector corresponding to the first related information into the encoding layer for encoding to obtain the first encoding result, input the search result and the second related information corresponding to the search result into the feature layer to obtain the feature vector corresponding to the search result and the feature vector corresponding to the second related information corresponding to the search result, and input the feature vector corresponding to the search result and the feature vector corresponding to the second related information corresponding to the search result into the encoding layer for encoding, and obtaining a second coding result corresponding to the search result.

Optionally, the encoding module 502 is specifically configured to, for each search result, input the search statement, the search result, the first relevant information, and second relevant information corresponding to the search result as target information into the feature association submodel, so as to determine an auxiliary correlation between the search statement and the search result on the premise that the correlation between at least some feature dimensions involved in the target information is determined, and obtain the ranking result for each search result according to the first encoding result, the second encoding result corresponding to each search result, and the auxiliary correlation.

Optionally, the encoding module 502 is specifically configured to, the feature association submodel includes: and determining each feature dimension combination by a full-scale feature association model, wherein each feature dimension combination comprises at least one feature dimension related to the target information, determining the association between the feature dimension combination and each other feature dimension combination by the full-scale feature association model aiming at each feature dimension combination, and determining the auxiliary association between the search result and the search statement according to the determined association between each feature dimension combination.

Optionally, the encoding module 502 is specifically configured to, the feature association submodel includes: setting a feature association model, determining feature information of each set feature dimension from the target information, wherein each set feature dimension is part of all feature dimensions related to the target information, inputting the feature information into the set feature association model to determine the association between each set feature dimension and each other set feature dimension for each set feature dimension, and determining the auxiliary association between the search result and the search statement according to the determined association between each set feature dimension.

Optionally, the training module 504 is specifically configured to, for each search result, determine, according to the first encoding result and the second encoding result corresponding to the search result, that a correlation between the search result and the search sentence is used as a comprehensive correlation when the first related information and the search sentence are integrally encoded, and the second related information corresponding to the search result and the search result are integrally encoded, determine a ranking score for the search result according to the comprehensive correlation and the auxiliary correlation, and determine the ranking result according to the ranking score corresponding to each search result.

Optionally, the training module 504 is specifically configured to determine a ranking result evaluation score of the ranking result as an original ranking evaluation score, adjust the ranking result according to a clicked search result determined based on the tag information to obtain an adjusted ranking result, determine a ranking result evaluation score corresponding to the adjusted ranking result as an optimal ranking evaluation score, and train the ranking model with a minimization of a deviation between the original ranking evaluation score and the optimal ranking evaluation score as an optimization objective.

Fig. 6 is a schematic diagram of an information displaying apparatus provided in the present specification, including:

a receiving module 600, configured to receive a search statement input by a user;

a determining module 602, configured to determine search results corresponding to the search statement;

a ranking module 604, configured to input each search result into a pre-trained ranking model, to obtain a ranking result corresponding to the search statement, where the ranking model is obtained by training through the model training method;

and a display module 606, configured to display information to the user according to the sorting result.

The present specification also provides a computer-readable storage medium storing a computer program, which can be used to execute a method of model training and a method of information presentation provided in fig. 1.

This description also provides a schematic block diagram of an electronic device corresponding to that of fig. 1, shown in fig. 7. As shown in fig. 7, at the hardware level, the electronic device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, but may also include hardware required for other services. The processor reads a corresponding computer program from the non-volatile memory into the memory and then runs the computer program to implement the model training method and the information displaying method described in fig. 1. Of course, besides the software implementation, the present specification does not exclude other implementations, such as logic devices or a combination of software and hardware, and the like, that is, the execution subject of the following processing flow is not limited to each logic unit, and may be hardware or logic devices.

In the 90 s of the 20 th century, improvements in a technology could clearly distinguish between improvements in hardware (e.g., improvements in circuit structures such as diodes, transistors, switches, etc.) and improvements in software (improvements in process flow). However, as technology advances, many of today's process flow improvements have been seen as direct improvements in hardware circuit architecture. Designers almost always obtain the corresponding hardware circuit structure by programming an improved method flow into the hardware circuit. Thus, it cannot be said that an improvement in the process flow cannot be realized by hardware physical modules. For example, a Programmable Logic Device (PLD), such as a Field Programmable Gate Array (FPGA), is an integrated circuit whose Logic functions are determined by programming the Device by a user. A digital system is "integrated" on a PLD by the designer's own programming without requiring the chip manufacturer to design and fabricate application-specific integrated circuit chips. Furthermore, nowadays, instead of manually making an Integrated Circuit chip, such Programming is often implemented by "logic compiler" software, which is similar to a software compiler used in program development and writing, but the original code before compiling is also written by a specific Programming Language, which is called Hardware Description Language (HDL), and HDL is not only one but many, such as abel (advanced Boolean Expression Language), ahdl (alternate Hardware Description Language), traffic, pl (core universal Programming Language), HDCal (jhdware Description Language), lang, Lola, HDL, laspam, hardward Description Language (vhr Description Language), vhal (Hardware Description Language), and vhigh-Language, which are currently used in most common. It will also be apparent to those skilled in the art that hardware circuitry that implements the logical method flows can be readily obtained by merely slightly programming the method flows into an integrated circuit using the hardware description languages described above.

The controller may be implemented in any suitable manner, for example, the controller may take the form of, for example, a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro) processor, logic gates, switches, an Application Specific Integrated Circuit (ASIC), a programmable logic controller, and an embedded microcontroller, examples of which include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320, the memory controller may also be implemented as part of the control logic for the memory. Those skilled in the art will also appreciate that, in addition to implementing the controller as pure computer readable program code, the same functionality can be implemented by logically programming method steps such that the controller is in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers and the like. Such a controller may thus be considered a hardware component, and the means included therein for performing the various functions may also be considered as a structure within the hardware component. Or even means for performing the functions may be regarded as being both a software module for performing the method and a structure within a hardware component.

The systems, devices, modules or units illustrated in the above embodiments may be implemented by a computer chip or an entity, or by a product with certain functions. One typical implementation device is a computer. In particular, the computer may be, for example, a personal computer, a laptop computer, a cellular telephone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

For convenience of description, the above devices are described as being divided into various units by function, and are described separately. Of course, the functions of the various elements may be implemented in the same one or more software and/or hardware implementations of the present description.

As will be appreciated by one skilled in the art, embodiments of the present description may be provided as a method, system, or computer program product. Accordingly, the description may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the description may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

The description has been presented with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the description. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

In a typical configuration, a computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.

The memory may include forms of volatile memory in a computer readable medium, Random Access Memory (RAM) and/or non-volatile memory, such as Read Only Memory (ROM) or flash memory (flash RAM). Memory is an example of a computer-readable medium.

Computer-readable media, including both non-transitory and non-transitory, removable and non-removable media, may implement information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), Read Only Memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), Digital Versatile Discs (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. As defined herein, a computer readable medium does not include a transitory computer readable medium such as a modulated data signal and a carrier wave.

It should also be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.

This description may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media including memory storage devices.

The embodiments in the present specification are described in a progressive manner, and the same and similar parts among the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments. In particular, for the system embodiment, since it is substantially similar to the method embodiment, the description is simple, and for the relevant points, reference may be made to the partial description of the method embodiment.

The above description is only an example of the present specification, and is not intended to limit the present specification. Various modifications and alterations to this description will become apparent to those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present specification should be included in the scope of the claims of the present specification.

Claims

1. A method of model training, comprising:

2. The method of claim 1, wherein the order model comprises an overall correlation submodel, the overall correlation submodel comprising a feature layer and an encoding layer;

3. The method of claim 2, wherein the order model further comprises a feature association submodel;

4. The method of claim 3, wherein the feature association submodel comprises: a full-scale feature correlation model;

5. The method of claim 3, wherein the feature association submodel comprises: setting a characteristic correlation model;

6. The method according to claim 3, wherein obtaining the ranking result for each search result according to the first encoding result, the second encoding result corresponding to each search result, and the auxiliary correlation specifically includes:

7. The method of claim 1, wherein training the ranking model according to the ranking results and the label information corresponding to each search result specifically comprises:

8. A method of information presentation, comprising:

receiving a search sentence input by a user;

determining each search result corresponding to the search statement;

inputting each search result into a pre-trained sequencing model to obtain a sequencing result corresponding to the search statement, wherein the sequencing model is obtained by training through the method of any one of claims 1 to 7;

and displaying information to the user according to the sequencing result.

9. An apparatus for model training, comprising:

10. An apparatus for information presentation, comprising:

the receiving module is used for receiving a search statement input by a user;

a ranking module, configured to input each search result into a pre-trained ranking model to obtain a ranking result corresponding to the search statement, where the ranking model is obtained by training according to any one of the methods of claims 1 to 7;

11. A computer-readable storage medium, characterized in that the storage medium stores a computer program which, when executed by a processor, implements the method of any of the preceding claims 1 to 7 or 8.

12. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of any of claims 1 to 7 or 8 when executing the program.