Search recommendation model training method, search result sorting method and device

Abstract

The present application discloses a method for training a search recommendation model in the field of artificial intelligence, a method and device for sorting search results, the training method comprising: obtaining a training sample set, the training sample set comprising a sample user behavior group sequence and a sample masked user behavior group sequence; training the search recommendation model by taking the training sample set as input data and the object of the sample user's response operation obtained by the masking process as the training target, and obtaining a trained search recommendation model, the search recommendation model is used to predict the label of the candidate recommendation object in the search results corresponding to the query field when the target user enters a query field, and the label is used to indicate the probability of the target user performing a response operation on the candidate recommendation object. The technical solution based on the present application can improve the accuracy of the search recommendation model, thereby enhancing the user experience.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and more specifically, to a method for training a search recommendation model, and a method and device for sorting search results. Background Art

[0002] Selection rate prediction refers to predicting the probability of a user choosing a product in a specific environment. For example, in search recommendation systems of applications such as app stores and online advertising, selection rate prediction based on the search terms (also known as query terms or query fields) entered by users plays a key role. For example, objects in the candidate set corresponding to the query terms can be sorted based on selection rate prediction to generate search results; selection rate prediction can improve user experience.

[0003] At present, in the search recommendation system, a ranking function can be learned based on the association between the search term and the objects in the candidate set, and the objects in the candidate set can be ranked to generate search results. However, the above ranking function only takes into account the static information of the current search, that is, the information of the search term; for different users, the same search ranking result is obtained by inputting the same search term; that is, the accuracy of the search results obtained based on the current search recommendation model is low, which cannot meet the needs of different users and leads to a poor user experience.

[0004] Therefore, how to improve the accuracy of search recommendation models and thus enhance user experience has become an urgent problem to be solved. Summary of the invention

[0005] The present application provides a method for training a search recommendation model, a method and a device for sorting search results, which can improve the accuracy of the search recommendation model, thereby improving the accuracy of the feedback search results.

[0006] In a first aspect, a method for training a search recommendation model is provided, comprising: obtaining a training sample set, wherein the training sample set includes a sample user behavior group sequence and a sample masked user behavior group sequence, the sample user behavior group sequence includes a first query field and an object of a response operation of a sample user in a search result corresponding to the first query field, and the sample masked user behavior group sequence includes a second query field and a sequence obtained by masking the object of a response operation of the sample user in the search result corresponding to the second query field; training a search recommendation model by using the training sample set as input data and using the masked object of the sample user response operation as a training target to obtain a trained search recommendation model, wherein the search recommendation model is used to predict a label of a candidate recommendation object in the search result corresponding to the query field when a target user inputs a query field, and the label is used to indicate the probability that the target user responds to the candidate recommendation object.

[0007] It should be understood that the above sample user behavior group sequence may refer to a group of data of the sample users sorted in chronological order; for example, the behavior log data of the sample users.

[0008] In one possible implementation, the training sample set may be data obtained from a user behavior log of a sample user, specifically including the sample user's historical query fields, and the sample user's response operations to candidate recommendation objects in the search results corresponding to the historical query fields; for example, the sample's click operations, download operations, purchase operations, and browsing operations on candidate recommendations in the search results corresponding to a certain query field.

[0009] In a possible implementation, the candidate recommendation object may include any one of a document, a service product, an advertisement, and an application.

[0010] In one possible implementation, the label of the candidate recommendation object may refer to an evaluation score of the candidate recommendation object corresponding to the query field, wherein a candidate recommendation object with a higher evaluation score may indicate a greater probability that the user will respond to the candidate recommendation object based on the user's historical behavior data, and the candidate recommendation object may be sorted to a higher position in the search results.

[0011] In an embodiment of the present application, in the process of training the search recommendation model, data including user query fields and user response operations can be used so that the search recommendation model can learn the association between query fields and user response operations based on historical behavior logs of different users; that is, the search recommendation model in the present application can be effectively trained by utilizing user historical behavior data, so that the trained search recommendation model can predict the label of the candidate recommendation object in the search results corresponding to the query field when the user enters the query field; the user's search intention can be identified based on the label of the candidate recommendation object, thereby improving the accuracy of the feedback search results, that is, improving the accuracy of the search recommendation model.

[0012] In a possible implementation, the training sample set may include multiple sample data, and the multiple sample data may be historical behavior group sequences of the user obtained based on historical behavior data of the same sample user.

[0013] For example, multiple historical behavior group sequences of user A are obtained from the user behavior log of sample user A, and one of the multiple historical behavior group sequences includes a historical query field of sample user A, and an object of response operation of sample user A in the search results corresponding to the historical query field; through the multiple historical behavior group sequences of sample user A, the search recommendation model can improve the accuracy of the search recommendation model by learning the association relationship between the historical query field and the object of historical response operation in the multiple historical behavior group sequences of sample user A, thereby improving the accuracy of the search recommendation model when sample user A enters the current query field.

[0014] It should be understood that the sample data in the above training sample set can be data generated based on the operation behavior of real users; or, the sample data included in the training sample set can also be data constructed by an algorithm, that is, the sample data may not be generated based on the operation behavior of real users. In one possible implementation, the sample masking user behavior group sequence may refer to masking (MASK) the original user behavior group sequence obtained; for example, a certain proportion of user behavior group sequences can be randomly sampled from the original user behavior group sequence; for the sampled user behavior group sequence, the operation object of the user response behavior therein is masked, and the specific method of masking can be to replace the original object with a special 'MASK' symbol.

[0015] In combination with the first aspect, in certain implementations of the first aspect, the search recommendation model predicts the label of the candidate recommendation object based on the query field input by the target user and the historical behavior group sequence of the target user, and the historical behavior group sequence of the target user is obtained based on the historical query field of the target user and the historical behavior data corresponding to the historical query field, and the historical behavior data corresponding to the historical query field is the object of the target user's response operation to the search results corresponding to the historical query field.

[0016] In combination with the first aspect, in certain implementations of the first aspect, the sample user behavior group sequence also includes identification information, the identification information is used to indicate an association relationship between the first query field and an object of a response operation of the sample user, and the identification information includes a time identifier.

[0017] In one possible implementation, obtaining a training sample set includes: obtaining data of the first query field and the object of the sample user's response operation; and obtaining the sample user behavior group sequence by binding the first query field and the object of the sample user's response operation corresponding to the first query field.

[0018] In an embodiment of the present application, in order to learn the association between query fields and user response behaviors in the process of training the search recommendation model, that is, to learn the query fields and the user's response operations to the candidate recommendation objects in the search results corresponding to the query fields; the relatively independent query field data and user response field data in the user behavior log can be bound to obtain a training sample data, namely, a user behavior group sequence.

[0019] In a possible implementation manner, the binding process of the first query field and the sample user response operation corresponding to the first query field includes:

[0020] Add identification information to the first query field and the object of the response operation of the sample user corresponding to the first query field, wherein the identification information includes a time identifier.

[0021] In combination with the first aspect, in some implementations of the first aspect, the search recommendation model is a Bidirectional Encoding Representation Transformation BERT model, further comprising: vectorizing the sample user behavior group sequence and the sample covered user behavior group sequence to obtain a vector sequence;

[0022] The method uses the training sample set as input data, including: inputting the vector sequence into the BERT model.

[0023] In an embodiment of the present application, the search recommendation model can be a bidirectional encoding representation transform BERT model; further, in order to make the format of the training sample set recognizable by the search recommendation model, the data in the training sample set can be vectorized, that is, the multivariate group sequence can be converted into a vector sequence.

[0024] In combination with the first aspect, in some implementations of the first aspect, the response operation of the sample user includes one or more of a click operation, a download operation, a purchase operation, and a browse operation of the sample user.

[0025] In a second aspect, a method for sorting search results is provided, characterized by comprising: obtaining a sequence of user behavior groups to be processed of a user, wherein the sequence of user behavior groups to be processed includes a current query field of the user and a sequence obtained by masking an object of a response operation of the user;

[0026] Inputting the sequence of behavior groups to be processed into a pre-trained search recommendation model to obtain labels of candidate recommendation objects in the candidate recommendation object set corresponding to the current query field, wherein the labels are used to indicate the probability of the user performing a response operation on the candidate recommendation objects in the candidate recommendation object set;

[0027] The search ranking result corresponding to the current query field is obtained according to the label of the candidate recommendation object, wherein the search recommendation model is used to predict the label of the candidate recommendation object in the search results corresponding to the query field when the target user inputs the query field, and the search recommendation model is obtained by training with a training sample set as input data and obtaining the objects of sample user response operations that are masked as training targets, the training sample set includes a sample user behavior group sequence and a sample masked user behavior group sequence, the sample user behavior group sequence includes a first query field and the objects of the sample user's response operations in the search results corresponding to the first query field, and the sample masked user behavior group sequence includes a second query field and a sequence obtained after masking the objects of the sample user's response operations in the search results corresponding to the second query field.

[0028] It should be understood that the above sample user behavior group sequence may refer to a group of data of the sample users sorted in chronological order; for example, the behavior log data of the sample users.

[0029] In one possible implementation, the training sample set may be data obtained from a user behavior log of a sample user, specifically including the sample user's historical query fields, and the sample user's response operations to candidate recommendation objects in the search results corresponding to the historical query fields; for example, the sample's click operations, download operations, purchase operations, and browsing operations on candidate recommendations in the search results corresponding to a certain query field.

[0030] In a possible implementation, the candidate recommendation object may be a document, a service product, an advertisement, an application program, etc.

[0031] In one possible implementation, the label of the candidate recommendation object may refer to an evaluation score of the candidate recommendation object corresponding to the query field, wherein a candidate recommendation object with a higher evaluation score may indicate a greater probability that the user will respond to the candidate recommendation object based on the user's historical behavior data, and the candidate recommendation object may be sorted to a higher position in the search results.

[0032] In an embodiment of the present application, the label of the candidate recommendation object in the candidate recommendation object set corresponding to the current query field can be obtained through a pre-trained search recommendation model, and the label can be used to indicate the probability that the user responds to the candidate recommendation object in the candidate recommendation object set; the search ranking result corresponding to the current query field is obtained according to the label of the candidate recommendation object; in the process of training the search recommendation model, data including the user query field and the user's response operation can be adopted so that the search recommendation model can learn the association between the query field and the user's response operation according to the historical behavior logs of different users; that is, the search recommendation model in the present application can be effectively trained by using the user's historical behavior data, so that the trained search recommendation model can predict the label of the candidate recommendation object in the search results corresponding to the query field when the user enters the query field; the user's search intention can be identified according to the label of the candidate recommendation object, thereby improving the accuracy of the search result ranking.

[0033] In one possible implementation, masking the user behavior group sequence with a sample may refer to masking (MASK) the obtained original user behavior group sequence; for example, a certain proportion of user behavior group sequences may be randomly sampled from the original user behavior group sequence; for the sampled user behavior group sequence, the operation objects of the user response behaviors therein are masked, and the specific method of masking may be replacing the original objects with a special 'MASK' symbol.

[0034] In combination with the second aspect, in certain implementations of the second aspect, the search recommendation model predicts the label of the candidate recommendation object based on the query field input by the target user and the historical behavior group sequence of the target user, and the historical behavior group sequence of the target user is obtained based on the historical query field of the target user and the historical behavior data corresponding to the historical query field, and the historical behavior data corresponding to the historical query field is the object of the target user's response operation to the search results corresponding to the historical query field.

[0035] In a possible implementation manner, the sample user behavior group sequence is obtained by binding the acquired first query field of the sample user and the object of the response operation of the sample user.

[0036] In an embodiment of the present application, in order to learn the association between query fields and user response behaviors in the process of training the search recommendation model, that is, to learn the query fields and the user's response operations to the candidate recommendation objects in the search results corresponding to the query fields; the relatively independent query field data and user response field data in the user behavior log can be bound to obtain a training sample data, namely, a user behavior group sequence.

[0037] In a possible implementation, the sample user behavior group sequence is obtained by adding identification information to the acquired first query field of the sample user and the object of the response operation of the sample user, wherein the identification information includes a time identifier.

[0038] In combination with the second aspect, in certain implementations of the second aspect, the sample user behavior group sequence also includes identification information, the identification information is used to indicate an association relationship between the first query field and an object of a response operation of the sample user, and the identification information includes a time identifier.

[0039] In combination with the second aspect, in certain implementations of the second aspect, the pre-trained search recommendation model is a bidirectional encoding representation transform BERT model, and the training sample set refers to a set obtained by vectorizing the sample user behavior group sequence and the sample masked user behavior group sequence.

[0040] In an embodiment of the present application, the search recommendation model can be a bidirectional encoding representation transform BERT model; further, in order to make the format of the training sample set recognizable by the search recommendation model, the data in the training sample set can be vectorized, that is, the multivariate group sequence can be converted into a vector sequence.

[0041] In combination with the second aspect, in some implementations of the second aspect, the user's response operation includes one or more of the user's click operation, download operation, purchase operation, and browse operation.

[0042] In a third aspect, a training device for a search recommendation model is provided, comprising: an acquisition unit, configured to acquire a training sample set, wherein the training sample set comprises a sample user behavior group sequence and a sample masked user behavior group sequence, the sample user behavior group sequence comprises a first query field and an object of a response operation of a sample user in a search result corresponding to the first query field, and the sample masked user behavior group sequence comprises a second query field and a sequence obtained by masking the object of a response operation of the sample user in a search result corresponding to the second query field;

[0043] A processing unit is used to train a search recommendation model by using the training sample set as input data and the objects of the sample user response operations that have been masked as training targets, so as to obtain a trained search recommendation model, wherein the search recommendation model is used to predict the labels of candidate recommendation objects in the search results corresponding to the query field when the target user inputs a query field, and the labels are used to represent the probability that the target user will respond to the candidate recommendation object.

[0044] It should be understood that the expansion, limitation, explanation and description of the relevant contents in the above-mentioned first aspect also apply to the same contents in the third aspect.

[0045] In combination with the third aspect, in certain implementations of the third aspect, the search recommendation model predicts the label of the candidate recommendation object based on the query field input by the target user and the historical behavior group sequence of the target user, and the historical behavior group sequence of the target user is obtained based on the historical query field of the target user and the historical behavior data corresponding to the historical query field, and the historical behavior data corresponding to the historical query field is the object of the target user's response operation to the search results corresponding to the historical query field.

[0046] In combination with the third aspect, in certain implementations of the third aspect, the sample user behavior group sequence also includes identification information, the identification information is used to indicate an association relationship between the first query field and an object of a response operation of the sample user, and the identification information includes a time identifier.

[0047] In combination with the third aspect, in some implementations of the third aspect, the search recommendation model is a bidirectional encoding representation transform BERT model, and the processing unit is further used to: vectorize the sample user behavior group sequence and the sample covered user behavior group sequence to obtain a vector sequence;

[0048] The processing unit is specifically used to: input the vector sequence into the BERT model.

[0049] In combination with the third aspect, in certain implementations of the third aspect, the response operation of the sample user includes one or more of a click operation, a download operation, a purchase operation, and a browse operation of the sample user.

[0050] In a fourth aspect, a device for sorting search results is provided, comprising: an acquisition unit, for acquiring a sequence of user behavior groups to be processed, wherein the sequence of user behavior groups to be processed includes the current query field of the user and a sequence obtained by masking the object of the user's response operation; a processing unit, for inputting the sequence of behavior groups to be processed into a pre-trained search recommendation model to obtain a label of a recommended object in a set of candidate recommended objects corresponding to the current query field, wherein the label is used to indicate a probability that the user responds to a candidate recommended object in the set of candidate recommended objects; and obtaining a search sorting result corresponding to the query field according to the label of the candidate recommended object, wherein the search recommendation model is used to pre-train a search recommendation model when a target user enters a query field. Measure the labels of candidate recommended objects in the search results corresponding to the query field, the search recommendation model is obtained by taking a training sample set as input data and obtaining masked sample user response operation objects as training targets, the training sample set includes a sample user behavior group sequence and a sample masked user behavior group sequence, the training sample set includes a sample user behavior group sequence and a sample masked user behavior group sequence, the sample user behavior group sequence includes a first query field and the objects of the sample user's response operation in the search results corresponding to the first query field, the sample masked user behavior group sequence includes a second query field and a sequence obtained after masking the objects of the sample user's response operation in the search results corresponding to the second query field.

[0051] It should be understood that the expansion, limitation, explanation and description of the relevant contents in the above-mentioned second aspect also apply to the same contents in the fourth aspect.

[0052] In combination with the fourth aspect, in certain implementations of the fourth aspect, the search recommendation model predicts the label of the candidate recommendation object based on the query field input by the target user and the historical behavior group sequence of the target user, and the historical behavior group sequence of the target user is obtained based on the historical query field of the target user and the historical behavior data corresponding to the historical query field, and the historical behavior data corresponding to the historical query field is the object of the target user's response operation to the search results corresponding to the historical query field.

[0053] In combination with the fourth aspect, in certain implementations of the fourth aspect, the sample user behavior group sequence also includes identification information, the identification information is used to indicate an association relationship between the first query field and an object of a response operation of the sample user, and the identification information includes a time identifier.

[0054] In combination with the fourth aspect, in certain implementations of the fourth aspect, the pre-trained search recommendation model is a bidirectional encoding representation transform BERT model, and the training sample set refers to a set obtained by vectorizing the sample user behavior group sequence and the sample masked user behavior group sequence.

[0055] In combination with the fourth aspect, in certain implementations of the fourth aspect, the user's response operation includes one or more of the user's click operation, download operation, purchase operation, and browse operation.

[0056] In a fifth aspect, a training device for a search recommendation model is provided, comprising an input / output interface, a processor, and a memory. The processor is used to control the input / output interface to send and receive information, the memory is used to store a computer program, and the processor is used to call and run the computer program from the memory, so that the training device executes the training method in the first aspect and any one of the implementations of the first aspect.

[0057] Optionally, the training device may be a terminal device / server, or a chip in the terminal device / server.

[0058] Optionally, the memory may be located inside the processor, for example, it may be a cache memory in the processor. The memory may also be located outside the processor, so as to be independent of the processor; for example, it may be an internal memory of the training device.

[0059] In a sixth aspect, a device for sorting search results is provided, comprising an input / output interface, a processor, and a memory. The processor is used to control the input / output interface to send and receive information, the memory is used to store a computer program, and the processor is used to call and run the computer program from the memory, so that the device executes the method in the second aspect and any one of the implementations of the second aspect.

[0060] Optionally, the above-mentioned device may be a terminal device / server, or a chip in the terminal device / server.

[0061] Optionally, the memory may be located inside the processor, for example, a cache in the processor. The memory may also be located outside the processor, so as to be independent of the processor, for example, an internal memory of the device.

[0062] In a seventh aspect, a computer program product is provided, the computer program product comprising: a computer program code, when the computer program code is run on a computer, the computer executes the methods in the above aspects.

[0063] It should be noted that the above-mentioned computer program code can be stored in whole or in part on the first storage medium, wherein the first storage medium can be packaged together with the processor or separately packaged with the processor, and the embodiments of the present application do not specifically limit this.

[0064] In an eighth aspect, a computer-readable medium is provided, wherein the computer-readable medium stores a program code, and when the computer program code is executed on a computer, the computer executes the methods in the above aspects. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 is a schematic diagram of a search system provided in an embodiment of the present application;

[0066] Figure 2 It is a schematic diagram of a system architecture provided by an embodiment of the present application;

[0067] Figure 3 is a schematic diagram of the hardware structure of a chip provided in an embodiment of the present application;

[0068] Figure 4 It is a system architecture of a method for training a search recommendation model and a method for ranking search results using an embodiment of the present application;

[0069] Figure 5 is a schematic flow chart of a training method for a search recommendation model provided in an embodiment of the present application;

[0070] Figure 6 is a schematic diagram of a sequence search architecture based on a bidirectional encoding representation change provided in an embodiment of the present application;

[0071] Figure 7 is a schematic diagram of a user behavior group sequence provided in an embodiment of the present application;

[0072] Figure 8 is a schematic diagram of offline training of a search recommendation model provided in an embodiment of the present application;

[0073] Fig. 9 is a schematic flow chart of a method for sorting search results provided in an embodiment of the present application;

[0074] Fig.10 It is a schematic diagram of online prediction of a search recommendation model provided in an embodiment of the present application;

[0075] Fig.11 This is a schematic diagram of searching for recommended objects in an application market provided by an embodiment of the present application;

[0076] Fig.12 is a schematic block diagram of a training device for a search recommendation model provided in an embodiment of the present application;

[0077] Fig.13 is a schematic block diagram of a device for sorting search results provided in an embodiment of the present application;

[0078] Fig.14 is a schematic block diagram of a training device for a search recommendation model provided in an embodiment of the present application;

[0079] Fig.15 It is a schematic block diagram of a device for sorting search results provided in an embodiment of the present application. DETAILED DESCRIPTION

[0080] The following will describe the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0081] First, the concepts involved in the embodiments of the present application are briefly described.

[0082] 1. Personalized search system

[0083] A personalized search system is a system that uses machine learning algorithms to analyze users' historical data, predict new query requests, and provide personalized search results.

[0084] For example, when a user inputs search term A, the personalized search system can analyze the user's historical behavior data using a machine learning algorithm to predict the user's association with the products in the system and the input search term. Furthermore, the products are sorted in descending order according to the predicted association or a function based on the association, and the search system displays the products in different positions in order as search results for the user and provides feedback to the user.

[0085] 2. Offline training

[0086] Offline training refers to a module in a personalized recommendation system that iteratively updates the recommendation model parameters according to the user's historical data using a machine learning algorithm until the set requirements are met.

[0087] 3. Online inference

[0088] Online prediction refers to predicting the user's preference for the recommended product under the current search term based on the offline trained model and the characteristics of the user, product and context, and predicting the probability of the user selecting the recommended product.

[0089] 4. Sequence Search

[0090] Sequential search is proposed to meet the needs of most search scenarios. It assumes that multiple query terms of a user within a period of time may be interrelated. Contextual information can be used to infer the user's response to the search results when the user initiates a new round of search. For example, after entering the search system, the user searches within the system and performs various operations on the feedback search results, such as querying, browsing, clicking, and downloading. Personalized search results can be returned based on the prediction of the user's response to the search results.

[0091] For example, Figure 1 It is a schematic diagram of a search system provided in an embodiment of the present application.

[0092] like Figure 1 As shown, after the user enters the search system 100, the user can trigger a search request in the search system 100 (for example, the user can enter a query term, which can also be called a search term); the search system can input the query term and other relevant information into the prediction model, and then predict the correlation between the products in the search system and the query term. Furthermore, the products can be sorted in descending order according to the predicted correlation or a function based on the correlation, and the search system can display the products in different positions in order as search results for the user; then the user can perform user actions on the feedback search results, such as browsing, clicking to download, etc. At the same time, the user's actual behavior will be stored in the log as training data, and the parameters of the prediction model will be continuously updated through the offline training module to improve the prediction effect of the model.

[0093] For example, when a user opens the application market in a smart terminal and enters a query word in the search bar, the search system of the application market is triggered. The search system of the application market predicts the probability of a user downloading a given application (APP) based on the user's historical download records, user click records and other user responses, as well as the application's own characteristics, such as time, location and other environmental feature information. Based on the calculation results, the search system of the application market can display the search results, i.e., candidate APPs, in descending order according to the predicted probability values, thereby increasing the download probability of the candidate APPs.

[0094] Among them, the above-mentioned smart terminal can refer to a mobile or fixed terminal device; for example, the smart terminal can be a mobile phone, a tablet personal computer (TPC), a media player, a smart TV, a laptop computer (LC), a personal digital assistant (PDA), a personal computer (PC), etc., and the embodiments of the present application do not impose any limitations on this.

[0095] For example, an APP with a higher predicted user selection rate may be displayed at a front search result position, and an APP with a lower predicted user selection rate may be displayed at a back search result position.

[0096] Furthermore, the user's response to the search results will also be logged and the parameters of the prediction model will be trained and updated through the offline training module.

[0097] For example, in the online prediction stage, the model can construct an input sequence based on the current query terms, historical query terms, and download results, and directly output ranked search results.

[0098] The search recommendation model in the above-mentioned offline training and the online prediction model can be a neural network model. The following introduces the relevant terms and concepts of the neural network that may be involved in the embodiments of the present application.

[0099] 5. Neural Networks

[0100] A neural network can be composed of neural units. A neural unit can be a s The output of the operation unit with the intercept 1 as input can be:

[0101]

[0102] Where, s = 1, 2, ... n, n is a natural number greater than 1, W s For x s , b is the bias of the neural unit. f is the activation function of the neural unit, which is used to introduce nonlinear characteristics into the neural network to convert the input signal in the neural unit into the output signal. The output signal of the activation function can be used as the input of the next convolutional layer, and the activation function can be a sigmoid function. A neural network is a network formed by connecting multiple single neural units mentioned above, that is, the output of one neural unit can be the input of another neural unit. The input of each neural unit can be connected to the local receptive field of the previous layer to extract the features of the local receptive field. The local receptive field can be an area composed of several neural units.

[0103] 6. Deep Neural Networks

[0104] A deep neural network (DNN), also known as a multi-layer neural network, can be understood as a neural network with multiple hidden layers. According to the position of different layers, the neural network inside the DNN can be divided into three categories: input layer, hidden layer, and output layer. Generally speaking, the first layer is the input layer, the last layer is the output layer, and the layers in between are all hidden layers. The layers are fully connected, that is, any neuron in the i-th layer must be connected to any neuron in the i+1-th layer.

[0105] Although DNN looks complicated, the work of each layer is not complicated. In simple terms, it can be expressed as the following linear relationship: in, is the input vector, is the output vector, is the offset vector, W is the weight matrix (also called coefficient), and α() is the activation function. Each layer is just an input vector After such a simple operation, the output vector Since DNN has many layers, the coefficient W and the offset vector The number of these parameters is also relatively large. The definitions of these parameters in DNN are as follows: Take the coefficient W as an example: Assume that in a three-layer DNN, the linear coefficient from the 4th neuron in the second layer to the 2nd neuron in the third layer is defined as The superscript 3 represents the layer number of the coefficient W, while the subscripts correspond to the third layer index 2 of the output and the second layer index 4 of the input.

[0106] In summary, the coefficients from the kth neuron in the L-1th layer to the jth neuron in the Lth layer are defined as

[0107] It should be noted that the input layer does not have a W parameter. In a deep neural network, more hidden layers allow the network to better describe complex situations in the real world. Theoretically, the more parameters a model has, the higher its complexity and the greater its "capacity", which means it can complete more complex learning tasks. Training a deep neural network is the process of learning the weight matrix, and its ultimate goal is to obtain the weight matrix of all layers of the trained deep neural network (the weight matrix formed by many layers of vector W).

[0108] 7. Loss Function

[0109] In the process of training a deep neural network, because we hope that the output of the deep neural network is as close as possible to the value we really want to predict, we can compare the predicted value of the current network with the target value we really want, and then update the weight vector of each layer of the neural network according to the difference between the two (of course, there is usually an initialization process before the first update, that is, pre-configuring parameters for each layer in the deep neural network). For example, if the predicted value of the network is high, adjust the weight vector to make it predict a lower value, and keep adjusting until the deep neural network can predict the target value we really want or a value very close to the target value we really want. Therefore, it is necessary to pre-define "how to compare the difference between the predicted value and the target value", which is the loss function or objective function, which are important equations used to measure the difference between the predicted value and the target value. Among them, taking the loss function as an example, the higher the output value (loss) of the loss function, the greater the difference, so the training of the deep neural network becomes a process of minimizing this loss as much as possible.

[0110] 8. Back Propagation Algorithm

[0111] Neural networks can use the error back propagation (BP) algorithm to correct the size of the parameters in the initial neural network model during the training process, so that the reconstruction error loss of the neural network model becomes smaller and smaller. Specifically, the forward transmission of the input signal to the output will generate error loss, and the parameters in the initial neural network model are updated by back propagating the error loss information, so that the error loss converges. The back propagation algorithm is a back propagation movement dominated by error loss, which aims to obtain the optimal parameters of the neural network model, such as the weight matrix.

[0112] Figure 2 A system architecture 200 provided by an embodiment of the present application is shown.

[0113] exist Figure 2 In the embodiment of the present invention, the data collection device 260 is used to collect training data. For the training method of the search recommendation model of the present invention, the search recommendation model can be further trained by training samples, that is, the training data collected by the data collection device 260 can be training samples.

[0114] For example, in an embodiment of the present application, a training sample set may include a sample user behavior group sequence and a sample masked user behavior group sequence. The sample user behavior group sequence includes a sample query field (for example, a first query field) and the object of the sample user's response operation in the search results corresponding to the sample query field. The sample masked user behavior group sequence refers to a sequence obtained after masking the object of the sample user's response operation.

[0115] After collecting the training data, the data collection device 260 stores the training data in the database 230 , and the training device 220 obtains the target model / rule 201 through training based on the training data maintained in the database 230 .

[0116] The following describes how the training device 220 obtains the target model / rule 201 based on the training data. The training device 220 processes the input sample training set, and compares the label of the object of the sample user response operation that is masked out and the true value, until the difference between the label of the object of the sample user response operation that is masked out and the true value is less than a certain threshold, thereby completing the training of the target model / rule 201.

[0117] For example, in an embodiment of the present application, the training device 220 can train the search recommendation model based on the training sample set. For example, the training sample set can be used as input data to obtain the object of the masked sample user response operation as the training target to train the search recommendation model; and then the trained search recommendation model is obtained, that is, the trained search recommendation model can be the target model / rule 201.

[0118] The target model / rule 201 can be used to predict the labels of candidate recommended objects in the search results corresponding to the query field when the user enters the query field. The target model / rule 201 in the embodiment of the present application can specifically be a bidirectional encoder representations from transformers (BERT) model, etc.

[0119] It should be noted that, in actual applications, the training data maintained in the database 230 may not all be collected by the data collection device 260, but may also be received from other devices.

[0120] It should also be noted that the training device 220 does not necessarily train the target model / rule 201 entirely based on the training data maintained by the database 230, and it is also possible to obtain training data from the cloud or other places for model training. The above description should not be regarded as a limitation on the embodiments of the present application.

[0121] The target model / rule 201 trained by the training device 220 can be applied to different systems or devices, such as Figure 2The execution device 210 shown in the figure may be a terminal, such as a mobile phone terminal, a tablet computer, a laptop computer, an augmented reality (AR) / virtual reality (VR), a vehicle terminal, etc., or a server, or a cloud server, etc. Figure 2 In the embodiment, the execution device 210 is configured with an input / output (I / O) interface 212 for data interaction with an external device. A user may input data to the I / O interface 212 through a client device 240. The input data may include: training samples input by the client device in the embodiment of the present application.

[0122] Preprocessing module 213 and preprocessing module 214 are used to preprocess the input data received by I / O interface 212; in the embodiment of the present application, preprocessing module 213 and preprocessing module 214 may be omitted (or only one of the preprocessing modules may be present), and computing module 211 may be used directly to process the input data.

[0123] When the execution device 210 preprocesses the input data, or when the computing module 211 of the execution device 210 performs calculations and other related processing, the execution device 210 can call the data, code, etc. in the data storage system 250 for corresponding processing, and can also store the data, instructions, etc. obtained from the corresponding processing into the data storage system 250.

[0124] Finally, the I / O interface 212 processes the results. For example, the trained search recommendation model can be used by the search system to online predict the labels of the candidate recommendation objects in the candidate recommendation object set corresponding to the user's current query field, and obtain the search ranking results corresponding to the current query field based on the labels of the candidate recommendation objects and return them to the client device 240 for provision to the user.

[0125] It is worth noting that the training device 220 can generate corresponding target models / rules 201 based on different training data for different goals or different tasks. The corresponding target models / rules 201 can be used to achieve the above goals or complete the above tasks, thereby providing users with the desired results.

[0126] exist Figure 2 In the case shown in FIG. 2 , in one case, the user can manually provide input data, and the manual provision can be operated through the interface provided by the I / O interface 212 .

[0127] In another case, the client device 240 can automatically send input data to the I / O interface 212. If the client device 240 needs to obtain the user's authorization to automatically send the input data, the user can set the corresponding permissions in the client device 240. The user can view the results output by the execution device 210 in the client device 240, and the specific presentation form can be a specific method such as display, sound, action, etc. The client device 240 can also be used as a data acquisition terminal to collect the input data of the input I / O interface 212 and the output results of the output I / O interface 212 as shown in the figure as new sample data, and store them in the database 230. Of course, it is also possible not to collect through the client device 240, but the I / O interface 212 directly stores the input data of the input I / O interface 212 and the output results of the output I / O interface 212 as new sample data in the database 130.

[0128] It is worth noting that Figure 2 This is only a schematic diagram of a system architecture provided by an embodiment of the present application, and the positional relationship between the devices, components, modules, etc. shown in the figure does not constitute any limitation. Figure 2 In the embodiment, the data storage system 250 is an external memory relative to the execution device 210; in other cases, the data storage system 250 can also be placed in the execution device 210.

[0129] Exemplarily, the search recommendation model in the present application may be a bidirectional encoder representations from transformers (BERT) model; or, it may be an enhanced representation through knowledge integration (EREIN) model; or it may be other models.

[0130] Figure 3 It is a schematic diagram of the hardware structure of a chip provided in an embodiment of the present application.

[0131] like Figure 3 As shown, the chip includes a neural-network processing unit 300 (NPU); the chip can be set as follows Figure 2 The execution device 210 shown in FIG. 1 is used to complete the calculation work of the calculation module 211. The chip can also be set in Figure 2 The training device 220 shown is used to complete the training work of the training device 220 and output the target model / rule 201.

[0132] The neural network processor 300 is mounted on the main central processing unit (CPU) as a coprocessor, and the main CPU assigns tasks; the core part of the NPU 300 is the operation circuit 303, and the controller 304 controls the operation circuit 303 to extract data from the memory (weight memory or input memory) and perform operations.

[0133] In some implementations, the computing circuit 303 includes multiple processing engines (PEs) inside.

[0134] In some implementations, the arithmetic circuit 303 is a two-dimensional systolic array; the arithmetic circuit 203 may also be a one-dimensional systolic array or other electronic circuitry capable of performing mathematical operations such as multiplication and addition.

[0135] In some implementations, the operation circuit 303 is a general-purpose matrix processor.

[0136] For example, assume there is an input matrix A, a weight matrix B, and an output matrix C. The operation circuit 303 takes the corresponding data of the matrix B from the weight memory 302 and caches it on each PE in the operation circuit 303. The operation circuit 303 obtains the matrix A data from the input memory 301 and performs matrix operations with the matrix B. The partial result or the final result of the matrix is ​​stored in the accumulator 208 (accumulator). The vector calculation unit 307 can further process the output of the operation circuit 303, such as vector multiplication, vector addition, exponential operation, logarithmic operation, size comparison, etc.

[0137] For example, the vector calculation unit 307 can be used for network calculations of non-convolutional / non-FC layers in a neural network, such as pooling, batch normalization, local response normalization, etc.

[0138] In some implementations, the vector computation unit 307 can store the vector of processed outputs to the unified memory 306. For example, the vector computation unit 307 can apply a nonlinear function to the output of the operation circuit 303; for example, accumulate a vector of values ​​to generate an activation value.

[0139] In some implementations, the vector calculation unit 307 generates normalized values, merged values, or both.

[0140] In some implementations, the vector of processed outputs can be used as activation input to the computational circuit 303, for example for use in subsequent layers in a neural network.

[0141] Exemplarily, the unified memory 306 can be used to store input data and output data. The weight data is directly stored in the input data in the external memory to the input memory 301 and / or the unified memory 306 through the storage unit access controller 305 (direct memory access controller, DMAC), the weight data in the external memory is stored in the weight memory 302, and the data in the unified memory 306 is stored in the external memory.

[0142] Exemplarily, the bus interface unit 310 (BIU) may be used to implement interaction between the main CPU, the DMAC and the instruction fetch memory 309 via a bus.

[0143] Exemplarily, an instruction fetch buffer 309 connected to the controller 304 may be used to store instructions used by the controller 304. The controller 304 may be used to call instructions cached in the instruction fetch buffer 309 to control the working process of the computing accelerator.

[0144] Generally, the unified memory 306, the input memory 301, the weight memory 302 and the instruction fetch memory 309 can all be on-chip memories; the external memory is a memory outside the NPU, which can be a double data rate synchronous dynamic random access memory (DDR SDRAM), a high bandwidth memory (HBM) or other readable and writable memory.

[0145] It should be noted that the operations in the search recommendation model in the embodiment of the present application can be performed by the operation circuit 303 or the vector calculation unit 307.

[0146] At present, in the search recommendation system, a ranking function can be learned based on the association between the search term and the objects in the candidate set, and the objects in the candidate set can be ranked to generate search results. However, the above ranking function only takes into account the static information of the current search, that is, the information of the search term; for different users, the same search ranking result is obtained by inputting the same search term; that is, the accuracy of the search results obtained based on the current search recommendation model is low, which cannot meet the needs of different users and leads to a poor user experience.

[0147] In view of this, by using data including user query fields and user response operations in the process of training the search recommendation model, the search recommendation model can learn the association between query fields and user response operations based on the historical behavior logs of different users; that is, the search recommendation model in the present application can be effectively trained by utilizing user historical behavior data, so that the trained search recommendation model can predict the labels of candidate recommendation objects in the search results corresponding to the query field when the user enters the query field; the user's search intention can be identified based on the label of the candidate recommendation object, thereby improving the accuracy of the feedback search results, that is, improving the accuracy of the search recommendation model.

[0148] Figure 4 The system architecture of the method for training the search recommendation model and the method for ranking search results of the embodiment of the present application is applied. The system architecture 400 may include a local device 420, a local device 430, an execution device 410, and a data storage system 450, wherein the local device 420 and the local device 430 may be connected to the execution device 410 via a communication network.

[0149] Among them, the execution device 410 can be implemented by one or more servers. Optionally, the execution device 410 can be used in conjunction with other computing devices; for example: data storage, routers, load balancers and other devices. The execution device 410 can be arranged at a physical site, or distributed at multiple physical sites. The execution device 410 can use the data in the data storage system 450, or call the program code in the data storage system 450 to implement the training method of the search recommendation model and the method for sorting search results of the embodiment of the present application.

[0150] Exemplarily, the data storage system 450 may be deployed in the local device 420 or the local device 430; for example, the data storage system 450 may be used to store a user's behavior log.

[0151] It should be noted that the above-mentioned execution device 410 can also be called a cloud device, and in this case the execution device 410 can be deployed in the cloud.

[0152] Specifically, the execution device 410 can perform the following process: obtaining a training sample set, wherein the training sample set includes a sample user behavior group sequence and a sample masked user behavior group sequence, the sample user behavior group sequence includes a first query field and an object of a response operation of a sample user in a search result corresponding to the first query field, and the sample masked user behavior group sequence includes a second query field and a sequence obtained by masking the object of a response operation of the sample user in the search result corresponding to the second query field; training a search recommendation model by using the training sample set as input data and using the masked object of the sample user response operation as a training target to obtain a trained search recommendation model, wherein the search recommendation model is used to predict a label of a candidate recommended object in the search result corresponding to the query field when the target user inputs a query field, and the label is used to represent the probability that the target user responds to the candidate recommended object.

[0153] By executing the above process, the device 410 can obtain a pre-trained search recommendation model through training. The search recommendation model can be effectively trained by utilizing the user's historical behavior data, so that the trained search recommendation model can predict the labels of candidate recommendation objects in the search results corresponding to the query field according to the user's interests and hobbies when the user enters the query field.

[0154] In a possible implementation, the training method of the execution device 410 may be an offline training method executed in the cloud.

[0155] For example, after the user operates their respective user devices (e.g., local device 420 and local device 430), the operation log can be stored in the data storage system 450, and the execution device 410 can call the data in the data storage system 450 to complete the training process of the search recommendation model. Among them, each local device can represent any computing device, such as a personal computer, a computer workstation, a smart phone, a tablet computer, a smart camera, a smart car or other types of cellular phones, a media consumption device, a wearable device, a set-top box, a game console, etc. The local device of each user can interact with the execution device 410 through a communication network of any communication mechanism / communication standard, and the communication network can be a wide area network, a local area network, a point-to-point connection, etc., or any combination thereof.

[0156] In one implementation, the local device 420 and the local device 430 can obtain relevant parameters of a pre-trained search recommendation model from the execution device 410, and place the search recommendation model on the local device 420 and the local device 430. When a user enters a query field, the search recommendation model can be used to predict the labels of candidate recommendation objects in the search results corresponding to the query field. The labels can indicate whether the user responds to the candidate recommendation object; for example, whether the user will click on the candidate recommendation object.

[0157] In another implementation, a pre-trained search recommendation model can be directly deployed on the execution device 410. The execution device 410 obtains the sequence of user behavior groups to be processed from the local device 420 and the local device 430, and obtains the labels of the candidate recommendation objects in the set of candidate recommendation objects corresponding to the current query field according to the pre-trained search recommendation model.

[0158] Exemplarily, the data storage system 450 may be deployed in the local device 420 or the local device 430 to store the user behavior logs of the local device.

[0159] Exemplarily, the data storage system 450 may be independent of the local device 420 or the local device 430 and deployed separately on a storage device. The storage device may interact with the local device to obtain the user's behavior log in the local device and store it in the storage device.

[0160] Combine the following Figures 5 to 11 The embodiments of the present application are described in detail.

[0161] Figure 5 It is a schematic flowchart of the training method of the search recommendation model provided in an embodiment of the present application. Figure 5 The training method 500 shown includes steps 510 to 520, and steps 510 to 520 are described in detail below.

[0162] Step 510: Obtain a training sample set.

[0163] Among them, the above-mentioned training sample set includes a sample user behavior group sequence and a sample masked user behavior group sequence. The sample user behavior group sequence may include a first query field and the object of the sample user's response operation in the search results corresponding to the first query field. The sample masked user behavior group sequence refers to a sequence obtained after masking the object of the sample user's response operation; for example, the sample masked user behavior group sequence may include a second query field and a sequence obtained after masking the object of the sample user's response operation in the search results corresponding to the second query field.

[0164] It should be understood that the above sample behavior group sequence may refer to a group of data of the sample users sorted in chronological order; for example, the behavior log data of the sample users.

[0165] For example, the training sample set may be data acquired in the data storage system 450 as shown in the figure.

[0166] For example, Figure 1 As shown, the training sample set may be data obtained from the user behavior log of the sample user, specifically, including the historical query field of the sample user, and the response operation of the sample user to the candidate recommendation object in the search results corresponding to the historical query field; for example, the click operation, download operation, purchase operation and browsing operation of the sample to the candidate recommendation in the search results corresponding to a certain query field. It should be understood that the sample user behavior group sequence may include the first query field and the object of the sample user's response operation in the search results corresponding to the first query field; wherein the first query field and the object of the sample user's response operation in the search results are associated; for example, when the sample user enters a certain query field, the sample user will obtain the search results corresponding to the sample field, and the search results may include multiple candidate recommendation objects associated with the query field; further, the sample user may respond to the candidate recommendation objects included in the search results according to his own interests and needs, such as clicking, downloading, purchasing or browsing some or all of the candidate objects in the search results, thereby generating data of the sample user's response operation and storing it in the sample user's behavior log. Exemplarily, the object of the above-mentioned user response operation may refer to a document, a service product, an advertisement, an application, etc.; the user's operation response includes one or more of the user's click operation, download operation, purchase operation and browse operation.

[0167] Furthermore, in an embodiment of the present application, in order to learn the association between query fields and user response behaviors in the process of training the search recommendation model, that is, to learn the query fields and the user's response operations to the candidate recommendation objects in the search results corresponding to the query fields; the relatively independent query field data in the user behavior log can be bound to the user response field data, so as to obtain a training sample data, namely, a user behavior group sequence.

[0168] Optionally, in a possible implementation, the above-mentioned acquisition of the training sample set includes: acquiring data of the first query field and the object of the sample user's response operation; and obtaining a sample user behavior group sequence by binding the first query field and the object of the sample user's response operation corresponding to the first query field.

[0169] Exemplarily, the sample user behavior group sequence further includes identification information, where the identification information is used to indicate an association relationship between the first query field and an object of a response operation of the sample user, and the identification information includes a time identifier.

[0170] For example, Figure 7 As shown, it may be a sequence including 4 training samples, wherein each training sample may include a time identifier, a query field, and an object of a user's response operation (e.g., a clicked object, a downloaded object); wherein the time identifier may serve as an identifier for binding the query field with the object of the user's response operation corresponding to the query field.

[0171] Exemplarily, masking a sample user behavior group sequence may refer to masking (MASK) the obtained original user behavior group sequence; for example, a certain proportion of user behavior group sequences may be randomly sampled from the original user behavior group sequence; for the sampled user behavior group sequence, the operation objects of the user response behaviors therein are masked, and the specific method of masking may be replacing the original objects with a special 'MASK' symbol.

[0172] For example, Figure 8 As shown, the clicked object c3 and the downloaded object d3 in the user behavior group sequence with the time identifier t3 are covered to generate a corresponding covered user behavior group sequence.

[0173] Step 520: Train the search recommendation model by using the training sample set as input data and the objects of the sample user response operations subjected to masking as training targets, thereby obtaining a trained search recommendation model.

[0174] Among them, the above-mentioned search recommendation model is used to predict the label of the candidate recommended object in the search results corresponding to the query field when the target user enters the query field. The label can be used to indicate whether the target user responds to the candidate recommended object.

[0175] Exemplarily, the label of the above-mentioned candidate recommendation object may refer to the evaluation score of the candidate recommendation object corresponding to the query field, wherein a candidate recommendation object with a higher evaluation score may indicate a greater probability that the user will respond to the candidate recommendation object based on the user's historical behavior data, and the candidate recommendation object can be sorted to a higher position in the search results.

[0176] Optionally, in a possible implementation, the search recommendation model predicts the label of the candidate recommendation object based on the query field input by the target user and the historical behavior group sequence of the target user, and the historical behavior group sequence of the target user is obtained based on the historical query field of the target user and the historical behavior data corresponding to the historical query field, and the historical behavior data corresponding to the historical query field is the object of the target user's response operation to the search result corresponding to the historical query field.

[0177] In an embodiment of the present application, the search recommendation model can be effectively trained by utilizing user historical behavior data, so that the trained search recommendation model can predict the labels of candidate recommendation objects in the search results corresponding to the query field when the user enters the query field; the user's search intent can be identified based on the labels of the candidate recommendation objects, thereby improving the accuracy of search result ranking.

[0178] Exemplarily, in an embodiment of the present application, the search recommendation model may be a bidirectional encoding representation transform BERT model; further, in order to make the format of the training sample set recognizable to the search recommendation model, the data in the training sample set may be vectorized, that is, the multivariate group sequence may be converted into a vector sequence.

[0179] Optionally, in a possible implementation, the above-mentioned search recommendation model can be a bidirectional encoding representation transform BERT model, and the training method also includes: vectorizing the sample user behavior group sequence and the sample masked user behavior group sequence included in the training sample set to obtain a vector sequence; and taking the training sample set as input data, including: inputting the vector sequence into the BERT model.

[0180] For example, the query field and the object of the response operation in the user behavior group sequence can be converted into a dense vector; the query object is the query field, and the object of the response operation is the search result selected by the user. In the above vectorization process, the vector of the object itself needs to be added to its corresponding behavior group vector and behavior type vector to fully express the user behavior group sequence.

[0181] For example, taking the user behavior group sequence g1{time stamp:t1,query:q1,click:{c11,c12},download:{d11}} as an example, the user behavior group sequence can be formatted as follows:

[0182] Vector(q1)=word2vect(q1)+group_embedding(t1)+action_embedding(query);

[0183] Vector(c11)=item_embedding(c11)+group_embedding(t1)+action_embedding(click);

[0184] Vector(d11)=item_embedding(d11)+group_embedding(t1)+action_embedding(download);

[0185] Among them, word2vec() represents the word vector conversion function, group_embedding() represents the group embedding vector composition function, action_embedding() represents the behavior embedding vector composition function, and item_embedding() represents the item embedding vector composition function.

[0186] In an embodiment of the present application, in the process of training the search recommendation model, by adopting data including user query fields and user response operations, the search recommendation model can learn the association between query fields and user response operations based on historical behavior logs of different users; that is, the search recommendation model in the present application can be effectively trained by utilizing user historical behavior data, so that the trained search recommendation model can predict the label of the candidate recommendation object in the search results corresponding to the query field when the user enters the query field; the user's search intention can be identified according to the label of the candidate recommendation object, thereby improving the accuracy of the feedback search results, that is, improving the accuracy of the search recommendation model.

[0187] Figure 6 It is a schematic diagram of a sequence search architecture based on bidirectional coding representation changes provided in an embodiment of the present application.

[0188] like Figure 6 As shown, the sequence search architecture 600 includes a user historical behavior log 601, a user multi-behavior sequence alignment and behavior group sequence generation module 602, and a user behavior sequence modeling module based on bidirectional encoding representation transformation 603. Among them, the sequence search architecture can more accurately capture the user's interests by jointly modeling the user's historical query sequence and the user's response sequence to the query result, so as to output search results that are more in line with the user's personal preferences in the future.

[0189] Exemplarily, the user's historical query sequence and the user's historical response sequence can be obtained from the user behavior log 601. The user's historical query sequence refers to the sequence formed by sorting the query fields entered by the user in the search box in time; the user's historical response sequence refers to the sequence formed by sorting the user's response behavior to the search results displayed by the platform in time, for example, the user's response behavior may include but is not limited to: click, download, purchase and other actions that reflect the user's interest tendency.

[0190] The user multi-behavior sequence alignment and behavior group sequence generation module 602 is used to align the user historical query sequence with the user historical response sequence and output the aligned user behavior group sequence. The alignment process refers to obtaining the association relationship between the user historical query sequence and the user historical response sequence obtained from the user behavior log 601.

[0191] It should be understood that the user historical query sequence or user historical response sequence included in the user behavior log 601 can be relatively independent data, but the user historical response sequence refers to the user response to the feedback search results after the corresponding user enters a specific query word; therefore, there is an association relationship between the user historical query sequence and the user historical response sequence, and the association relationship between the user historical query sequence and the user historical response sequence in the user behavior log data can be obtained through the above-mentioned alignment processing.

[0192] For example, obtaining a user historical query sequence A from the user behavior log 601, then aligning the user historical query sequence with the user historical response sequence refers to obtaining a user historical response sequence A corresponding to the user historical query sequence A, that is, the user historical response sequence A refers to the user's response behavior to the search results of the corresponding query field A displayed by the platform after the user enters the query field A in the search box.

[0193] For example, Figure 7 Schematic diagram of a user behavior group sequence provided by an embodiment of the present application. Figure 7 As shown, each user behavior group sequence can be obtained by aligning the user's historical query sequence with the user's historical response sequence. Each element of the user behavior group sequence can be an N-tuple, and an N-tuple can describe a complete search behavior of the user, such as what query field the user used and what response the user had to the corresponding search result. For example, the user's response can include click, download, browse or purchase, etc. Figure 7 The object operated by the response behavior shown in may be a document, a service product, an advertisement, an application, etc.

[0194] In the embodiments of the present application, Figure 7As shown, a user query sequence and a user response sequence can be bound together through a time stamp to generate a training sample data.

[0195] Among them, the user behavior sequence modeling module 603 based on bidirectional coding representation transformation is used to receive the output of the user multi-behavior sequence alignment and behavior group sequence generation module 602, and obtains a search recommendation model by modeling the user behavior group sequence based on the bidirectional coding representation transformation. That is, when a user enters a query word, search results based on the user's preferences can be obtained based on the search recommendation model; that is, in the search results, recommended items that the user may be interested in and respond to (for example, click) can be arranged at the front of the search results based on the user's historical behavior, and recommended items that the user may not be interested in can be arranged at the back of the search results.

[0196] Combine the following Figures 8 to 9 The offline training and online prediction process of the above search recommendation model is described in detail.

[0197] Offline training phase:

[0198] Figure 8 It is a schematic flowchart of the offline training of the search recommendation model provided in an embodiment of the present application. Figure 8 The training method 700 shown includes steps S701 to S704, and steps S701 to S704 are described in detail below.

[0199] It should be noted that Figure 8 The offline training process of the search recommendation model shown can be Figure 7 The user behavior sequence modeling module 603 based on bidirectional coding representation transformation is shown.

[0200] S701: Perform masking processing on the user behavior group sequence.

[0201] Exemplarily, the obtained original user behavior group sequence can be masked (MASK); for example, a certain proportion of user behavior group sequences can be randomly sampled from the original user behavior group sequence; for the sampled user behavior group sequence, the operation objects of the user response behaviors therein are masked, and the specific method of masking can be to replace the original objects with special 'MASK' symbols.

[0202] It should be understood that the training goal of the search recommendation model is to restore the original objects that are covered in the above user behavior group sequence.

[0203] The original user behavior group sequence may refer to a user behavior group sequence obtained according to a user query sequence and a user historical response sequence; that is, Figure 7 The user behavior group sequence shown; or, it may refer to Figure 6 The user multi-behavior sequence alignment and user behavior group sequence output by the behavior group sequence generation module 602 are shown.

[0204] For example, Figure 8 As shown, after the original user behavior group sequence is masked, the user behavior response corresponding to the time mark t3 can be masked.

[0205] S702: Vectorization processing of user behavior group sequences.

[0206] The above-mentioned user behavior group sequence vectorization processing refers to formatting the user behavior group sequence after masking processing and converting it into a format that can be processed by a bidirectional encoder representations from transformers (BERT).

[0207] For example, the query behavior and response line sequence in the user behavior group sequence can be converted into dense vectors as the objects of operation; the operation object of the query behavior is the query field, and the operation object of the response behavior is the search result selected by the user. In the above vectorization process, the vector of the operation object itself needs to be added to its corresponding behavior group vector and behavior type vector to fully express this user behavior group sequence.

[0208] For example, taking the user behavior group sequence g1{time stamp:t1,query:q1,click:{c11,c12},download:{d11}} as an example, the user behavior group sequence can be formatted as follows:

[0209] Vector(q1)=word2vect(q1)+group_embedding(t1)+action_embedding(query);

[0210] Vector(c11)=item_embedding(c11)+group_embedding(t1)+action_embedding(click);

[0211] Vector(d11)=item_embedding(d11)+group_embedding(t1)+action_embedding(download);

[0212] Among them, word2vec() represents the word vector conversion function, group_embedding() represents the group embedding vector composition function, action_embedding() represents the behavior embedding vector composition function, and item_embedding() represents the item embedding vector composition function.

[0213] Through the above format processing of the user behavior group sequence, the user behavior group sequence g1 can be converted into four vectors: Vector(q1), Vector(c11), Vector(c12), and Vector(d11).

[0214] S703: Train a search recommendation model; for example, the search recommendation model may be a bidirectional encoding representation transformation model.

[0215] For example, by using the masked vector output by the above S702 as training data, the search recommendation model learns the association between sequence points of different user behavior groups and updates the vector expression of each sequence group; the training goal of the search recommendation model is to restore the original objects masked in the above user behavior group sequence.

[0216] Exemplarily, a vector sequence after vectorization processing of the user behavior group sequence is input into the search recommendation model. During the training process, the search recommendation model can generate random vectors for the vectors of the masked positions, and continuously reversely update the model parameters of the search recommendation model through the difference with the true value label of the object at the masked position, that is, continuously update the input vector sequence; since the vectors of the masked positions will interact with the vectors of other positions, the input vector sequence will be continuously updated during the training process; the training goal is to make the label of the vector of the masked position output by the search recommendation model continuously close to the true value label of the position; after the training of the search recommendation model is completed, the information of the user's historical behavior group sequence is saved in these vectors.

[0217] S704: Calculate the loss function of the search recommendation model.

[0218] For example, by calculating the loss value of self-recovery of the object at the masked position, it is determined that the search recommendation model is to be iteratively trained, and the model parameters of the search recommendation model can be trained based on the loss value.

[0219] Exemplarily, the vector expression corresponding to the masked position object output by the model can be obtained, and the training goal is to continuously update the model parameters of the search training model so that the inner product of the vector expression of the masked position and the embedding vector of the corresponding actual operation object is as large as possible, and the inner product with the embedding vectors of other objects is as small as possible, so that the search training model completes the training process.

[0220] In an embodiment of the present application, in the process of training the search recommendation model, by adopting data including user query fields and user response operations, the search recommendation model can learn the association between query fields and user response operations based on historical behavior logs of different users; that is, the search recommendation model in the present application can be effectively trained by utilizing user historical behavior data, so that the trained search recommendation model can predict the label of the candidate recommendation object in the search results corresponding to the query field when the user enters the query field; the user's search intention can be identified according to the label of the candidate recommendation object, thereby improving the accuracy of the feedback search results, that is, improving the accuracy of the search recommendation model.

[0221] Online prediction stage:

[0222] Fig. 9 It is a schematic diagram of online prediction of the search recommendation model provided in an embodiment of the present application.

[0223] like Fig. 9 As shown, the query field currently input by the user is obtained, and the user behavior operation object to be predicted is replaced with the 'MASK' symbol to generate masked input data; the input data is formatted to obtain an input vector expression; and the final search result, i.e., the search ranking result, is obtained by performing an inner product operation on the input vector and the embedding vector of the candidate set object.

[0224] Fig.10 It is an exemplary flowchart of the method for sorting search results provided in an embodiment of the present application.

[0225] Fig.10 The method 800 shown includes steps S810 to S830, and steps S810 to S830 are described in detail below.

[0226] S810: Obtain a sequence of user behavior groups to be processed.

[0227] The sequence of user behavior groups to be processed includes the user's current query field and a sequence obtained by masking the object of the user's response operation.

[0228] For example, Fig. 9 As shown, when the user enters the current query field, that is, the latest query field, since it is not yet known which recommended objects in the search results corresponding to the current query field the user will respond to, the object of the user's response operation can be replaced with the 'MASK' symbol to generate a sequence of user behavior groups to be processed.

[0229] S820: Input the to-be-processed behavior group sequence into a pre-trained search recommendation model to obtain labels of candidate recommendation objects in a set of candidate recommendation objects corresponding to the current query field.

[0230] The label may be used to indicate the probability of a user responding to a candidate recommendation object in the candidate recommendation object set.

[0231] S830: Obtain search ranking results corresponding to the current query field according to the labels of the candidate recommendation objects.

[0232] Among them, the search recommendation model is used to predict the label of the candidate recommended object in the search results corresponding to the query field when the target user inputs a query field. The search recommendation model is obtained by training with a training sample set as input data and with the objects of sample user response operations obtained through masking as training targets. The training sample set includes a sample user behavior group sequence and a sample masked user behavior group sequence. The sample user behavior group sequence includes a first query field and the objects of the sample user's response operations in the search results corresponding to the first query field. The sample masked user behavior group sequence includes a second query field and a sequence obtained after masking the objects of the sample user's response operations in the search results corresponding to the second query field.

[0233] Exemplarily, the candidate recommendation objects may be documents, service products, advertisements, applications, and the like.

[0234] Exemplarily, the user's response operation may include one or more of a click operation, a download operation, a purchase operation, and a browse operation on the candidate recommendation object.

[0235] It should be noted that through Figure 5 or Figure 8 The training method shown pre-trains the search model to obtain a trained search recommendation model; when the user enters the current query field, that is, when the user enters the latest query field, the obtained sequence of user behavior groups to be processed can be input into the pre-trained search recommendation model, and the pre-trained search recommendation model can output the user's label of the recommended object in the candidate recommendation set corresponding to the current field; according to the label of the candidate recommendation object, it can be predicted whether the user will respond to the candidate recommendation object in the candidate recommendation set, for example, it can be predicted whether the user will click on the candidate recommendation object; further, the candidate recommendation objects in the candidate recommendation set can be sorted according to the label of the candidate recommendation object to obtain a search sorting result.

[0236] Exemplarily, the label of the above-mentioned candidate recommendation object may refer to the evaluation score of the candidate recommendation object corresponding to the query field, wherein a candidate recommendation object with a higher evaluation score may indicate a greater probability that the user will respond to the candidate recommendation object based on the user's historical behavior data, and the candidate recommendation object can be sorted to a higher position in the search results.

[0237] It should be noted that the sample user behavior group sequence may include the first query field and the object of the sample user's response operation in the search results corresponding to the first query field; wherein the first query field and the object of the sample user's response operation in the search results are associated; for example, when the sample user enters a query field, the sample user will obtain the search results corresponding to the sample field, and the search results may include multiple candidate recommended objects associated with the query field; further, the sample user can respond to the candidate recommended objects included in the search results according to his own interests and needs, such as clicking, downloading, purchasing or browsing some or all of the candidate objects in the search results, thereby generating data on the sample user's response operation and storing it in the sample user's behavior log. In an embodiment of the present application, in order to enable the association between the query field and the user's response behavior to be learned in the process of training the search recommendation model, that is, to learn the query field and the user's response operation to the candidate recommended object in the search results corresponding to the query field; the relatively independent query field data in the user behavior log can be bound with the user response field data, thereby obtaining a training sample data, namely, a user behavior group sequence.

[0238] Optionally, in a possible implementation manner, the sample user behavior group sequence further includes identification information, where the identification information is used to indicate an association relationship between the first query field and an object of a response operation of the sample user, and the identification information includes a time identifier.

[0239] For example, Figure 7 As shown, it may be a sequence including 4 training samples, wherein each training sample may include a time identifier, a query field, and a user's response operation (e.g., click, download); wherein the time identifier may serve as an identifier for binding the query field with the user's response operation corresponding to the query field.

[0240] Optionally, in one possible implementation, the search recommendation model predicts the label of the candidate recommendation object based on the query field input by the target user and the historical behavior group sequence of the target user, and the historical behavior group sequence of the target user is obtained based on the historical query field of the target user and the historical behavior data corresponding to the historical query field, and the historical behavior data corresponding to the historical query field is the object of the target user's response operation to the search results corresponding to the historical query field.

[0241] In an embodiment of the present application, the search recommendation model can be effectively trained by utilizing user historical behavior data, so that the trained search recommendation model can predict the labels of candidate recommendation objects in the search results corresponding to the query field when the user enters the query field; the user's search intent can be identified based on the labels of the candidate recommendation objects, thereby improving the accuracy of search result ranking.

[0242] Optionally, in a possible implementation, the pre-trained search recommendation model is a Bidirectional Encoding Representation Transformation (BERT) model, and the training sample set refers to a set obtained by vectorizing the sample user behavior group sequence and the sample masked user behavior group sequence.

[0243] In an embodiment of the present application, the search recommendation model can be a bidirectional encoding representation transform BERT model; further, in order to make the format of the training sample set recognizable by the search recommendation model, the data in the training sample set can be vectorized, that is, the multivariate group sequence can be converted into a vector sequence.

[0244] In an embodiment of the present application, the label of the candidate recommendation object in the candidate recommendation object set corresponding to the current query field can be obtained through a pre-trained search recommendation model, and the label can be used to indicate the probability that the user responds to the candidate recommendation object in the candidate recommendation object set; the search ranking result corresponding to the current query field is obtained according to the label of the candidate recommendation object; in the process of training the search recommendation model, data including the user query field and the user's response operation can be adopted so that the search recommendation model can learn the association between the query field and the user's response operation according to the historical behavior logs of different users; that is, the search recommendation model in the present application can be effectively trained by using the user's historical behavior data, so that the trained search recommendation model can predict the label of the candidate recommendation object in the search results corresponding to the query field when the user enters the query field; the user's search intention can be identified according to the label of the candidate recommendation object, thereby improving the accuracy of the search result ranking.

[0245] Fig.11 It is a schematic diagram of searching for recommended objects in the application market provided in an embodiment of the present application.

[0246] like Fig.11As shown, a user opens an application market of a smart terminal and can enter the current search term "stand-alone game" in the search box (for example, a query field); after searching using the search term, search results can be returned; for example, the search results may include App1, App2, App3, and App4; the user can click App2 in the search results, and at this time, the data of the user's response operation can be stored in the user behavior log; next time, when the user enters the same search term in the search box of the application market, the new search ranking results displayed are as follows: Fig.11 As shown; App2" can be sorted in a front position, where "App2" can represent the same App as App2, or can represent an App of the same type as App2.

[0247] It should be noted that the process of the user entering the search word in the search box may be that the user enters the search word by voice; or, the user may manually enter the search word on the screen of the smart terminal, and this application does not impose any limitation on this.

[0248] For example, taking the search term "stand-alone game" as an example, if the user has clicked "Plants vs. Zombies 2", the search recommendation model provided by the embodiment of the present application will rank "Plants vs. Zombies All-Stars" higher in the search results when the user enters "stand-alone game" or a search term similar to "word game" next time; or, if the user has clicked "stand-alone Landlord (Happy Edition)", the search recommendation model provided by the embodiment of the present application will rank "June Game Stand-alone Landlord" higher in the search results when the user enters "stand-alone game" or a search term similar to "word game" next time; or, if the user has clicked "stand-alone Lianliankan", the search recommendation model provided by the embodiment of the present application will rank "Happy Candy Crush" higher in the search results when the user enters "stand-alone game" or a search term similar to "word game" next time. Similarly, if the category of the App clicked by the user changes this time, the search recommendation model in the present application can also capture the changes in the user's recent interests and hobbies, thereby obtaining dynamically adjusted search ranking results.

[0249] It should be understood that the above are examples, and the search results corresponding to the search terms returned by the application market are not limited to the above examples.

[0250] It should be understood that the above examples are intended to help those skilled in the art understand the embodiments of the present application, rather than to limit the embodiments of the present application to the specific numerical values ​​or specific scenarios illustrated. Those skilled in the art can obviously make various equivalent modifications or changes based on the above examples, and such modifications or changes also fall within the scope of the embodiments of the present application.

[0251] Combination of the above Figures 1 to 11 , describes in detail the training method of the search recommendation model and the method of sorting search results in the embodiment of the present application, and will now combine Figures 12 to 15 , describe in detail the device embodiments of the present application.

[0252] It should be understood that the training device in the embodiments of the present application can execute the training method of the search recommendation model in the aforementioned embodiments of the present application, and the search result sorting device can execute the search result sorting method in the aforementioned embodiments of the present application, that is, the specific working processes of the following various products can refer to the corresponding processes in the aforementioned method embodiments.

[0253] Fig.11 It is a schematic block diagram of a training device for a search recommendation model provided in an embodiment of the present application.

[0254] It should be understood that the training device 900 can perform Figure 5 The training method of the search recommendation model shown in FIG. 9 includes: an acquisition unit 910 and a processing unit 920 .

[0255] The acquisition unit 910 is used to acquire a training sample set, wherein the training sample set includes a sample user behavior group sequence and a sample masked user behavior group sequence, wherein the sample user behavior group sequence includes a first query field and an object of a response operation of a sample user in a search result corresponding to the first query field, and the sample masked user behavior group sequence includes a second query field and a sequence obtained by masking the object of a response operation of the sample user in a search result corresponding to the second query field; the processing unit 920 is used to train a search recommendation model by using the training sample set as input data and the masked object of the sample user response operation as a training target to obtain a trained search recommendation model, wherein the search recommendation model is used to predict the label of a candidate recommended object in the search result corresponding to the query field when the target user inputs a query field, and the label is used to indicate the probability of whether the target user responds to the candidate recommended object.

[0256] Optionally, as an embodiment, the search recommendation model predicts the label of the candidate recommendation object based on the query field input by the target user and the historical behavior group sequence of the target user, and the historical behavior group sequence of the target user is obtained based on the historical query field of the target user and the historical behavior data corresponding to the historical query field, and the historical behavior data corresponding to the historical query field is the object of the target user's response operation to the search results corresponding to the historical query field.

[0257] Optionally, as an embodiment, the sample user behavior group sequence further includes identification information, where the identification information is used to indicate an association relationship between the first query field and an object of a response operation of the sample user, and the identification information includes a time identifier.

[0258] Optionally, as an embodiment, the search recommendation model is a Bidirectional Encoding Representation Transformation (BERT) model, and the processing unit 920 is further used to: perform vectorization processing on the sample user behavior group sequence and the sample covered user behavior group sequence to obtain a vector sequence;

[0259] The processing unit 920 is specifically used to: input the vector sequence into the BERT model.

[0260] Optionally, as an embodiment, the response operation of the sample user includes one or more of a click operation, a download operation, a purchase operation and a browse operation of the sample user.

[0261] Fig.12 It is a schematic block diagram of a device for sorting search results provided in an embodiment of the present application.

[0262] It should be understood that the apparatus 1000 may perform Fig.10 The method for sorting search results shown in the figure. The device 1000 includes: an acquisition unit 1010 and a processing unit 1020.

[0263] Among them, the acquisition unit 1010 is used to obtain the user's sequence of user behavior groups to be processed, wherein the sequence of user behavior groups to be processed includes the user's current query field and a sequence obtained by masking the object of the user's response operation; the processing unit 1020 is used to input the sequence of behavior groups to be processed into a pre-trained search recommendation model to obtain the label of the recommended object in the candidate recommendation object set corresponding to the current query field, and the label is used to indicate the probability that the user responds to the candidate recommendation object in the candidate recommendation object set; the search ranking result corresponding to the query field is obtained according to the label of the candidate recommendation object, wherein the search recommendation model is used to predict the query field when the target user enters the query field. The search recommendation model is obtained by taking a training sample set as input data and obtaining masked sample user response operation objects as training targets, the training sample set includes a sample user behavior group sequence and a sample masked user behavior group sequence, the training sample set includes a sample user behavior group sequence and a sample masked user behavior group sequence, the sample user behavior group sequence includes a first query field and the object of the sample user's response operation in the search results corresponding to the first query field, the sample masked user behavior group sequence includes a second query field and a sequence obtained after masking the object of the sample user's response operation in the search results corresponding to the second query field.

[0264] Optionally, as an embodiment, the search recommendation model predicts the label of the candidate recommendation object based on the query field input by the target user and the historical behavior group sequence of the target user, and the historical behavior group sequence of the target user is obtained based on the historical query field of the target user and the historical behavior data corresponding to the historical query field, and the historical behavior data corresponding to the historical query field is the object of the target user's response operation to the search results corresponding to the historical query field.

[0265] Optionally, as an embodiment, the sample user behavior group sequence further includes identification information, where the identification information is used to indicate an association relationship between the first query field and an object of a response operation of the sample user, and the identification information includes a time identifier.

[0266] Optionally, as an embodiment, the pre-trained search recommendation model is a Bidirectional Encoding Representation Transformation BERT model, and the training sample set refers to a set obtained by vectorizing the sample user behavior group sequence and the sample masked user behavior group sequence.

[0267] Optionally, as an embodiment, the user's response operation includes one or more of the user's click operation, download operation, purchase operation and browse operation.

[0268] It should be noted that the training device 900 and the device 1000 are implemented in the form of functional units. The term "unit" here can be implemented in the form of software and / or hardware, and is not specifically limited to this.

[0269] For example, a "unit" may be a software program, a hardware circuit, or a combination of the two that implements the above functions. The hardware circuit may include an application specific integrated circuit (ASIC), an electronic circuit, a processor (e.g., a shared processor, a dedicated processor, or a group processor, etc.) and a memory for executing one or more software or firmware programs, a combined logic circuit, and / or other suitable components that support the described functions.

[0270] Therefore, the units of each example described in the embodiments of the present application can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present application.

[0271] Fig.14 It is a schematic diagram of the hardware structure of the training device for the search recommendation model provided in an embodiment of the present application.

[0272] Fig.14 The training device 1100 shown (the training device 1100 may be a computer device) includes a memory 1101, a processor 1102, a communication interface 1103 and a bus 1104. The memory 1101, the processor 1102 and the communication interface 1103 are connected to each other through the bus 1104.

[0273] The memory 1101 may be a read-only memory (ROM), a static storage device, a dynamic storage device or a random access memory (RAM). The memory 1101 may store a program. When the program stored in the memory 1101 is executed by the processor 1102, the processor 1102 is used to execute the various steps of the training method of the search recommendation model of the embodiment of the present application; for example, executing Figure 5 The steps shown.

[0274] It should be understood that the training device shown in the embodiment of the present application can be a server, for example, it can be a server in the cloud, or it can also be a chip configured in a server in the cloud.

[0275] Processor 1102 can adopt a general central processing unit (CPU), a microprocessor, an application specific integrated circuit (ASIC) or one or more integrated circuits to execute relevant programs to implement the training method of the search recommendation model of the method embodiment of the present application.

[0276] The processor 1102 may also be an integrated circuit chip with signal processing capability. In the implementation process, each step of the training method of the search recommendation model of the present application may be completed by an integrated logic circuit of hardware in the processor 1102 or by instructions in the form of software.

[0277] The processor 1102 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, or discrete hardware components. The methods, steps, and logic block diagrams disclosed in the embodiments of the present application may be implemented or executed. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc. The steps of the method disclosed in the embodiments of the present application may be directly embodied as being executed by a hardware decoding processor, or may be executed by a combination of hardware and software modules in the decoding processor. The software module may be located in a mature storage medium in the art, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, or the like. The storage medium is located in the memory 1101, and the processor 1102 reads the information in the memory 1101, and completes the implementation of the present application in combination with its hardware. Fig.12 The functions required to be performed by the units included in the training device shown, or the functions required to perform the method embodiments of the present application Figure 5 The training method of the search recommendation model shown.

[0278] The communication interface 1103 uses a transceiver device such as but not limited to a transceiver to implement communication between the training device 1100 and other devices or a communication network.

[0279] Bus 1104 may include a path for transmitting information between various components of training device 1100 (eg, memory 1101 , processor 1102 , communication interface 1103 ).

[0280] Fig.15 It is a schematic diagram of the hardware structure of the device for sorting search results provided in an embodiment of the present application.

[0281] Fig.15 The device 1200 shown (the device 1200 may be a computer device) includes a memory 1201, a processor 1202, a communication interface 1203 and a bus 1204. The memory 1201, the processor 1202 and the communication interface 1203 are connected to each other through the bus 1204.

[0282] The memory 1201 may be a read-only memory (ROM), a static storage device, a dynamic storage device or a random access memory (RAM). The memory 1201 may store a program. When the program stored in the memory 1201 is executed by the processor 1202, the processor 1202 is used to execute the various steps of the search result sorting method of the embodiment of the present application; for example, executing Fig.10 The steps shown.

[0283] It should be understood that the device shown in the embodiment of the present application may be a smart terminal, or may also be a chip configured in the smart terminal.

[0284] Processor 1202 can adopt a general central processing unit (CPU), a microprocessor, an application specific integrated circuit (ASIC) or one or more integrated circuits to execute relevant programs to implement the method of sorting search results in the method embodiment of the present application.

[0285] The processor 1202 may also be an integrated circuit chip with signal processing capability. In the implementation process, each step of the search result sorting method of the present application may be completed by hardware integrated logic circuits in the processor 1202 or software instructions.

[0286] The processor 1202 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, or discrete hardware components. The methods, steps, and logic block diagrams disclosed in the embodiments of the present application may be implemented or executed. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc. The steps of the method disclosed in the embodiments of the present application may be directly embodied as being executed by a hardware decoding processor, or may be executed by a combination of hardware and software modules in the decoding processor. The software module may be located in a mature storage medium in the art, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, or the like. The storage medium is located in the memory 1201, and the processor 1202 reads the information in the memory 1201, and completes the implementation of the present application in combination with its hardware. Fig.13 The functions required to be performed by the units included in the device shown, or the functions required to perform the method embodiments of the present application Fig.10 The method used to sort the search results shown.

[0287] The communication interface 1203 uses a transceiver device such as but not limited to a transceiver to implement communication between the apparatus 1200 and other devices or a communication network.

[0288] The bus 1204 may include a path for transmitting information between various components of the device 1200 (eg, the memory 1201 , the processor 1202 , and the communication interface 1203 ).

[0289] It should be noted that although the training apparatus 1100 and the apparatus 1200 shown above only show a memory, a processor, and a communication interface, in the specific implementation process, those skilled in the art should understand that the training apparatus 1100 and the apparatus 1200 may also include other devices necessary for normal operation. At the same time, those skilled in the art should understand that according to specific needs, the training apparatus 1100 and the apparatus 1200 may also include hardware devices for implementing other additional functions.

[0290] In addition, those skilled in the art should understand that the training device 1100 and the device 1200 may also only include the components necessary to implement the embodiments of the present application, and do not necessarily include Fig.14 or Fig.15 All devices shown in .

[0291] The embodiment of the present application also provides a chip, which includes a transceiver unit and a processing unit. The transceiver unit can be an input / output circuit or a communication interface; the processing unit is a processor or a microprocessor or an integrated circuit integrated on the chip; the chip can execute the training method of the search recommendation model in the above method embodiment.

[0292] The present application also provides a chip, which includes a transceiver unit and a processing unit. The transceiver unit may be an input / output circuit or a communication interface; the processing unit may be a processor or a microprocessor or an integrated circuit integrated on the chip; and the chip may execute the search result sorting method in the above method embodiment.

[0293] An embodiment of the present application also provides a computer-readable storage medium having instructions stored thereon, which, when executed, execute the training method of the search recommendation model in the above method embodiment.

[0294] An embodiment of the present application also provides a computer-readable storage medium having instructions stored thereon, which, when executed, execute the method for sorting search results in the above method embodiment.

[0295] An embodiment of the present application also provides a computer program product comprising instructions, which, when executed, execute the training method of the search recommendation model in the above method embodiment.

[0296] The embodiment of the present application also provides a computer program product comprising instructions, which, when executed, perform the method for sorting search results in the above method embodiment.

[0297] It should be understood that the processor in the embodiment of the present application may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0298] It should also be understood that the memory in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic random access memory (DRAM), synchronous DRAM (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM).

[0299] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

[0300] It should be understood that the term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. A and B can be singular or plural. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship, but it may also indicate an "and / or" relationship. Please refer to the context for specific understanding.

[0301] In this application, "at least one" means one or more, and "plurality" means two or more. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can be represented by: a, b, c, ab, ac, bc, or abc, where a, b, c can be single or multiple.

[0302] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0303] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0304] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0305] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0306] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0307] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0308] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0309] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A training method for a search recommendation model, It is characterized in that include: Acquire a training sample set, wherein the training sample set includes a sample user behavior group sequence and a sample masked user behavior group sequence, the sample user behavior group sequence includes a first query field and an object of a response operation of a sample user in a search result corresponding to the first query field, and the sample masked user behavior group sequence includes a second query field and a sequence obtained by masking the object of a response operation of the sample user in a search result corresponding to the second query field; The search recommendation model is trained by taking the training sample set as input data and the masked object of the sample user response operation as the training target, thereby obtaining a trained search recommendation model, wherein the search recommendation model is used to predict the label of the candidate recommendation object in the search results corresponding to the query field when the target user inputs a query field, and the label is used to represent the probability of the target user responding to the candidate recommendation object.

2. The training method according to claim 1, It is characterized in that The search recommendation model predicts the label of the candidate recommendation object based on the query field input by the target user and the historical behavior group sequence of the target user. The historical behavior group sequence of the target user is obtained based on the historical query field of the target user and the historical behavior data corresponding to the historical query field. The historical behavior data corresponding to the historical query field is the object of the target user's response operation to the search result corresponding to the historical query field.

3. The training method according to claim 1 or 2, It is characterized in that The sample user behavior group sequence further includes identification information, where the identification information is used to indicate an association relationship between the first query field and an object of a response operation of the sample user, and the identification information includes a time identifier.

4. The training method according to any one of claims 1 to 3, It is characterized in that The search recommendation model is a bidirectional encoding representation transformation BERT model, and further includes: Performing vectorization processing on the sample user behavior group sequence and the sample covered user behavior group sequence to obtain a vector sequence; The method uses the training sample set as input data, including: The vector sequence is input into the BERT model.

5. The training method according to any one of claims 1 to 4, It is characterized in that The response operation of the sample user includes one or more of a click operation, a download operation, a purchase operation, and a browse operation of the sample user.

6. A method for sorting search results, It is characterized in that include: Acquire a user behavior group sequence to be processed of the user, wherein the user behavior group sequence to be processed includes the current query field of the user and a sequence obtained by masking the object of the user's response operation; Inputting the sequence of behavior groups to be processed into a pre-trained search recommendation model to obtain labels of candidate recommendation objects in the candidate recommendation object set corresponding to the current query field, wherein the labels are used to indicate the probability of the user performing a response operation on the candidate recommendation objects in the candidate recommendation object set; Obtain the search ranking result corresponding to the current query field according to the label of the candidate recommendation object, Among them, the search recommendation model is used to predict the label of the candidate recommended object in the search results corresponding to the query field when the target user inputs a query field. The search recommendation model is obtained by training with a training sample set as input data and obtaining the masked sample user response operation objects as training targets. The training sample set includes a sample user behavior group sequence and a sample masked user behavior group sequence. The sample user behavior group sequence includes a first query field and the object of the sample user's response operation in the search results corresponding to the first query field. The sample masked user behavior group sequence includes a second query field and a sequence obtained after masking the object of the sample user's response operation in the search results corresponding to the second query field.

7. The method according to claim 6, It is characterized in that The search recommendation model predicts the label of the candidate recommendation object based on the query field input by the target user and the historical behavior group sequence of the target user. The historical behavior group sequence of the target user is obtained based on the historical query field of the target user and the historical behavior data corresponding to the historical query field. The historical behavior data corresponding to the historical query field is the object of the target user's response operation to the search result corresponding to the historical query field.

8. The method according to claim 6 or 7, It is characterized in that The sample user behavior group sequence further includes identification information, where the identification information is used to indicate an association relationship between the first query field and an object of a response operation of the sample user, and the identification information includes a time identifier.

9. The method according to any one of claims 6 to 8, It is characterized in that The pre-trained search recommendation model is a Bidirectional Encoding Representation Transformation BERT model, and the training sample set refers to a set obtained by vectorizing the sample user behavior group sequence and the sample masked user behavior group sequence.

10. The method according to any one of claims 6 to 9, It is characterized in that The user's response operation includes one or more of the user's click operation, download operation, purchase operation and browse operation.

11. A training device for a search recommendation model, It is characterized in that include: an acquisition unit, configured to acquire a training sample set, wherein the training sample set includes a sample user behavior group sequence and a sample masked user behavior group sequence, the sample user behavior group sequence includes a first query field and an object of a response operation of a sample user in a search result corresponding to the first query field, and the sample masked user behavior group sequence includes a second query field and a sequence obtained by masking the object of a response operation of the sample user in a search result corresponding to the second query field; A processing unit is used to train a search recommendation model by using the training sample set as input data and the objects of the sample user response operations that have been masked as training targets, so as to obtain a trained search recommendation model, wherein the search recommendation model is used to predict the labels of candidate recommendation objects in the search results corresponding to the query field when the target user inputs a query field, and the labels are used to represent the probability that the target user will respond to the candidate recommendation object.

12. The training device according to claim 11, It is characterized in that The search recommendation model predicts the label of the candidate recommendation object based on the query field input by the target user and the historical behavior group sequence of the target user. The historical behavior group sequence of the target user is obtained based on the historical query field of the target user and the historical behavior data corresponding to the historical query field. The historical behavior data corresponding to the historical query field is the object of the target user's response operation to the search result corresponding to the historical query field.

13. A training device according to claim 11 or 12, It is characterized in that The sample user behavior group sequence further includes identification information, where the identification information is used to indicate an association relationship between the first query field and an object of a response operation of the sample user, and the identification information includes a time identifier.

14. A training device as claimed in any one of claims 11 to 13, It is characterized in that The search recommendation model is a Bidirectional Encoding Representation Transformation BERT model, and the processing unit is further used for: Performing vectorization processing on the sample user behavior group sequence and the sample covered user behavior group sequence to obtain a vector sequence; The processing unit is specifically used for: The vector sequence is input into the BERT model.

15. A training device as claimed in any one of claims 11 to 14, It is characterized in that The response operation of the sample user includes one or more of a click operation, a download operation, a purchase operation, and a browse operation of the sample user.

16. A device for sorting search results, It is characterized in that include: An acquisition unit, configured to acquire a sequence of user behavior groups to be processed of a user, wherein the sequence of user behavior groups to be processed includes a current query field of the user and a sequence obtained by masking an object of a response operation of the user; A processing unit is used to input the sequence of behavior groups to be processed into a pre-trained search recommendation model, obtain the label of the recommended object in the candidate recommended object set corresponding to the current query field, and the label is used to indicate the probability that the user responds to the candidate recommended object in the candidate recommended object set; obtain the search ranking result corresponding to the query field according to the label of the candidate recommended object, Among them, the search recommendation model is used to predict the label of the candidate recommended object in the search results corresponding to the query field when the target user inputs a query field. The search recommendation model is obtained by training with a training sample set as input data and obtaining the masked sample user response operation objects as training targets. The training sample set includes a sample user behavior group sequence and a sample masked user behavior group sequence. The sample user behavior group sequence includes a first query field and the object of the sample user's response operation in the search results corresponding to the first query field. The sample masked user behavior group sequence includes a second query field and a sequence obtained after masking the object of the sample user's response operation in the search results corresponding to the second query field.

17. The device according to claim 16, It is characterized in that The search recommendation model predicts the label of the candidate recommendation object based on the query field input by the target user and the historical behavior group sequence of the target user. The historical behavior group sequence of the target user is obtained based on the historical query field of the target user and the historical behavior data corresponding to the historical query field. The historical behavior data corresponding to the historical query field is the object of the target user's response operation to the search result corresponding to the historical query field.

18. The device according to claim 16 or 17, It is characterized in that The sample user behavior group sequence further includes identification information, where the identification information is used to indicate an association relationship between the first query field and an object of a response operation of the sample user, and the identification information includes a time identifier.

19. The device according to any one of claims 16 to 18, It is characterized in that The pre-trained search recommendation model is a Bidirectional Encoding Representation Transformation BERT model, and the training sample set refers to a set obtained by vectorizing the sample user behavior group sequence and the sample masked user behavior group sequence.

20. The device according to any one of claims 16 to 19, It is characterized in that The user's response operation includes one or more of the user's click operation, download operation, purchase operation and browse operation.

21. A training device for a search recommendation model, It is characterized in that The system comprises at least one processor and a memory, wherein the at least one processor is coupled to the memory and is configured to read and execute instructions in the memory to perform the training method according to any one of claims 1 to 5.

22. A device for sorting search results, It is characterized in that The system comprises at least one processor and a memory, wherein the at least one processor is coupled to the memory and configured to read and execute instructions in the memory to perform the method according to any one of claims 6 to 10.

23. A computer readable medium, It is characterized in that The computer-readable medium stores a program code, and when the computer program code is executed on a computer, the computer is enabled to perform the training method according to any one of claims 1 to 5.

24. A computer readable medium, It is characterized in that The computer readable medium stores a program code, and when the computer program code is executed on a computer, the computer is caused to execute the method according to any one of claims 6 to 10.

Images