Search text processing method and device

Abstract

The invention provides a search text processing method and device. The method comprises the steps of obtaining intention features of an input text sent by a target client and at least one candidate search text matched with the input text; for any candidate search text, acquiring generalization features of any candidate search text; obtaining a combined text and a first statistical feature of the combined text; determining a target score of any candidate search text according to a user feature associated with the target client, an intention feature of the input text, a generalization feature and a first statistical feature; and determining at least one target search text from the at least one candidate search text according to the target score of any candidate search text, and sending the at least one target search text to the target client, thereby avoiding recommendation and display only according to the statistical characteristics of the candidate search text, relieving the Mattai effect of text drop-down, and improving the user experience. And the search experience of the user is improved.

Description

Technical Field

[0001] The present disclosure relates to the field of data processing technology, and in particular to a method and device for processing search text. Background Art

[0002] In the traffic-driving drop-down scenario, the user enters text in the search box, and the system first finds some matching texts in the vocabulary based on the input text, and then scores these texts using the model. These words are sorted by score and displayed to the user as drop-down text.

[0003] In the related art, a deep neural network model (DNN) is used as a scoring model. However, the scoring model is only displayed based on the statistical characteristics of the text. For example, it tends to recommend and display texts that are ranked high and are frequently displayed and clicked. It is difficult to recommend and display texts that are ranked low and some newly added texts, causing the Matthew effect to continue to intensify, affecting the user's search experience. Summary of the invention

[0004] The present disclosure aims to solve one of the technical problems in the related art at least to some extent.

[0005] To this end, the first purpose of the present disclosure is to propose a method for processing search text, so as to jointly determine the target score of any candidate search text based on user characteristics, intention characteristics of the input text, generalized characteristics of any candidate search text and the first statistical characteristics of the combined text corresponding to any candidate search text, thereby determining at least one target search text from at least one candidate search text based on the target score, thereby exploring more potential high-quality search texts, avoiding recommendation and display based solely on the statistical characteristics of the candidate search text, alleviating the Matthew effect of the drop-down text, and improving the user's search experience.

[0006] The second objective of the present disclosure is to provide a processing device for searching text.

[0007] The third objective of the present disclosure is to provide an electronic device.

[0008] A fourth objective of the present disclosure is to provide a computer-readable storage medium.

[0009] A fifth object of the present disclosure is to provide a computer program product.

[0010] To achieve the above-mentioned purpose, an embodiment of the first aspect of the present disclosure proposes a method for processing search text, including: receiving input text sent by a target client, and obtaining the intent feature of the input text, and obtaining at least one candidate search text matching the input text; for any candidate search text in the at least one candidate search text, obtaining the generalized feature of the any candidate search text, wherein the generalized feature is used to indicate the category and / or source to which the any candidate search text belongs; obtaining a combined text and a first statistical feature of the combined text, wherein the combined text is obtained by combining the any candidate search text and the input text; determining a target score of any candidate search text according to user features associated with the target client, the intent feature of the input text, the generalized feature of any candidate search text, and the first statistical feature of the combined text corresponding to the any candidate search text; determining at least one target search text from the at least one candidate search text according to the target score of the any candidate search text, and sending the at least one target search text to the target client.

[0011] The search text processing method of the embodiment of the present disclosure receives input text sent by a target client, obtains the intention feature of the input text, and obtains at least one candidate search text matching the input text; for any candidate search text among the at least one candidate search text, obtains the generalized feature of the any candidate search text, wherein the generalized feature is used to indicate the category and / or source to which the any candidate search text belongs; obtains a combined text and a first statistical feature of the combined text, wherein the combined text is obtained by combining the any candidate search text and the input text; and obtains a first statistical feature of the combined text according to user features associated with the target client, the intention feature of the input text, the generalized feature of the any candidate search text, and the first statistical feature of the combined text corresponding to the any candidate search text. A statistical feature is used to determine the target score of any candidate search text; at least one target search text is determined from the at least one candidate search text according to the target score of any candidate search text, and the at least one target search text is sent to the target client, thereby jointly determining the target score of any candidate search text according to user characteristics, intention characteristics of the input text, generalization characteristics of any candidate search text and the first statistical feature of the combined text corresponding to any candidate search text, and then determining at least one target search text from at least one candidate search text according to the target score, thereby realizing the exploration and discovery of more potential high-quality search texts, avoiding recommendation and display based only on the statistical characteristics of the candidate search text, alleviating the Matthew effect of the drop-down text, and improving the user's search experience.

[0012] To achieve the above-mentioned purpose, the second aspect of the present disclosure proposes a search text processing device, including: a receiving module, used to receive an input text sent by a target client, and obtain the intention feature of the input text, and obtain at least one candidate search text matching the input text; a first acquisition module, used to obtain the generalization feature of any candidate search text in the at least one candidate search text, wherein the generalization feature is used to indicate the category and / or source to which the any candidate search text belongs; a second acquisition module, used to obtain a combined text and a first statistical feature of the combined text, wherein the combined text is obtained by combining the any candidate search text and the input text; a first determination module, used to determine the target score of any candidate search text according to the user feature associated with the target client, the intention feature of the input text, the generalization feature of any candidate search text and the first statistical feature of the combined text corresponding to the any candidate search text; a second determination module, used to determine at least one target search text from the at least one candidate search text according to the target score of the any candidate search text, and send the at least one target search text to the target client.

[0013] To achieve the above-mentioned purpose, the third aspect embodiment of the present disclosure proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the search text processing method described in the first aspect embodiment of the present disclosure is implemented.

[0014] To achieve the above-mentioned purpose, the fourth aspect of the present disclosure proposes a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method for processing search text as described in the first aspect of the present disclosure is implemented.

[0015] To achieve the above-mentioned purpose, the fifth aspect of the present disclosure proposes a computer program product, which, when an instruction processor in the computer program product executes, implements the search text processing method as described in the first aspect of the present disclosure.

[0016] Additional aspects and advantages of the present disclosure will be given in part in the following description and in part will be obvious from the following description or learned through practice of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The above and / or additional aspects and advantages of the present disclosure will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0018] Figure 1 A flowchart of a text search processing method provided by an embodiment of the present disclosure;

[0019] Figure 2 A flowchart of a text search processing method provided by an embodiment of the present disclosure;

[0020] Figure 3 A flowchart of a text search processing method provided by an embodiment of the present disclosure;

[0021] Figure 4 A schematic diagram of the structure of the prediction model provided by the embodiment of the present disclosure;

[0022] Figure 5 A schematic diagram of the structure of a text search processing device provided by an embodiment of the present disclosure;

[0023] Figure 6 The invention is a block diagram of an electronic device for searching text according to an exemplary embodiment. DETAILED DESCRIPTION

[0024] Embodiments of the present disclosure are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present disclosure, and should not be construed as limiting the present disclosure.

[0025] It should be noted that the collection, collection, updating, analysis, processing, use, transmission, storage and other aspects of user personal information involved in the technical solution of this disclosure are in compliance with the provisions of relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. Necessary measures are taken for user personal information to prevent illegal access to user personal information data and maintain the security of user personal information, network security and national security.

[0026] In the related technology, an exploration model or a sorting model is used to recommend and display the drop-down text, wherein the exploration model may include: Thompson Sampling (TS), Upper Confidence Bound (UCB), Linear Thompson Sampling (LinearTS), and the sorting model may include: Logistic Regression (LR), Gradient Boosting Decision Tree (GBDT), Deep Neural Network (DNN).

[0027] However, the above TS and UCB methods only estimate the distribution of the global click rate of words, and provide the same drop-down words to different users, and cannot provide personalized results. Although the Linear TS method can provide personalized results to users, its architecture is relatively complex.

[0028] LR is essentially a linear classifier that is prone to underfitting and low precision. GBDT is a serial process with high computational complexity. Although the DNN model can achieve high precision, it focuses on the short-term benefits of words and lacks the ability to explore long-tail words with missing data.

[0029] Therefore, in view of the above problems, the present disclosure proposes a method and device for processing search text. It should be noted that the method for processing search text in the embodiment of the present disclosure can be applied to a server.

[0030] The following describes the search text processing method and device according to the embodiments of the present disclosure with reference to the accompanying drawings.

[0031] Figure 1 A flowchart of a text search processing method provided in an embodiment of the present disclosure.

[0032] like Figure 1 As shown, the method for processing the search text may include the following steps:

[0033] Step 101: receive input text sent by a target client, obtain intention features of the input text, and obtain at least one candidate search text matching the input text.

[0034] In an embodiment of the present disclosure, when a user enters text in a search box of a target client, the target client may send the input text entered by the user to a server. After receiving the input text, the server may obtain at least one candidate search text similar to the input text. For example, the server may obtain at least one candidate search text that is semantically similar to the input text, or may obtain at least one candidate search text that is similar to each text fragment in the input text.

[0035] As an example, the intent features of the input text can be used to indicate the completeness, part of speech, length, and breadth of the input text. In the embodiments of the present disclosure, the input text can be analyzed to determine the completeness, part of speech, length, and breadth of the input text.

[0036] Among them, it should be noted that the higher the completeness of the input text, the more specific the intention of the input text; the more precise the part of speech of the input text, the more specific the intention of the input text; the longer the length of the input text, the more specific the intention of the input text; the number of search texts recalled, exposed or clicked by the input text can be counted, and the more search texts recalled, exposed or clicked by the input text, the broader the input text is and the less specific the intention of the input text is.

[0037] Step 102: for any candidate search text among at least one candidate search text, obtain a generalized feature of any candidate search text.

[0038] The generalized features are used to indicate the category and / or source associated with any candidate search text.

[0039] In order to explore the uncertainty of each candidate search text, in an embodiment of the present disclosure, for any candidate search text, a generalized feature of any candidate search text can be obtained, wherein the generalized feature can indicate the category and / or source to which the corresponding candidate search text belongs. For example, taking the method applied to an online sales platform or an online shopping platform as an example, the first-level category of the online sales platform or the online shopping platform is food, and the second-level category can be fruits, snacks or vegetables, etc. The category associated with any candidate search text may include the first-level category and at least one second-level category corresponding to the first-level category; in addition, the source to which the candidate text belongs may be the brand associated with the candidate text. For example, taking the candidate text as a commodity as an example, the source associated with the candidate text may be a brand related to the commodity.

[0040] Step 103: Obtain the combined text and the first statistical feature of the combined text.

[0041] The combined text is obtained by combining any candidate search text and the input text.

[0042] Furthermore, a combined text obtained by combining any candidate search text and the input text may be obtained, and the first statistical feature is used to indicate statistical information such as the number of clicks and exposure of the combined text.

[0043] Step 104, determining a target score for any candidate search text based on user features associated with the target client, intention features of the input text, generalized features of any candidate search text, and first statistical features of the combined text corresponding to any candidate search text.

[0044] In order to avoid making recommendations and displays based solely on the statistical features of the candidate search texts, alleviate the Matthew effect of the drop-down text, and provide personalized recommendations to users, in the embodiments of the present disclosure, the target score of any candidate search text can be jointly determined based on the user features associated with the target client, the intention features of the input text, the generalized features of any candidate search text, and the first statistical features of the combined text corresponding to any candidate search text.

[0045] Among them, user characteristics include historical behavior characteristics and user attribute characteristics. Historical behavior characteristics can be used to indicate user historical behavior information, and user attribute characteristics can be used to indicate user attribute information such as gender, age, interests and hobbies. According to the login information of the target client user, the attribute information of each user associated with the target client is determined.

[0046] Step 105 : determining at least one target search text from at least one candidate search text according to the target score of any candidate search text, and sending the at least one target search text to a target client.

[0047] Furthermore, at least one target search text is determined from at least one candidate search text according to the target score of any candidate search text. For example, at least one candidate search text can be sorted from high to low according to the target score of any candidate search text to obtain a sorting sequence of at least one candidate search text. From the sorting sequence, multiple candidate search texts ranked first are obtained as target search texts. Then, the at least one target search text is sent to the target client, and the target client can display the at least one target search text according to.

[0048] In summary, by receiving the input text sent by the target client, and obtaining the intention feature of the input text, and obtaining at least one candidate search text matching the input text; for any candidate search text in the at least one candidate search text, obtaining the generalized feature of any candidate search text, wherein the generalized feature is used to indicate the category and / or source associated with any candidate search text; obtaining a combined text and a first statistical feature of the combined text, wherein the combined text is obtained based on the combination of any candidate search text and the input text; determining the target classification of any candidate search text based on the user features associated with the target client, the intention features of the input text, the generalized features of any candidate search text, and the first statistical feature of the combined text corresponding to any candidate search text value; according to the target score of any candidate search text, at least one target search text is determined from at least one candidate search text, and the at least one target search text is sent to the target client, thereby jointly determining the target score of any candidate search text according to user characteristics, intention characteristics of the input text, generalized characteristics of any candidate search text and the first statistical characteristics of the combined text corresponding to any candidate search text, so as to determine at least one target search text from at least one candidate search text according to the target score, thereby realizing the exploration and discovery of more potential high-quality search texts, avoiding recommendation and display based only on the statistical characteristics of the candidate search text, alleviating the Matthew effect of the drop-down text, and realizing personalized recommendation, thereby improving the user's search experience.

[0049] In order to clearly illustrate how the target score of any candidate search text is determined in the above embodiment based on user characteristics associated with the target client, intention characteristics of the input text, generalized characteristics of any candidate search text, and first statistical characteristics of the combined text corresponding to any candidate search text, the present disclosure proposes another method for processing search text.

[0050] Figure 2 A flowchart of a text search processing method provided in an embodiment of the present disclosure.

[0051] like Figure 2 As shown, the method for processing the search text may include the following steps:

[0052] Step 201: receiving an input text sent by a target client, acquiring an intention feature of the input text, and acquiring at least one candidate search text matching the input text.

[0053] Step 202: for any candidate search text among at least one candidate search text, obtain a generalized feature of any candidate search text.

[0054] The generalized features are used to indicate the category and / or source associated with any candidate search text.

[0055] Step 203: Obtain the combined text and the first statistical feature of the combined text.

[0056] The combined text is obtained by combining any candidate search text and the input text.

[0057] Step 204: Obtain the second statistical feature of any candidate search text and the third statistical feature of the input text.

[0058] In the disclosed embodiment, the second statistical feature is used to indicate the statistical information such as the number of clicks and exposures of any candidate search text, and the statistical information of any candidate search text can be obtained by counting the behavior information of each client. The third statistical feature is used to indicate the statistical information such as the number of clicks and exposures of the input text, and the statistical information of the input text can be obtained by counting the behavior information of each client.

[0059] Step 205 , determining a first score of any candidate search text according to the user feature, the first statistical feature, the generalized feature and the second statistical feature corresponding to any candidate search text, and the third statistical feature corresponding to the input text.

[0060] In the disclosed embodiment, deep learning technology can be used to predict the first score of any candidate search text based on user features, first statistical features, generalized features, and second statistical features corresponding to any candidate search text, and third statistical features corresponding to the input text. The first score can be used to indicate the click rate of any candidate search text.

[0061] Among them, it should be noted that, in order to reduce the amount of data calculation, before determining the first score of any candidate search text based on user characteristics, the first statistical characteristics, the generalized characteristics and the second statistical characteristics corresponding to any candidate search text, and the third statistical characteristics corresponding to the input text, the first statistical characteristics, the second statistical characteristics and the third statistical characteristics may be normalized, and for some identification information (such as the words and product IDs that the user has clicked on in the past), abnormal values ​​of the identification information (such as abnormal ID values) may be filtered, and the corresponding feature information may be determined by searching and matching.

[0062] Step 206: Determine a second score of any candidate search text according to the generalization feature of any candidate search text, the user feature, and the intention feature of the input text.

[0063] In order to avoid the Matthew effect of drop-down text, the uncertainty of each candidate search text can be explored. In the embodiment of the present disclosure, the exploration technology (Exploit&Explore) can be adopted, and the second score of any candidate search text can be determined according to the generalization characteristics, user characteristics and intention characteristics of any candidate search text, wherein the second score is used to indicate the uncertainty of any candidate search text, and a candidate search text with greater uncertainty may correspond to a larger second score.

[0064] Step 207: Obtain a target weight, and use the target weight to perform a weighted summation on the first score and the second score of any candidate search text to obtain a target score for any candidate search text.

[0065] As an example, the target weight may be preset according to actual needs or manual experience.

[0066] As another example, the target weights may be determined from model parameters of a trained prediction network.

[0067] As a possible implementation of the embodiment of the present disclosure, the target weight may be obtained by the following steps:

[0068] 1. Obtain training samples, where the training samples include sample input text and sample recommended search text of a sample client.

[0069] Among them, in the embodiments of the present disclosure, there is no restriction on the method of obtaining training samples. For example, the training samples can be obtained from an existing training set, collected online, synthesized manually, and so on.

[0070] 2. Label the sample recommended search text and obtain the labeling score of the sample recommended search text.

[0071] Among them, the annotation score is used to indicate whether the sample recommended search text is clicked.

[0072] 3. Obtain sample generalization features of sample recommendation search text and sample intent features of sample input text;

[0073] Among them, the acquisition of sample generalization features can be found in Figure 1 In step 102 of the embodiment, the acquisition of sample intention features can be seen in Figure 1 The step 101 of the embodiment is not described in detail in this disclosure.

[0074] 4. Obtain the sample combination text and the first sample statistical features of the sample combination text.

[0075] The sample combined text is obtained by combining the sample input text and the sample recommended search text.

[0076] In the embodiment of the present disclosure, the first sample statistical feature of the sample combination text can be obtained by referring to Figure 1 The present disclosure will not elaborate on step 103 of the illustrated embodiment.

[0077] 5. Input the sample user features associated with the sample client, the sample intent features of the sample input text, the sample generalization features of the sample recommended search text, and the first sample statistical features of the sample combined text information into the initial prediction model to obtain the prediction score of the sample recommended search text output by the initial prediction model; adjust the model parameters of the initial prediction model according to the difference between the annotation score and the prediction score to determine the target weight from the adjusted model parameters.

[0078] Wherein, when the initial prediction model is used to predict the prediction score of the sample recommended search text, the prediction can be performed in the following manner: the second sample statistical feature of the sample recommended search text and the third sample statistical feature of the sample input text can be obtained; the sample user feature associated with the sample client, the third sample statistical feature of the sample input text, the sample intention feature of the sample input text, the second sample statistical feature of the sample recommended search text, the sample generalization feature of the sample recommended search text and the first sample statistical feature of the sample combination text are input into the initial prediction model to obtain the prediction score of the sample recommended search text output by the initial prediction model, wherein the acquisition of the second sample statistical feature, the third sample statistical feature and the sample user feature can be referred to. Figure 2 The embodiments shown in the present disclosure will not be described in detail here.

[0079] Then, the target weight is used to perform a weighted sum of the first score and the second score of any candidate search text to obtain the target score of any candidate search text, which can be specifically expressed as follows:

[0080] Score = (1-α)Tower CTR +αTower EE (Feature EE ).

[0081] Among them, Score represents the target score, Tower CTR Indicates the first score, Tower EE represents the second score, and α represents the target weight.

[0082] Step 208: Determine at least one target search text from at least one candidate search text according to the target score of any candidate search text, and send the at least one target search text to a target client.

[0083] In summary, by obtaining a combined text and a first statistical feature of the combined text; determining a first score of any candidate search text according to user features, the first statistical features, the generalized features and the second statistical features corresponding to any candidate search text, and the third statistical features corresponding to the input text; determining a second score of any candidate search text according to the generalized features of any candidate search text, the user features and the intention features of the input text; obtaining a target weight, and using the target weight, weighted summing the first score and the second score of any candidate search text to obtain a target score of any candidate search text, thereby, according to user features, the first statistical features and the second statistical features corresponding to any candidate search text, and the third statistical features and the intention features corresponding to the input text, the click rate of any candidate search text can be estimated, according to the generalized features, the user features and the intention features of any candidate search text, the uncertainty estimation of any candidate search text can be achieved, thereby determining the target score of any candidate search text according to the click rate and uncertainty of any candidate search text, avoiding the Matthew effect of the drop-down text, and realizing personalized recommendation, thereby improving the user's search experience.

[0084] In order to clearly illustrate how the above embodiment determines the first score and the second score of any candidate search text, the present disclosure proposes another method for processing search text.

[0085] Figure 3 A flowchart of a text search processing method provided in an embodiment of the present disclosure.

[0086] like Figure 3 As shown, the method for processing the search text may include the following steps:

[0087] Step 301: receive input text sent by a target client, obtain intention features of the input text, and obtain at least one candidate search text matching the input text.

[0088] Step 302, for any candidate search text among at least one candidate search text, obtain the generalized features of any candidate search text, obtain the combined text and the first statistical features of the combined text, the second statistical features of any candidate search text and the third statistical features of the input text.

[0089] The generalized feature is used to indicate the category and / or source associated with any candidate search text, and the combined text is obtained by combining any candidate search text with the input text.

[0090] Step 303: Acquire historical behavior features and user attribute features in the user features.

[0091] In the embodiments of the present disclosure, the historical behavior features in the user features may be used to indicate the user's historical behavior information, such as the products or texts that the user has clicked on in the past.

[0092] As an example, at least one historical search keyword associated with the target client is obtained, and any historical search keyword among the at least one historical search keyword is encoded to obtain a first encoding feature of any historical search keyword; text information corresponding to any candidate search keyword is encoded to obtain a second encoding feature; the first encoding feature and the second encoding feature corresponding to any historical search keyword are inner-producted to obtain a weight sequence; the weight sequence is used to perform weighted summation on the first encoding feature corresponding to at least one historical search keyword to obtain a historical behavior feature.

[0093] For example, suppose there are N historical search keywords, and the first encoding feature of the N historical search keywords is h 1 ,...h N , weight sequence e = [s T h 1 ,...s T h N ], using the softmax function to perform weight sequence

[0094] In the disclosed embodiment, the user attribute feature may be used to indicate attribute information such as user gender, age, interests and hobbies. In the disclosed embodiment, the attribute information of each user may be determined based on the login information of the target client user.

[0095] Step 304 , concatenate the historical behavior feature, the user attribute feature, the third statistical feature of the input text, the first statistical feature corresponding to any candidate search text, the generalized feature, and the second statistical feature to obtain a first concatenated feature.

[0096] In order to improve the accuracy of click-through rate prediction of candidate search texts, historical behavior features, user attribute features, the third statistical features of the input text, the first statistical features corresponding to any candidate search text, the generalized features and the second statistical features can be spliced ​​together to obtain a first spliced ​​feature. Based on the first spliced ​​feature, the first score of each candidate search text is predicted.

[0097] Step 305: input the first concatenated feature into the multi-layer neural network of the first network branch of the prediction model to obtain a first score of any candidate search text.

[0098] As a possible implementation of the embodiment of the present disclosure, a multi-layer neural network of the first network branch of the prediction model may be used, and based on the first concatenated feature, the first score of each candidate search text may be predicted, which may be specifically expressed as follows:

[0099] feature concat =concatenate(historical behavior feature, user attribute feature, third statistical feature of input text, first statistical feature, second statistical feature);

[0100] Tower CTR =MLP(feature concat );

[0101] Among them, feature concat Represents the first splicing feature, Tower CTR Indicates the first score.

[0102] Step 306: Concatenate the user feature, the generalized feature of any candidate search text, and the intention feature of the input text to obtain a second concatenated feature.

[0103] At the same time, in order to explore the uncertainty of any candidate search text, uncertainty estimation can be performed based on user features, generalized features of any candidate search text and intention features of the input text. As an example, user features, generalized features of any candidate search text and intention features of the input text can be spliced ​​to obtain a second spliced ​​feature. Based on the second spliced ​​feature, the uncertainty of any candidate search text is estimated.

[0104] Step 307: Use the multi-layer neural network of the second network branch of the prediction model to extract features from the second concatenated feature information to obtain user intention features.

[0105] In the embodiment of the present disclosure, the multi-layer neural network of the second network branch of the prediction model can be used to extract the second splicing feature information to obtain the user intention feature, which can be specifically expressed as follows:

[0106] User intention feature = MLP (user feature, generalized feature of any candidate search text, intention feature of input text).

[0107] Step 308: Determine the predicted click rate of any candidate search text based on the user intent feature, and predict the click rate, which is input into the activation function of the second network branch to obtain a second score output by the activation function.

[0108] As an example, the user intention feature is input into the sparse variational Gaussian process network (SVGP) of the second network branch to obtain multiple output click-through rates (logit) of any candidate search text output by the sparse variational Gaussian process network; the mean and variance of the multiple output click-through rates of any candidate search text are determined; according to the mean and variance of the multiple output click-through rates of any candidate search text, a Gaussian distribution map of the multiple output click-through rates is generated, and sampling is performed from the Gaussian distribution map of any candidate search text to obtain the predicted click-through rate of any candidate search text. Specifically, it can be expressed as the following expression:

[0109] μ,σ 2 =SVGP (user intent feature);

[0110] Then, a relevant sampling algorithm (such as Thompson sampling) is used to obtain the predicted click rate:

[0111] Predicted click rate = sample from N(μ,σ 2 );

[0112] Therefore, the predicted click rate is input into the activation function of the second network branch to obtain a second score output by the activation function.

[0113] It should be noted that the greater the variance of the output click rate, the greater the uncertainty of the candidate search text. The greater the uncertainty of the candidate search text, the greater the corresponding second score, which can correspond to a greater display opportunity.

[0114] In the embodiments of the present disclosure, Figure 4 As shown, the prediction model has two network branches. The first network branch is modeled using a deep neural network, and can predict a first score for each candidate search text. The first score is used to indicate the click rate of the candidate search text. The second network branch can predict a second score for each candidate search text. The second score is used to indicate the uncertainty of the candidate search text. The second network branch can correct the mean and variance of each candidate search text through user intent features (user feedback behavior). When the number of clicks / exposures of the candidate search text gradually increases, the variance of the candidate search text tends to stabilize. The smaller the uncertainty, the smaller the second score.

[0115] Step 309: Obtain a target weight, and use the target weight to perform a weighted summation on the first score and the second score of any candidate search text to obtain a target score for any candidate search text.

[0116] Step 310: Determine at least one target search text from at least one candidate search text according to the target score of any candidate search text, and send the at least one target search text to a target client.

[0117] In summary, by obtaining historical behavior features and user attribute features in user features; splicing the historical behavior features, user attribute features, the third statistical features of the input text, the first statistical features corresponding to any candidate search text, the generalized features and the second statistical features to obtain a first spliced ​​feature; inputting the first spliced ​​feature into the multi-layer neural network of the first network branch of the prediction model to obtain a first score for any candidate search text; splicing the user features, the generalized features of any candidate search text and the intention features of the input text to obtain a second spliced ​​feature; using the multi-layer neural network of the second network branch of the prediction model to perform Feature extraction is performed to obtain user intention features; based on the user intention features, the predicted click rate of any candidate search text is determined, and the predicted click rate is input into the activation function of the second network branch to obtain the second score output by the activation function, thereby, through the first network branch and the second network branch of the prediction network, the first score and the second score of the candidate search text can be predicted, and the target score for ranking the candidate search text is determined according to the first score and the second score, so that when performing search text recommendation, the uncertainty of the candidate search text is considered while the click rate of the candidate search text is considered, thereby alleviating the Matthew effect of the drop-down text and improving the user's search experience.

[0118] In order to implement the above embodiment, the present disclosure also proposes a processing device for searching text.

[0119] Figure 5 A schematic diagram of the structure of a text search processing device provided in an embodiment of the present disclosure.

[0120] like Figure 5 As shown, the text search processing device 500 includes: a receiving module 510, a first acquisition module 520, a second acquisition module 530, a first determination module 540 and a second determination module 550.

[0121] Among them, the receiving module 510 is used to receive the input text sent by the target client, and obtain the intention feature of the input text, and obtain at least one candidate search text matching the input text; the first acquisition module 520 is used to obtain the generalized feature of any candidate search text in at least one candidate search text, wherein the generalized feature is used to indicate the category and / or source to which any candidate search text belongs; the second acquisition module 530 is used to obtain the combined text and the first statistical feature of the combined text, wherein the combined text is obtained based on the combination of any candidate search text and the input text; the first determination module 540 is used to determine the target score of any candidate search text based on the user features associated with the target client, the intention feature of the input text, the generalized feature of any candidate search text and the first statistical feature of the combined text corresponding to any candidate search text; the second determination module 550 is used to determine at least one target search text from at least one candidate search text based on the target score of any candidate search text, and send the at least one target search text to the target client.

[0122] As a possible implementation method of the embodiment of the present disclosure, the first determination module 540 is specifically used to: obtain the second statistical feature of any candidate search text and the third statistical feature of the input text; determine the first score of any candidate search text based on user features, the first statistical feature, the generalization feature and the second statistical feature corresponding to any candidate search text, and the third statistical feature corresponding to the input text; determine the second score of any candidate search text based on the generalization feature, the user feature and the intention feature of the input text of any candidate search text; obtain the target weight, and use the target weight to perform weighted summation of the first score and the second score of any candidate search text to obtain the target score of any candidate search text.

[0123] As a possible implementation method of the embodiment of the present disclosure, the first determination module 540 is also used to: obtain historical behavior features and user attribute features in user features; splice the historical behavior features, the user attribute features, the third statistical features of the input text, the first statistical features corresponding to any candidate search text, the generalized features and the second statistical features to obtain a first spliced ​​feature; input the first spliced ​​feature into the multi-layer neural network of the first network branch of the prediction model to obtain a first score for any candidate search text.

[0124] As a possible implementation method of the embodiment of the present disclosure, the first determination module 540 is also used to: obtain at least one historical search keyword associated with the target client, and encode any historical search keyword among the at least one historical search keyword to obtain a first encoding feature of any historical search keyword; encode any candidate search text to obtain a second encoding feature; perform an inner product of the first encoding feature and the second encoding feature corresponding to any historical search keyword to obtain a weight sequence; use the weight sequence to perform weighted summation of the first encoding feature corresponding to at least one historical search keyword to obtain a historical behavior feature.

[0125] As a possible implementation method of the embodiment of the present disclosure, the first determination module 540 is also used to: splice user features, generalized features of any candidate search text and intention features of the input text to obtain second spliced ​​features; use a multi-layer neural network of a second network branch of a prediction model to perform feature extraction on the second spliced ​​feature information to obtain user intention features; determine a predicted click-through rate of any candidate search text based on the user intention features; and input the predicted click-through rate into an activation function of the second network branch to obtain a second score output by the activation function.

[0126] As a possible implementation method of the embodiment of the present disclosure, the first determination module 540 is also used to: input the user intention feature into the sparse variational Gaussian process network of the second network branch to obtain multiple output click rates of any candidate search text output by the sparse variational Gaussian process network; determine the mean and variance of the multiple output click rates of any candidate search text; generate a Gaussian distribution map of the multiple output click rates based on the mean and variance of the multiple output click rates of any candidate search text, and sample from the Gaussian distribution map of any candidate search text to obtain the predicted click rate of any candidate search text.

[0127] As a possible implementation method of the embodiment of the present disclosure, the first determination module is also used to: obtain training samples, wherein the training samples include sample input text and sample recommended search text of the sample client; annotate the sample recommended search text to obtain an annotation score of the sample recommended search text, wherein the annotation score is used to indicate whether the sample recommended search text is clicked; obtain sample generalization features of the sample recommended search text and sample intent features of the sample input text; obtain sample combination text and a first sample statistical feature of the sample combination text, wherein the sample combination text is obtained by combining the sample input text and the sample recommended search text; input the sample user features associated with the sample client, the sample intent features of the sample input text, the sample generalization features of the sample recommended search text, and the first sample statistical features of the sample combination text information into the initial prediction model to obtain the prediction score of the sample recommended search text output by the initial prediction model; and adjust the model parameters of the initial prediction model according to the difference between the annotation score and the prediction score to determine the target weight from the adjusted model parameters.

[0128] As a possible implementation method of the embodiment of the present disclosure, the first determination module is also used to: obtain the second sample statistical feature of the sample recommended search text and the third sample statistical feature of the sample input text; input the sample user feature associated with the sample client, the third sample statistical feature of the sample input text, the sample intention feature of the sample input text, the second sample statistical feature of the sample recommended search text, the sample generalization feature of the sample recommended search text and the first sample statistical feature of the sample combined text into the initial prediction model to obtain the prediction score of the sample recommended search text output by the initial prediction model.

[0129] The search text processing device of the disclosed embodiment receives input text sent by a target client, obtains the intention feature of the input text, and obtains at least one candidate search text matching the input text; obtains the generalized feature of any candidate search text among the at least one candidate search text, wherein the generalized feature is used to indicate the category and / or source to which any candidate search text belongs; obtains a combined text and a first statistical feature of the combined text, wherein the combined text is obtained by combining any candidate search text and the input text; determines based on user features associated with the target client, the intention feature of the input text, the generalized feature of any candidate search text, and the first statistical feature of the combined text corresponding to any candidate search text A target score for any candidate search text; based on the target score for any candidate search text, at least one target search text is determined from at least one candidate search text, and the at least one target search text is sent to the target client, thereby jointly determining the target score for any candidate search text based on user features, the intent features of the input text, the generalized features of any candidate search text, and the first statistical features of the combined text corresponding to any candidate search text, thereby determining at least one target search text from at least one candidate search text based on the target score, thereby achieving the exploration and discovery of more potential high-quality search texts, avoiding recommendation and display based solely on the statistical features of the candidate search text, alleviating the Matthew effect of the drop-down text, and improving the user's search experience.

[0130] It should be noted that the above explanation of the embodiment of the search text processing method is also applicable to the search text processing device of this embodiment, and will not be repeated here.

[0131] In order to implement the above embodiments, the present application also proposes an electronic device, such as Figure 6 As shown, Figure 6 The invention is a block diagram of an electronic device for searching text according to an exemplary embodiment.

[0132] like Figure 6 As shown, the electronic device 600 includes:

[0133] The memory 610 and the processor 620, and the bus 630 connecting different components (including the memory 610 and the processor 620), the memory 610 stores a computer program, and when the processor 620 executes the program, the method for processing the search text described in the embodiment of the present disclosure is implemented.

[0134] Bus 630 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processor or a local bus using any of a variety of bus architectures. By way of example, these architectures include, but are not limited to, an Industry Standard Architecture (ISA) bus, a Micro Channel Architecture (MAC) bus, an Enhanced ISA bus, a Video Electronics Standards Association (VESA) local bus, and a Peripheral Component Interconnect (PCI) bus.

[0135] The electronic device 600 typically includes a variety of electronic device readable media. These media can be any available media that can be accessed by the electronic device 600, including volatile and non-volatile media, removable and non-removable media.

[0136] The memory 610 may also include computer system readable media in the form of volatile memory, such as random access memory (RAM) 640 and / or cache memory 650. The electronic device 600 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the storage system 660 may be used to read and write non-removable, non-volatile magnetic media ( Figure 6 not shown, usually called a "hard drive"). Although Figure 6 Not shown, a disk drive for reading and writing to a removable non-volatile disk (e.g., a "floppy disk"), and an optical drive for reading and writing to a removable non-volatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical medium) may be provided. In these cases, each drive may be connected to bus 630 via one or more data medium interfaces. Memory 610 may include at least one program product having a set (e.g., at least one) of program modules that are configured to perform the functions of the various embodiments of the present disclosure.

[0137] A program / utility 680 having a set (at least one) of program modules 670 may be stored, for example, in the memory 610, such program modules 670 including but not limited to an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment. The program modules 670 generally perform the functions and / or methods of the embodiments described in the present disclosure.

[0138] The electronic device 600 may also communicate with one or more external devices 690 (e.g., keyboards, pointing devices, displays, etc.), one or more devices that enable a user to interact with the electronic device 600, and / or any device that enables the electronic device 600 to communicate with one or more other computing devices (e.g., network cards, modems, etc.). Such communication may be performed through an input / output (I / O) interface 692. Furthermore, the electronic device 600 may also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) through a network adapter 693. Figure 6 As shown, the network adapter 693 communicates with other modules of the electronic device 600 via the bus 630. It should be understood that although Figure 6 Not shown, other hardware and / or software modules may be used in conjunction with electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0139] The processor 620 executes various functional applications and data processing by running the programs stored in the memory 610 .

[0140] It should be noted that the implementation process and technical principles of the electronic device of this embodiment refer to the aforementioned explanation of the search text processing method of the embodiment of the present disclosure, and will not be repeated here.

[0141] In order to implement the above embodiments, the present disclosure further proposes a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method for processing the search text described in the above embodiments is implemented.

[0142] In order to implement the above embodiments, the present disclosure further provides a computer program product. When an instruction processor in the computer program product is executed, the method for processing search text described in the above embodiments is executed.

[0143] In the description of this specification, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0144] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0145] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A method for processing a search text, characterized in that: include: Receiving an input text sent by a target client, and obtaining an intention feature of the input text, and obtaining at least one candidate search text matching the input text; For any candidate search text among the at least one candidate search text, obtaining a generalization feature of the any candidate search text, wherein the generalization feature is used to indicate a category and / or source associated with the any candidate search text; Acquire a combined text and a first statistical feature of the combined text, wherein the combined text is obtained by combining any one of the candidate search texts and the input text; Determining a target score for any candidate search text according to a user feature associated with a target client, an intention feature of the input text, a generalization feature of any candidate search text, and a first statistical feature of a combined text corresponding to any candidate search text; At least one target search text is determined from the at least one candidate search text according to the target score of any one of the candidate search texts, and the at least one target search text is sent to the target client.

2. The method according to claim 1, characterized in that The step of determining a target score of any candidate search text according to a user feature associated with a target client, an intention feature of the input text, a generalization feature of any candidate search text, and a first statistical feature of a combined text corresponding to any candidate search text comprises: Acquire a second statistical feature of any candidate search text and a third statistical feature of the input text; Determine a first score of any candidate search text according to the user feature, the first statistical feature, the generalized feature, and the second statistical feature corresponding to any candidate search text, and the third statistical feature corresponding to the input text; Determining a second score of any candidate search text according to the generalization feature of any candidate search text, the user feature, and the intention feature of the input text; A target weight is obtained, and the target weight is used to perform a weighted summation on the first score and the second score of any candidate search text to obtain a target score of any candidate search text.

3. The method according to claim 2, characterized in that The step of determining a first score of any candidate search text according to the user feature, the first statistical feature, the generalized feature, and the second statistical feature corresponding to any candidate search text, and the third statistical feature corresponding to the input text comprises: Acquire historical behavior characteristics and user attribute characteristics from the user characteristics; The historical behavior feature, the user attribute feature, the third statistical feature of the input text, the first statistical feature, the generalized feature, and the second statistical feature corresponding to any candidate search text are concatenated to obtain a first concatenated feature; The first concatenated feature is input into a multi-layer neural network of a first network branch of a prediction model to obtain a first score of any candidate search text.

4. The method according to claim 3, characterized in that The acquiring of the historical behavior features in the user features includes: Acquire at least one historical search keyword associated with the target client, and encode any one of the at least one historical search keyword to obtain a first encoding feature of the any one of the historical search keywords; Encoding any of the candidate search texts to obtain a second encoding feature; Perform an inner product of the first coding feature and the second coding feature corresponding to any one of the historical search keywords to obtain a weight sequence; The weight sequence is used to perform weighted summation on the first coding feature corresponding to the at least one historical search keyword to obtain the historical behavior feature.

5. The method according to claim 2, characterized in that: The determining the second score of any candidate search text according to the generalization feature of any candidate search text, the user feature and the intention feature of the input text comprises: Concatenate the user feature, the generalized feature of any candidate search text, and the intention feature of the input text to obtain a second concatenated feature; Using a multi-layer neural network of a second network branch of the prediction model to extract features from the second splicing feature information to obtain user intention features; Determining a predicted click rate of any candidate search text based on the user intention feature; The predicted click rate is input into the activation function of the second network branch to obtain a second score output by the activation function.

6. The method according to claim 5, characterized in that The step of predicting the click rate of any candidate search text based on the user intention feature includes: Inputting the user intention feature into the sparse variational Gaussian process network of the second network branch to obtain multiple output click rates of any candidate search text output by the sparse variational Gaussian process network; Determine the mean and variance of multiple output click rates of any candidate search text; A Gaussian distribution graph of the multiple output click rates of any candidate search text is generated according to the means and variances of the multiple output click rates, and sampling is performed from the Gaussian distribution graph of any candidate search text to obtain a predicted click rate of any candidate search text.

7. The method according to claim 2, characterized in that The obtaining of the target weight comprises: Acquire a training sample, wherein the training sample includes a sample input text and a sample recommended search text of a sample client; Annotating the sample recommended search text to obtain an annotation score of the sample recommended search text, wherein the annotation score is used to indicate whether the sample recommended search text is clicked; Acquire a sample generalization feature of the sample recommended search text and a sample intention feature of the sample input text; Acquire a sample combined text and a first sample statistical feature of the sample combined text, wherein the sample combined text is obtained by combining the sample input text and the sample recommended search text; Inputting the sample user features associated with the sample client, the sample intent features of the sample input text, the sample generalization features of the sample recommended search text, and the first sample statistical features of the sample combined text information into the initial prediction model to obtain the prediction score of the sample recommended search text output by the initial prediction model; According to the difference between the labeled score and the predicted score, the model parameters of the initial prediction model are adjusted to determine the target weight from the adjusted model parameters.

8. The method according to claim 7, characterized in that The step of inputting the sample user features associated with the sample client, the sample intent features of the sample input text, the sample generalization features of the sample recommended search text, and the first sample statistical features of the sample combined text information into the initial prediction model to obtain the prediction score of the sample recommended search text output by the initial prediction model includes: Acquire a second sample statistical feature of the sample recommended search text and a third sample statistical feature of the sample input text; The sample user features associated with the sample client, the third sample statistical features of the sample input text, the sample intention features of the sample input text, the second sample statistical features of the sample recommended search text, the sample generalization features of the sample recommended search text and the first sample statistical features of the sample combined text are input into the initial prediction model to obtain the prediction score of the sample recommended search text output by the initial prediction model.

9. A text search processing device, characterized in that: include: A receiving module, configured to receive an input text sent by a target client, obtain an intention feature of the input text, and obtain at least one candidate search text matching the input text; A first acquisition module, configured to acquire, for any candidate search text among the at least one candidate search text, a generalization feature of the any candidate search text, wherein the generalization feature is used to indicate a category and / or source to which the any candidate search text belongs; A second acquisition module is used to acquire a combined text and a first statistical feature of the combined text, wherein the combined text is obtained by combining any one of the candidate search texts and the input text; A first determination module, configured to determine a target score of any candidate search text according to a user feature associated with a target client, an intention feature of the input text, a generalization feature of any candidate search text, and a first statistical feature of a combined text corresponding to any candidate search text; The second determination module is used to determine at least one target search text from the at least one candidate search text according to the target score of any candidate search text, and send the at least one target search text to the target client.

10. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the method for processing search text as claimed in any one of claims 1 to 8 is implemented.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for processing a search text as claimed in any one of claims 1 to 8 is implemented.

Images