CN111814058A - Push method, device, electronic device and storage medium based on user intent - Google Patents

Abstract

Embodiments of the present application disclose a user-intent-based push method, apparatus, electronic device, and computer-readable medium, which relate to the technical field of computer applications. The method includes: acquiring a search term input by a user and a set of mapping relationships of content to be pushed, wherein the set of mapping relationships of content to be pushed includes: a mapping relationship between content to be pushed and keywords; and based on a pre-trained semantic understanding model, respectively acquiring Describe the search semantic vector corresponding to the search word and the content semantic vector corresponding to the keyword; calculate the similarity between the search semantic vector and the content semantic vector, and determine the target content from the content semantic vector according to the similarity Semantic vector; according to the semantic vector of the target content, obtain the content to be pushed corresponding to the semantic vector of the target content, and push it. Therefore, the user's search intent can be mined according to the user's search terms, so as to effectively push content.

Description

Push method, device, electronic device and storage medium based on user intent

technical field

The embodiments of the present application relate to the technical field of computer applications, and more particularly, to a push method, apparatus, electronic device, and storage medium based on user intent.

Background technique

With the rapid development of the mobile era, the information on the Internet continues to increase, and users often search for the information they need in the massive information through search engines. Search engine promotion is also the most effective Internet advertising channel at present. However, the current method of advertising delivery relies on manually constructing a label system to select suitable advertisements for delivery among a large number of advertisements. This method relies on domain experience, is highly subjective, and cannot meet complex and changeable advertising requirements. Therefore, how to accurately obtain the user's intention information and push advertisements according to the user's intention to improve the push efficiency is an urgent problem to be solved.

SUMMARY OF THE INVENTION

In view of the above problems, embodiments of the present application provide a push method, device, electronic device, and storage medium based on user intent, which can effectively push content.

In a first aspect, an embodiment of the present application provides a push method based on user intent, the method includes: acquiring a search term input by a user and a set of mapping relationships of content to be pushed, where the set of mapping relationships of content to be pushed includes: a set of mapping relationships to be pushed The mapping relationship between content and keywords; based on the pre-trained semantic understanding model, respectively obtain the search semantic vector corresponding to the search word and the content semantic vector corresponding to the keyword; calculate the search semantic vector and the content The similarity of the semantic vectors, the target content semantic vector is determined from the content semantic vector according to the similarity; according to the target content semantic vector, the content to be pushed corresponding to the target content semantic vector is obtained and pushed.

In a second aspect, an embodiment of the present application further provides a push device based on user intent, the device includes: an information acquisition module that obtains a search term input by a user and a set of mapping relationships of content to be pushed, where the mapping relationship of content to be pushed is obtained. The set includes: the mapping relationship between the content to be pushed and the keywords; the vector acquisition module, based on the pre-trained semantic understanding model, respectively acquires the search semantic vector corresponding to the search word and the content semantic vector corresponding to the keyword; determining module, calculating the similarity between the search semantic vector and the content semantic vector, and determining the target content semantic vector from the content semantic vector according to the similarity; the processing module, according to the target content semantic vector, obtains and Push the content to be pushed corresponding to the semantic vector of the target content.

In a third aspect, embodiments of the present application further provide an electronic device, including: one or more processors; a memory; and one or more application programs, wherein the one or more application programs are stored in the memory and configured to be executed by the one or more processors, the one or more application programs configured to perform the above method.

In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, where a program code is stored in the computer-readable storage medium, and the program code can be invoked by a processor to execute the above method.

Embodiments of the present application disclose a user-intent-based push method, apparatus, electronic device, and computer-readable medium, which relate to the technical field of computer applications. The method includes: acquiring a search term input by a user and a set of mapping relationships of content to be pushed, wherein the set of mapping relationships of content to be pushed includes: a mapping relationship between content to be pushed and keywords; and based on a pre-trained semantic understanding model, respectively acquiring Describe the search semantic vector corresponding to the search word and the content semantic vector corresponding to the keyword; calculate the similarity between the search semantic vector and the content semantic vector, and determine the target content from the content semantic vector according to the similarity Semantic vector; according to the semantic vector of the target content, obtain the content to be pushed corresponding to the semantic vector of the target content, and push it. Therefore, the user's search intent can be mined according to the user's search terms, so as to effectively push content.

Description of drawings

In order to illustrate the technical solutions in the embodiments of the present application more clearly, the following briefly introduces the drawings that are used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. not all examples. Based on the embodiments of the present application, all other embodiments and drawings obtained by persons of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

FIG. 1 shows a schematic diagram of an application environment suitable for this embodiment of the present application.

FIG. 2 shows a schematic flowchart of a user-intent-based push method provided by an embodiment of the present application.

FIG. 3 shows a schematic flowchart of a push method based on user intent provided by another embodiment of the present application.

FIG. 4 shows a schematic flowchart of a push method based on user intent provided by another embodiment of the present application.

FIG. 5 shows a schematic flowchart of a push method based on user intent provided by still another embodiment of the present application.

FIG. 6 shows a block diagram of a module of a push device based on user intention provided by an embodiment of the present application.

FIG. 7 shows a block diagram of an electronic device for implementing the push method based on user intent according to the embodiment of the present application;

FIG. 8 shows a block diagram of a computer-readable storage medium for implementing the push method based on user intent according to the embodiment of the present application.

Detailed ways

In order to make those skilled in the art better understand the solutions of the present application, the following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the accompanying drawings in the embodiments of the present application. It should be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application.

At present, with the continuous development of the Internet, the number of Internet users has exploded. Advertising users often search for the information they need in the massive information on the Internet. Search engines have gradually become an indispensable tool, and advertising has gradually used the Internet as a carrier to spread. At present, search engines generally search according to the search text input by the user to obtain search results related to the search text, and provide the search results to the user for viewing. Therefore, how to obtain the user's search intent according to the user's search text, to achieve effective advertising becomes very important.

However, currently understanding the user's search intent is usually to use a text classification model to classify the user's search text, obtain the label corresponding to the user's search text, and obtain the association between the user's search text and the advertisement according to the label of the advertisement to be placed, so as to perform advertisement placement. The inventor has studied the difficulties in the current method of advertising, and found that due to the complex and changeable needs of advertisers, the corresponding labeling system is also extremely complex, usually requiring hundreds or thousands of classification labels, and the text classification model is supervised. Model, each category needs to be labeled from massive user search data to obtain training sample data for training. Therefore, obtaining the labels of training sample data requires a lot of manpower, and the accuracy of the classification model depends on the training samples. Quality and quantity of data.

Manually constructing a labeling system relies on personal domain knowledge, and the existing human error may affect the recognition ability of the final model. Although it is possible to obtain a batch of keywords corresponding to labels by means of data analysis, and to label the user search text by means of keyword matching to quickly obtain training sample data, this method has two drawbacks. On the one hand, because the words themselves are ambiguous, the same words may express completely different user intentions in different scenarios, and there may be a lot of interference in the training sample data obtained by the keyword matching method. The recognition ability of the text classification model is also poor; on the other hand, the distribution space of the training sample data obtained by keyword matching is relatively narrow, and only the representation of part of the text data in the current label can be found, and it is easy to lose a large amount of the current label. The corpus space indicates that the corpus space of the search text that can be recognized by the model obtained through such training sample data is also relatively narrow, and the model recognition effect is not good, resulting in an unsatisfactory final advertisement placement.

The inventor has studied the difficulties of the current advertising delivery methods, and more importantly, comprehensively considered the advertising delivery requirements of actual scenarios, and proposed that the present application implement a user-intent-based push method, device, electronic device and storage medium, and mine users according to the user's search terms. search intent to effectively push content.

In order to better understand a user-intent-based push method, device, electronic device, and storage medium provided by the embodiments of the present application, an application environment applicable to the embodiments of the present application is first described below.

Referring to FIG. 1, FIG. 1 shows a schematic diagram of an application environment suitable for the embodiment of the present application. The push method based on the user's intention provided by the embodiment of the present application can be applied to the polymorphic interaction system 10 as shown in FIG. 1 . The polymorphic interaction system 10 includes a terminal device 100 and a server 200 , and the server 200 is connected in communication with the terminal device 100 . The server 200 may be a traditional server or a cloud server, which is not specifically limited herein.

In some embodiments, if the user logs in through an account at the user terminal, all the information corresponding to the account may be stored in the storage space of the server 100 . The server 100 may be a single server, a server cluster, a local server, or a cloud server. There are multiple application programs installed in the user terminal, and the server 100 can push some content to the user terminal. The content is pushed to the user corresponding to the user terminal.

The server 100 can be connected to multiple user terminals, and can push the content to be pushed to all user terminals, or select a user terminal according to some policies, and push the content to be pushed to the selected user terminal. . The specific policy may be determined according to the content to be pushed and the user corresponding to each user terminal. In this embodiment of the present application, the content to be pushed may be advertisement information, for example, commodity discount information of a certain e-commerce application.

The above application environment is only an example for convenience of understanding, and it can be understood that the embodiments of the present application are not limited to the above application environment.

The user-intent-based push method, device, terminal device, and storage medium provided by the embodiments of the present application will be described in detail below through specific embodiments.

Please refer to FIG. 2. FIG. 2 shows a schematic flowchart of a user-intent-based push method provided by an embodiment of the present application. The user-intent-based push method provided by this embodiment can be applied to the server in the above-mentioned system, that is, the method The execution body of the .server may be the above-mentioned server, and the method is used to improve the accuracy of the content pushed for the user. Specifically, as shown in FIG. 2 , the method includes: S110 to S140.

S110: Obtain a set of mapping relationships between the search term input by the user and the content to be pushed.

The input form of the search term input by the user may be text or voice, etc. Specifically, if the input form of the user is text, the search term in the form of text input by the user is obtained, and if the input form of the user is voice, the voice information Search terms converted to text.

The set of mapping relationships of the content to be pushed includes: the mapping relationship between the content to be pushed and the keywords, and the content to be pushed is a thing waiting to be pushed or a virtual thing, which can be in the form of text, video, or pictures The form is not limited here. For example, the content to be pushed may be news, list, video, etc. to be pushed, or may be recommended products, stores, services, and the like. The mapping relationship between the push content and the keywords may be that one keyword corresponds to one push content, or one keyword may correspond to multiple push contents, which is not limited herein. As a way, the set of content mapping relationships to be pushed can be stored in various data structures. For example, according to different types of data structures, the set of content vector mapping relationships can be stored in the form of lists, tables, hash tables, arrays, trees, and the like.

In some embodiments, the keyword in the content mapping relationship set to be pushed may be the title corresponding to the content to be pushed, because the title is usually a refinement and summary of the content to be pushed, and can be used to represent the main content of the content to be pushed. The set of mapping relationships between the content and the title may be preset and stored locally in the terminal device or in the server.

In other embodiments, the keyword in the content mapping relationship set to be pushed may also be a piece of key text of the content to be pushed. Specifically, when the content to be pushed includes text content, if the content to be pushed includes an abstract, the key word of the content to be pushed may be the abstract of the content to be pushed; if the content to be pushed does not include an abstract, the key word of the content to be pushed may be the abstract of the content to be pushed It can be the first paragraph or the last text of the push content, because the first paragraph of the text content usually summarizes the content of the page, and the last paragraph of the text content usually summarizes the content of the page.

S120: Based on the pre-trained semantic understanding model, respectively obtain the search semantic vector corresponding to the search word and the content semantic vector corresponding to the keyword.

Among them, the semantic vector is a vector used to represent the semantics of the text information obtained by mapping the text information into the preset vector space. By using the search word as the input of the pre-trained semantic understanding model, the search semantics corresponding to the search word can be obtained. Vector, by taking the keyword as the input of the pre-trained semantic understanding model, the content semantic vector corresponding to the keyword can be obtained. Among them, the semantic understanding model can be a supervised model or an unsupervised model, which is a model pre-trained from text data.

Specifically, the semantic understanding model includes but is not limited to: one of Deep Neural Networks (DNN), Recurrent Neural Networks (RNN), Convolutional Neural Networks (CNN), Transformer, or combination. In some embodiments, the semantic understanding model may be a transformer-based bidirectional encoding representation network model (Bidirectional Encoder Representation from Transformers, BERT), a word vector model (doc2Vec), and the like.

As a method, the search semantic vector corresponding to the search term and the content semantic vector corresponding to the keyword can be obtained according to the specified time frequency, so as to obtain the change of the user's search intention and push the information more effectively, wherein the specified time frequency It can be set according to actual needs. For example, when the time frequency is one day, the search semantic vector corresponding to the search word and the content semantic vector corresponding to the keyword can be obtained every day based on the pre-trained semantic understanding model.

S130: Calculate the similarity between the search semantic vector and the content semantic vector, and determine the target content semantic vector from the content semantic vector according to the similarity.

After obtaining the search semantic vector corresponding to the search word and the content semantic vector corresponding to the keyword, the similarity between the semantic vector and the content vector can be calculated, where the similarity is used to represent the correlation between the search semantic vector and the content semantic vector, that is, The correlation between the search term and the keyword, specifically, the higher the similarity, the more the keyword can express the user intent reflected by the search term.

In some embodiments, based on a pre-trained semantic understanding model, respectively obtaining the search semantic vector corresponding to the search term and the content semantic vector corresponding to the keyword may be performed by a server, and the obtained search semantic vector and content semantic vector may be Stored in the offline server or the local database of the terminal device, the corresponding push device or terminal device in the offline server can calculate the similarity between the search semantic vector and the content semantic vector in real time through the stored data of the search semantic vector and content semantic vector.

Specifically, the target semantic vector is determined from the content semantic vector according to the similarity, and the content semantic vector whose similarity satisfies a specified condition may be used as the target semantic vector. The specified condition is a condition related to the actual push scene, and the target semantic vector may be a vector, or multiple vectors. In one way, the specified condition may be greater than a specific threshold, and accordingly, a content semantic vector corresponding to greater than a specific threshold is used as the target semantic vector. The specific condition may also be to obtain N content semantic vectors with the largest similarity, where N is a preset integer greater than 0, and correspondingly, the N content semantic vectors with the largest similarity are used as the target semantic vector.

As an embodiment, keywords of the content to be pushed can be optimized according to the similarity between the search semantic vector and the content semantic vector, so as to better attract users and improve the conversion rate of the content to be pushed. For example, when the content to be pushed is advertisement information, the advertiser can obtain the user group that the advertisement information wants to reach. The keyword corresponding to the content semantic vector of interest, so as to guide the advertiser to optimize the keyword in the advertisement information according to the keyword, and improve the conversion effect of the advertisement.

S140: Acquire the content to be pushed corresponding to the semantic vector of the target content according to the semantic vector of the target content, and push the content.

After the target semantic vector is obtained, the corresponding keyword can be obtained according to the target semantic vector, the content to be pushed corresponding to the keyword is obtained based on the mapping relationship set of the content to be pushed, and the content to be pushed is pushed.

As a method, the similarity corresponding to each target content semantic vector can be obtained, and the content to be pushed corresponding to the target semantic vector can be sorted in descending order of similarity, so that the target semantic vector with higher similarity can be sorted. The corresponding content to be pushed is arranged in the front, which is convenient for users to select.

As a method, the content semantic vector can also be specified in advance, the similarity between the content semantic vector and each search semantic vector can be calculated separately, and the search semantic vector whose similarity satisfies the specified conditions can be obtained. The user of the word pushes, and the push information is the content to be pushed corresponding to the specified content semantic vector.

In the method for pushing information based on user intent provided by the embodiments of the present application, after obtaining a set of mapping relationships between a search term input by a user and the content to be pushed, and the set of mapping relationships between the content to be pushed includes the mapping relationship between the content to be pushed and keywords, based on The pre-trained semantic understanding model obtains the search semantic vector corresponding to the search word and the content semantic vector corresponding to the keyword, calculates the similarity between the search semantic vector and the content semantic vector, and determines the target content semantic vector from the content semantic vector according to the similarity , so that according to the semantic vector of the target content, the content to be pushed corresponding to the semantic vector of the target content is obtained and pushed. Obtaining the content to be pushed by searching for the similarity between the semantic vector and the content semantic vector can make the pushed content more in line with the user's search intent, thereby improving the accuracy of the pushed content.

Please refer to FIG. 3. FIG. 3 shows a schematic flowchart of a user-intent-based push method provided by another embodiment of the present application, which is applied to the server in the above-mentioned system, that is, the execution body of the method may be the above-mentioned server. For improving the accuracy of the content pushed for the user, specifically, as shown in FIG. 3 , the method includes: S210 to S260.

S210: Acquire a set of mapping relationships between the search term input by the user and the content to be pushed.

Step S210 can refer to step S110.

In some embodiments, the keywords of the content to be pushed may be keywords included in the title of the content to be pushed, and the user's historical search data may be obtained through the terminal burying log, wherein the user's historical search data may include the user's historical search terms The to-be-pushed content browsed by the user corresponding to the historical search term. As one way, the mapping relationship between the content to be pushed and the title of the content to be pushed can be obtained according to the content to be pushed browsed by the user, so that the mapping relationship is stored in the content mapping relationship set to be pushed; When the content to be pushed is obtained, the mapping relationship between the content to be pushed and the title of the content to be pushed is stored.

S220: Obtain the feature vector of the search word and the feature vector of the keyword respectively.

In the embodiment of the present application, the semantic understanding model used to obtain the semantic vector is the BERT model, and the feature vector of the search word and the feature vector of the keyword are obtained as the input of the BERT model, and the feature vector is the feature vector generated by the text information through the encoder, The dimension of the data can be reduced by encoding the text, but the vector cannot obtain different semantic representations according to the context. By inputting the feature vector into the BERT model, the corresponding semantic vector can be obtained. The semantic vector can be semantically represented according to the context of the sentence.

In some embodiments, obtaining the feature vector of the search word and the feature vector of the keyword respectively includes obtaining the text vector, position vector, and initial word vector of the search word, and combining the text vector, position vector, and initial word vector of the search word with The word vectors are fused to form the feature vector of the search word; the text vector, the position vector and the initial word vector of the keyword are obtained, and the text vector, the position vector and the initial word vector of the keyword are fused to form the Describe the feature vector of the content to be pushed.

Among them, the text vector (Token Embeddings) contains the preset identifiers [CLS] and [SEP], and the hidden layer vector representation corresponding to each word. For each search word or keyword, it will be set at the beginning of the sentence. The [CLS] identifier is used to represent the beginning of the sentence, and the [SEP] identifier is used as a separation between two adjacent or juxtaposed sentences, and the symbol [SEP] can be set at the end of the sentence, the initial word vector (Segment Embeddings) It is used to represent different sentences, and the position vector (Position Embeddings) is used to represent the position of each word of the sentence in the sentence.

Among them, vector fusion means that vector fusion refers to converting multiple vectors into one vector. According to different fusion methods, vector fusion may include vector splicing, vector addition, and the like. As a way, in the BERT model and the variant model of the BERT model, the position vector is obtained in the process of using the sentence for training. The text vector, position vector, and initial word vector can be vector spliced, and then go through a full connection. layer to output. By fusing a variety of vectors to obtain feature vectors, the semantics of the sentence can be better perceived in combination with the context.

S230: The feature vector of the search word is used as the input of the bidirectional coding representation network, and the bidirectional coding representation network is used to obtain the search semantic vector corresponding to the search term.

Among them, the bidirectional encoding representation network is a transformer-based bidirectional encoding representation network (BidirectionalEncoder Representation from Transformers, BERT). The BERT model uses the context at the same time, and uses a bidirectional Transformer to represent the text. The Transformer consists of an encoder and a decoder. The encoder It is mainly composed of multiple encoding modules. Each encoding module includes a self-attention layer and a feedforward neural network layer. Similar to the encoder, the decoder is also composed of multiple decoding modules. Each decoding module includes a self-attention layer and a feedforward neural network layer. Feedforward neural network layer, and an additional layer of encoding and decoding.

The BERT model is a typical two-stage model, which can be divided into a pre-training stage and a fine-tuning stage. In the pre-training stage, the Transformer is mainly used as a feature extractor to learn a large amount of unlabeled text to learn linguistic knowledge, and finally the representation of the text is obtained. is the semantic vector corresponding to the text. In the fine-tuning stage, according to the text semantic knowledge learned in the pre-training stage, the model is fine-tuned and learned based on the actual downstream business needs, so as to adapt to the actual downstream business needs. In this embodiment, the pre-training stage of the BERT model is used, and the feature vector that combines the text vector, position vector, and initial word vector of the search word is used as the input of the BERT model, and the search semantics that can be semantically represented according to the context of the sentence is obtained. vector.

As an implementation manner, user search behavior data can be obtained through the search engine application program or the buried log data of the webpage, wherein the user search behavior data includes the search term input by the user and the waiting list clicked by the user in the search result of the search term. Push content, since the content to be pushed that the user clicks can represent the user's search intent, the ALBERT model can be trained by using the search term input by the user and the content to be pushed clicked by the user in the search results of the search term as weakly supervised data. To the association between the user's search term and the content to be pushed, this association can be fully interactively matched during model training, and a better text representation can be obtained.

As an implementation manner, a variant model of the BERT model, the ALBERT model, may be used. Unlike the BERT model that uses a 12-layer Transformer, the ALBERT model uses only 4-layer Transformers. The amount of model parameters is reduced by factoring the embedding parameterization and sharing parameters across layers. The training parameters are about 4 million, and the model size is only is 14M. In addition, the ALBERT model uses Sentence Order Prediction (SOP) to replace the Next Oder Prediction (NOP) of the BERT model, which enhances the model's ability to learn the continuity of sentences and improves the ability of self-supervised learning tasks , By removing dropout, a lot of temporary variables can be saved, which can effectively improve the memory utilization during model training, improve the efficiency of the model, and reduce the scale of training data required. Compared with the BERT model, although the accuracy of the ALBERT model will be slightly reduced by 1% to 2%, the training speed and prediction speed of the model will be improved by 2 to 3 times.

Since the user's search behavior is real-time, that is, for some events, the search engine needs to meet the user's search needs within a limited time. The search semantic vector corresponding to the search word. On the other hand, using the ALBERT model can use the user search behavior data closer to the current date as the training data, so as to better meet the real-time nature of user search behavior. For example, for the same number of samples, using the BERT model requires 1 day of training time, so when training the BERT model, only the buried log data 1 day ago can be used for training, while using the ALBERT model only requires 2 hours of training time. The buried log data 2 hours ago can be obtained for training, and the training data closer to the current time can better reflect the user's time-sensitive search intent.

S240: The feature vector of the content to be pushed is used as the input of the bidirectional coding representation network, and the bidirectional coding representation network is used to obtain the content semantic vector corresponding to the keyword.

It can be understood that the method for obtaining the content semantic vector corresponding to the keyword is the same as the method for obtaining the search semantic vector corresponding to the search word. For details, please refer to step S230.

S250: Calculate the similarity between the search semantic vector and the content semantic vector, and determine the target content semantic vector from the content semantic vector according to the similarity.

S260: Acquire the content to be pushed corresponding to the semantic vector of the target content according to the semantic vector of the target content, and push the content.

It should be noted that, for parts not described in detail in this embodiment, reference may be made to the foregoing embodiments, and details are not described herein again.

In the method for pushing information based on user intent provided by the embodiment of the present application, the semantic understanding model is a transformer-based bidirectional coding representation network BERT model, and after obtaining the search term input by the user and the set of mapping relationships of the content to be pushed, the set of mapping relationships of the content to be pushed is obtained. After including the mapping relationship between the content to be pushed and the keywords, the feature vector of the search word and the feature vector of the keyword are obtained respectively, and the feature vector of the search word is used as the input of the two-way encoding to represent the network, and the network is represented by the two-way encoding, and the search is obtained. The search semantic vector corresponding to the word, the feature vector of the keyword is used as the input of the two-way encoding to represent the network, and the network is represented by the two-way encoding, and the content semantic vector corresponding to the keyword is obtained. By calculating the similarity between the search semantic vector and the content semantic vector, according to The similarity determines the target content semantic vector from the content semantic vector, and then obtains the content to be pushed corresponding to the target content semantic vector according to the target content semantic vector, and pushes it. Through the BERT model, the semantic vectors corresponding to the search words and keywords, which are semantically represented according to the context of the sentence, can be obtained respectively, so that the user's search intent can be obtained according to the semantic vector and pushed, without the need for manual annotation and training expectations. On the one hand, it saves The manpower for labeling, on the other hand, also avoids the problem of inaccurate push caused by too thick label dimensions.

Please refer to FIG. 4. FIG. 4 shows a schematic flowchart of a user-intent-based push method provided by another embodiment of the present application, which is applied to the server in the above-mentioned system, that is, the execution body of the method may be the above-mentioned server. For improving the accuracy of the content pushed for the user, specifically, as shown in FIG. 4 , the method includes: S310 to S360.

S310: Obtain a set of mapping relationships between the search term input by the user and the content to be pushed.

S320: Based on the pre-trained semantic understanding model, obtain the search semantic vector corresponding to the search word and the content semantic vector corresponding to the keyword, respectively.

S330: Calculate the vector distance between the search semantic vector and the content semantic vector.

Among them, the vector distance can be Euclidean Distance, Manhattan Distance, Chebyshev Distance, Standardized Euclidean Distance, Cosine of Included Angle, etc. limited.

As an embodiment, the data of the search semantic vector and the content semantic vector can be stored in the offline server, and the information push device in the online server can obtain the data of the offline server's search semantic vector and content semantic vector through the network, and calculate the search semantic vector The vector distance between the vector and the content semantic vector.

In some embodiments, the vector distance may be the cosine distance of the included angle between the search semantic vector and the content semantic vector. Specifically, the vector length of the search semantic vector and the content semantic vector may be calculated to obtain the vector length of the search semantic and the content semantic vector. The vector length of semantics; calculate the vector inner product of the search semantic vector and the content semantic vector; calculate the cosine distance between the search semantic vector and the content semantic vector based on the vector inner product and the vector length, as the difference between the search semantic vector and the content semantic vector. vector distance.

S340: Determine the similarity between the search semantic vector and the content semantic vector according to the vector distance.

The similarity between the search semantic vector and the content semantic vector can be determined according to the vector distance. The greater the vector distance, the greater the difference between the search semantic vector and the content semantic vector, and accordingly, the lower the similarity between the search semantic vector and the content semantic vector.

As a way, the normalized vector distance can be used as the similarity between the search semantic vector and the content semantic vector. As another way, a pre-trained semantic understanding model can be integrated with a loss function that calculates the similarity according to the vector distance, and the vector distance can be converted into a vector similarity expressed in probabilistic form through the softmax function.

S350: Use the content semantic vector whose similarity satisfies the specified condition as the target content semantic vector.

The specified condition may be set according to a policy of pushing information, and the target semantic vector may be one vector or multiple vectors. In one way, the specified condition may be greater than a specific threshold, and accordingly, a content semantic vector corresponding to greater than a specific threshold is used as the target semantic vector. The specific condition may also be to obtain N content semantic vectors with the largest similarity, where N is a preset integer greater than 0, and correspondingly, the N content semantic vectors with the largest similarity are used as the target semantic vector.

S360: Acquire the content to be pushed corresponding to the semantic vector of the target content according to the semantic vector of the target content, and push the content.

It should be noted that, for parts not described in detail in this embodiment, reference may be made to the foregoing embodiments, and details are not described herein again.

The method for pushing information based on user intent provided by the embodiment of the present application acquires a set of mapping relationships between a search term input by a user and content to be pushed, and the set of mapping relationships for content to be pushed includes a mapping relationship between content to be pushed and keywords, and then based on a pre- The trained semantic understanding model obtains the search semantic vector corresponding to the search word and the content semantic vector corresponding to the keyword, calculates the vector distance between the search semantic vector and the content semantic vector, and determines the search semantic vector and the content semantic vector according to the vector distance. and the content semantic vector whose similarity satisfies the specified condition is used as the target content semantic vector, so as to obtain the content to be pushed corresponding to the target content semantic vector according to the target content semantic vector, and push it. The calculation of the vector distance is simple and easy to implement, and the similarity between the search semantic vector and the content semantic vector can be obtained quickly and in real time, so as to efficiently push content.

Please refer to FIG. 5. FIG. 5 shows a schematic flowchart of a user-intent-based push method provided by another embodiment of the present application, which is applied to the server in the above-mentioned system, that is, the execution body of the method may be the above-mentioned server. For improving the accuracy of the content pushed for the user, specifically, as shown in FIG. 5 , the method includes: S410 to S450.

S410: Acquire page text information of the page browsed by the user.

The user browsing page may be the text information of the page browsed in the search result after the user conducts a search. As a method, the user browsing page may also be in the scenario of information flow, the user is mainly for the purpose of browsing, and there is no The browsing page generated by the explicit retrieval behavior is not limited here. By acquiring the page text information of the page the user browses, the mapping relationship between the user and the page text information can be obtained.

As a method, the page text information of the user's browsing page can be obtained through the search history behavior data generated when the user searches. When browsing a page, the page viewed by each click can reflect the user's browsing tendency, that is to say, the text information of each page browsed by the user can reflect the user's intention information. For example, most of the pages that users browse are related to maternal and infant forums. It can be speculated that the user may be a novice parent who is concerned about the field of infants and young children, so the user may be more interested in the pushed content of infant and young children products. For another example, if it is obtained through analysis of the pages browsed by the user that the text content of the page has fewer words, the user may be more inclined to push short text content.

S420: Obtain a page semantic vector corresponding to the page text information based on the pre-trained semantic understanding model.

The pre-trained semantic understanding model may be a BERT model, an ALBERT model, or other machine learning models, which are not limited here. The page semantic vector corresponding to the page text information can be obtained by inputting the page enben information into the pre-trained semantic understanding model, wherein the page semantic vector is a vector used to represent the semantic content of the page text information of the page browsed by the user, It reflects the intent information of users' browsing behavior after searching.

S430: Calculate the cosine distance between the search semantic vector and the content semantic vector to obtain a first similarity.

The first similarity is used to represent the similarity between the search semantic vector and the content semantic vector, and the cosine distance between the search semantic vector and the content semantic vector is specifically calculated. For the process of obtaining the first similarity, please refer to steps S330 to S340.

S440: Calculate the cosine distance between the page semantic vector and the content semantic vector to obtain a second similarity.

The second similarity is used to represent the similarity between the page semantic vector and the content semantic vector. It can be understood that the process of calculating the cosine distance between the search semantic vector and the content semantic vector to obtain the second similarity is similar to step S430. Please refer to steps S330 to S340.

S450: Determine the semantic vector of the target content according to the first similarity and the second similarity.

As an implementation manner, different weight values may be assigned to the first similarity and the second similarity corresponding to the content semantic vector respectively, and the weighted sum of the first similarity and the second similarity may be used as the comprehensive similarity, and the The content semantic vector corresponding to the comprehensive similarity satisfying the specified condition is taken as the target content semantic vector.

As another implementation manner, the target content semantic vector may be specified, the search semantic vector corresponding to the first similarity that satisfies the first specified condition is obtained, and then the first user group corresponding to the search semantic vector is obtained; the second specified condition is obtained. The content semantic vector corresponding to the second similarity degree of , and then the second user group corresponding to the content semantic vector is obtained. By taking the intersection of the first user group and the second user group, the target user for information push is obtained, and the target content semantic vector is pushed to the target user.

It should be noted that, for parts not described in detail in this embodiment, reference may be made to the foregoing embodiments, and details are not described herein again.

The user-intent-based information push method provided by the embodiment of the present application acquires the search term input by the user, and based on the pre-trained semantic understanding model, obtains the search semantic vector corresponding to the search term and the content semantic vector corresponding to the keyword respectively, and can also use Obtain the page text information of the page the user browses, obtain the page semantic vector corresponding to the page text information based on the pre-trained semantic understanding model, calculate the cosine distance between the search semantic vector and the content semantic vector, obtain the first similarity, and calculate the page semantic vector and content. The cosine distance of the semantic vector is used to obtain the second similarity, and the semantic vector of the target content is determined according to the first similarity and the second similarity, so that the content to be pushed corresponding to the semantic vector of the target content is pushed. By obtaining the page text information of the user's browsing page, the user's intention can be understood from more dimensions, thereby realizing a more effective information push method.

Please refer to FIG. 6 , which shows a structural block diagram of a user-intent-based push device 600 provided by an embodiment of the present application. The device is applied to the server in the above system. The device may include: an information acquisition module 610 , a vector acquisition module 600 module 620 , determination module 630 and processing module 640 .

The information acquisition module 610 acquires the search term input by the user and a set of mapping relationships of the content to be pushed, where the set of mapping relationships of the content to be pushed includes: the mapping relationship between the content to be pushed and the keywords.

Further, the information acquisition module 610 includes: a search data acquisition sub-module, a data storage sub-module and a video display sub-module, wherein:

The search data acquisition sub-module is configured to acquire user historical search data, where the user historical search data includes historical search terms of the user and the content to be pushed browsed by the user corresponding to the historical search terms.

A data storage submodule, configured to store the mapping relationship between the content to be pushed and the title of the content to be pushed in the content mapping relationship set to be pushed.

The vector obtaining module 620, based on the pre-trained semantic understanding model, obtains the search semantic vector corresponding to the search term and the content semantic vector corresponding to the keyword, respectively.

Further, the vector acquisition model 620 includes: a feature vector acquisition sub-module, a search semantic vector acquisition sub-module and a content semantic vector acquisition sub-module, wherein:

The feature vector obtaining submodule is used to obtain the feature vector of the search word and the feature vector of the keyword respectively.

Further, the feature vector acquisition sub-module includes: a search word vector fusion unit and a keyword vector fusion unit, wherein:

The search word vector fusion unit is used to obtain the text vector, position vector and initial word vector of the search word, and fuse the text vector, position vector and initial word vector of the search word to form the feature vector of the search word .

And a keyword vector fusion unit, used to obtain the text vector, position vector, and initial word vector of the keyword, and fuse the text vector, position vector, and initial word vector of the keyword to form the content to be pushed. Feature vector.

The search semantic vector acquisition sub-module is configured to use the feature vector of the search word as the input of the bidirectional coding representation network, and obtain the search semantic vector corresponding to the search term through the bidirectional coding representation network.

The content semantic vector acquisition sub-module is configured to use the feature vector of the keyword as the input of the bidirectional coding representation network, and obtain the content semantic vector corresponding to the keyword through the bidirectional coding representation network.

The determining module 630 calculates the similarity between the search semantic vector and the content semantic vector, and determines a target content semantic vector from the content semantic vector according to the similarity.

Further, the determination module 630 includes: a distance calculation sub-module, a similarity determination sub-module and a target vector determination sub-module, wherein:

A distance calculation submodule, configured to calculate a vector distance between the search semantic vector and the content semantic vector.

Further, the distance calculation submodule includes: a length calculation unit, an inner product calculation unit and a cosine distance calculation unit, wherein:

A length calculation unit, configured to calculate the vector length of the search semantic vector and the content semantic vector.

An inner product calculation unit, configured to calculate the vector inner product of the search semantic vector and the content semantic vector.

A cosine distance calculation unit, configured to calculate the cosine distance between the search semantic vector and the content semantic vector based on the vector inner product and the vector length, as the distance between the search semantic vector and the content semantic vector vector distance.

The similarity determination sub-module is configured to determine the similarity between the search semantic vector and the content semantic vector according to the vector distance.

The target vector determination sub-module uses the content semantic vector whose similarity satisfies a specified condition as the target content semantic vector.

The processing module 640 obtains the content to be pushed corresponding to the semantic vector of the target content according to the semantic vector of the target content, and pushes it.

Further, the device may further include: a text information acquisition module, a page semantic acquisition module, a first calculation module, a second calculation module, and a comprehensive determination module, wherein:

The text information acquisition module is used to acquire the page text information of the page browsed by the user.

The page semantic acquisition module is configured to acquire the page semantic vector corresponding to the page text information based on the pre-trained semantic understanding model.

The first calculation module is configured to calculate the cosine distance between the search semantic vector and the content semantic vector to obtain a first similarity.

The second calculation module is configured to calculate the cosine distance between the page semantic vector and the content semantic vector to obtain a second similarity.

A comprehensive determination module is configured to determine the target content semantic vector according to the first similarity and the second similarity.

Those skilled in the art can clearly understand that, for the convenience and brevity of description, for the specific working process of the above-described devices and modules, reference may be made to the corresponding processes in the foregoing method embodiments, which will not be repeated here.

In several embodiments provided in this application, the coupling between the modules may be electrical, mechanical or other forms of coupling.

In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically alone, or two or more modules may be integrated into one module. The above-mentioned integrated modules can be implemented in the form of hardware, and can also be implemented in the form of software function modules.

Please refer to FIG. 7 , which shows a structural block diagram of an electronic device provided by an embodiment of the present application. The electronic device may be the above-mentioned server 100 . The electronic device 700 in the present application may include one or more of the following components: a processor 710, a memory 720, and one or more application programs, wherein the one or more application programs may be stored in the memory 720 and configured by One or more processors 710 execute, and one or more programs are configured to perform the methods described in the foregoing method embodiments.

Processor 710 may include one or more processing cores. The processor 710 uses various interfaces and lines to connect various parts of the entire electronic device 700, and executes by running or executing the instructions, programs, code sets or instruction sets stored in the memory 720, and calling the data stored in the memory 120. Various functions of the electronic device 100 and processing data. Optionally, the processor 710 may employ at least one of a digital signal processing (Digital Signal Processing, DSP), a Field-Programmable Gate Array (Field-Programmable Gate Array, FPGA), and a Programmable Logic Array (Programmable Logic Array, PLA). implemented in hardware. The processor 710 may integrate one or a combination of a central processing unit (Central Processing Unit, CPU), a graphics processing unit (Graphics Processing Unit, GPU), a modem, and the like. Among them, the CPU mainly handles the operating system, user interface and application programs, etc.; the GPU is used for rendering and drawing of the display content; the modem is used to handle wireless communication. It can be understood that, the above-mentioned modem may not be integrated into the processor 710, and is implemented by a communication chip alone.

The memory 720 may include random access memory (Random Access Memory, RAM), or may include read-only memory (Read-Only Memory). Memory 720 may be used to store instructions, programs, codes, sets of codes, or sets of instructions. The memory 720 may include a stored program area and a stored data area, wherein the stored program area may store instructions for implementing an operating system, instructions for implementing at least one function (such as a touch function, a sound playback function, an image playback function, etc.) , instructions for implementing the following method embodiments, and the like. The storage data area may also store data created by the terminal 100 during use (such as phone book, audio and video data, chat record data) and the like.

Please refer to FIG. 8 , which shows a structural block diagram of a computer-readable storage medium provided by an embodiment of the present application. The computer-readable medium 800 stores program codes, and the program codes can be invoked by the processor to execute the methods described in the above method embodiments.

The computer readable storage medium 800 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read Only Memory), EPROM, hard disk, or ROM. Optionally, the computer-readable storage medium 800 includes a non-transitory computer-readable storage medium. Computer readable storage medium 800 has storage space for program code 810 to perform any of the method steps in the above-described methods. These program codes can be read from or written to one or more computer program products. Program code 810 may be compressed, for example, in a suitable form.

In summary, the user-intent-based push method, device, electronic device and computer-readable medium provided by the present application acquire the search term input by the user and the set of mapping relationships of the content to be pushed, and the set of mapping relationships of the content to be pushed includes: The mapping relationship between the content to be pushed and the keywords; based on the pre-trained semantic understanding model, respectively obtain the search semantic vector corresponding to the search word and the content semantic vector corresponding to the keyword; calculate the search semantic vector and all According to the similarity of the content semantic vector, the target content semantic vector is determined from the content semantic vector according to the similarity; according to the target content semantic vector, the content to be pushed corresponding to the target content semantic vector is obtained and pushed. Therefore, the user's search intent can be mined according to the user's search terms, so as to effectively push content.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand: it can still be Modifications are made to the technical solutions described in the foregoing embodiments, or some technical features thereof are equivalently replaced; and these modifications or replacements do not drive the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims (10)

1. A pushing method based on user intention is characterized by comprising the following steps:

acquiring a search word input by a user and a mapping relation set of contents to be pushed, wherein the mapping relation set of the contents to be pushed comprises: mapping relation between the content to be pushed and the keywords;

respectively acquiring a search semantic vector corresponding to the search word and a content semantic vector corresponding to the keyword based on a pre-trained semantic understanding model;

calculating the similarity of the search semantic vector and the content semantic vector, and determining a target content semantic vector from the content semantic vector according to the similarity;

and acquiring the content to be pushed corresponding to the target content semantic vector according to the target content semantic vector, and pushing.

2. The method according to claim 1, wherein the semantic understanding model is a bidirectional coding representation network based on a converter, and the obtaining of the search semantic vector corresponding to the search word and the content semantic vector corresponding to the keyword respectively based on the pre-trained semantic understanding model comprises:

respectively obtaining the feature vector of the search word and the feature vector of the keyword;

taking the feature vector of the search word as the input of the bidirectional coding representation network, and obtaining a search semantic vector corresponding to the search word through the bidirectional coding representation network;

and taking the feature vector of the keyword as the input of the bidirectional coding representation network, and obtaining a content semantic vector corresponding to the keyword through the bidirectional coding representation network.

3. The method according to claim 2, wherein the obtaining the feature vector of the search term and the feature vector of the keyword respectively comprises:

acquiring a text vector, a position vector and an initial word vector of the search word, and fusing the text vector, the position vector and the initial word vector of the search word to form a feature vector of the search word;

and acquiring a text vector, a position vector and an initial word vector of the keyword, and fusing the text vector, the position vector and the initial word vector of the keyword to form a feature vector of the content to be pushed.

4. The method of claim 1, wherein the calculating a similarity between the search semantic vector and the content semantic vector, and determining a target content semantic vector from the content semantic vectors according to the similarity comprises:

calculating a vector distance between the search semantic vector and the content semantic vector;

determining the similarity of the search semantic vector and the content semantic vector according to the vector distance;

and taking the content semantic vector with the similarity meeting a specified condition as the target content semantic vector.

5. The method of claim 4, wherein the calculating a vector distance between the search semantic vector and the content semantic vector comprises:

calculating the vector length of the search semantic vector and the content semantic vector;

calculating a vector inner product of the search semantic vector and the content semantic vector;

calculating a cosine distance between the search semantic vector and the content semantic vector based on the vector inner product and the vector length as a vector distance between the search semantic vector and the content semantic vector.

6. The method of claim 1, further comprising:

acquiring page text information of a user browsing page;

acquiring a page semantic vector corresponding to the page text information based on a pre-trained semantic understanding model;

calculating the cosine distance between the search semantic vector and the content semantic vector to obtain a first similarity;

calculating the cosine distance between the page semantic vector and the content semantic vector to obtain a second similarity;

and determining the semantic vector of the target content according to the first similarity and the second similarity.

7. The method according to any one of claims 1 to 6, wherein the keywords of the content to be pushed are keywords included in a title of the content to be pushed, and the obtaining of the search term input by the user and the mapping relationship set of the content to be pushed comprises:

acquiring historical search data of a user, wherein the historical search data of the user comprises historical search words of the user and the content to be pushed browsed by the user corresponding to the historical search words;

and storing the mapping relation between the content to be pushed and the title of the content to be pushed in the mapping relation set of the content to be pushed.

8. A push device based on user intent, comprising:

the information acquisition module is used for acquiring a search word input by a user and a mapping relation set of contents to be pushed, wherein the mapping relation set of the contents to be pushed comprises: mapping relation between the content to be pushed and the keywords;

the vector acquisition module is used for respectively acquiring a search semantic vector corresponding to the search word and a content semantic vector corresponding to the keyword based on a pre-trained semantic understanding model;

the determining module is used for calculating the similarity between the search semantic vector and the content semantic vector and determining a target content semantic vector from the content semantic vector according to the similarity;

and the processing module is used for acquiring the content to be pushed corresponding to the target content semantic vector according to the target content semantic vector and pushing the content.

9. An electronic device, comprising:

one or more processors;

a memory;

one or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications configured to perform the method of any of claims 1-7.

10. A computer-readable medium having stored program code executable by a processor, the program code causing the processor to perform the method of any one of claims 1-7 when executed by the processor.

Images