CN114329199A - Material recall method and device - Google Patents
Material recall method and device Download PDFInfo
- Publication number
- CN114329199A CN114329199A CN202111612538.4A CN202111612538A CN114329199A CN 114329199 A CN114329199 A CN 114329199A CN 202111612538 A CN202111612538 A CN 202111612538A CN 114329199 A CN114329199 A CN 114329199A
- Authority
- CN
- China
- Prior art keywords
- user
- search
- time period
- materials
- intention
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Abstract
A method and a device for recalling materials are provided, wherein a plurality of time periods with different durations are set, and the time period with the longer duration comprises the time period with the longer duration; for each time period, calculating user search portrait characteristics of the time period based on search information of the user and a user portrait of the user in the time period; according to the length sequence of the time periods with the duration, sequentially fusing and calculating the user search portrait characteristics of the time periods to obtain the search intention input characteristics of the user, and fusing and calculating the correlation of the user search portrait characteristics based on the time periods to the search intention input characteristics of the user; inputting the search intention input characteristics of the user into an intention classification model to obtain the probability values of the search intentions of the user on different types of materials; and recalling the materials from the alternative material set according to the materials with the probability values of the search intentions of the users to the materials of different types, thereby accurately determining the search intentions of the users and recalling the materials.
Description
Technical Field
The present application relates to the field of computing network technologies, and in particular, to a method and an apparatus for material recall.
Background
With the development of computer networks, more and more network platforms are loaded in the computer networks to provide various services for users. At present, the most direct interaction mode between a network platform and a user is to search in a set of alternative materials by a search engine set by the network platform according to search information sent by the user, and distribute the materials according with the search information to the user. The core problem of pushing materials for a user by a network platform is how to accurately understand the search intention of the user, and the materials can be quickly and accurately distributed to the user only on the basis of fully understanding the search intention of the user.
With the proposal and development of the technical concept of user portrait, the user portrait plays an important role in understanding the search intention of the user, and the search intention of the user can be determined based on the combination of the user portrait and the search information of the user. Particularly for a vertical domain network platform, different users may have different search intentions for the same search information, and the search intentions of the same user in different time periods may also be different, so that the determination of the search intention of the user by combining the search information of the user with the user representation is a currently common technical route.
However, in the stage of determining the search intention of the user in the material recalling process, the user portrait based on the user portrait changes along with the time, and the determined search intention of the user is inaccurate only by combining the accumulated information of the set time period of the user portrait with the search information of the user, and the recalled material is also inaccurate based on the search intention of the inaccurate user.
Disclosure of Invention
In view of this, embodiments of the present disclosure provide a method and an apparatus for recalling a material, which can combine user images at different time periods with search information of a user to accurately determine a search intention of the user and recall the material.
The embodiment of the application is realized as follows:
a method of material recall, the method comprising:
setting a plurality of time periods with different durations, wherein the long time period comprises a long time period;
for each time period, calculating user search portrait characteristics of the time period based on search information of a user and a user portrait of the user in the time period;
according to the length sequence of the time periods with the duration, sequentially carrying out fusion calculation on the user search portrait features of the time periods to obtain search intention input features of the user, wherein the fusion calculation is carried out on the basis of the relevance of the user search portrait features of the time periods to the search intention input features of the user;
inputting the search intention input characteristics of the user into an intention classification model to obtain the probability values of the search intentions of the user to different types of materials;
and recalling the materials for the user from the alternative material set according to the materials with the probability values of the search intentions of the user on the different types of materials.
Preferably, the setting of the plurality of time periods with different durations is three time periods, including:
the time-sharing system comprises a first time period, a second time period and a third time period, wherein the duration of the first time period is greater than that of the second time period, and the second time period is greater than that of the third time period.
Preferably, said calculating user search representation characteristics for said time period comprises:
and inputting the user search information and the user image of the time period into a set feed-forward neural network, and calculating to obtain the user search image characteristics of the time period.
Preferably, the sequentially performing fusion calculation on the user search portrait features of the time period to obtain the search intention input features of the user includes:
a. selecting a current time period according to the time period length sequence, multiplying the search information and the splicing vector of the user portrait of the current time period by a set first weight vector, and calculating by adopting an activation function to obtain a first gating value;
b. multiplying the search information and the splicing vector of the user portrait in the current time period by a set second weight vector, and calculating by adopting an activation function to obtain a second gating value;
c. adding the product of the residual probability value of the first gating value and the user search portrait characteristics of the current time period and the product of the second gating value and the search intention input characteristics of the time period in the sequence of the current time period to obtain a sum value, and taking the sum value as the search intention input characteristics of the current time period;
d. and b, determining whether the current time period is the time period with the shortest time length in the set time period, if so, taking the search intention input characteristic of the current time period as the search intention input characteristic of the user, if not, taking the next time period in the sequence of the current time period as the current time period, and returning to the step a to continue executing.
Preferably, inputting the search intention input features of the user into an intention classification model, and obtaining the probability values of the search intention of the user on different types of materials includes:
extracting search information characteristics of the search information of the user by adopting a neural network to obtain the search information characteristics of the user;
inputting the search information characteristics of the user and the search intention input characteristics of the user into a first intention classification model, and outputting to obtain the probability values of the search intentions of the user to different types of materials.
Preferably, before obtaining the probability value of the search intention of the user for different types of materials, the method further comprises:
and filtering the search intention probability values of the different types of materials output by the intention classification model according to a set filtering rule to obtain the search intention probability values of the different types of materials of the user.
Preferably, the method further comprises:
recalling materials for the user from the set of alternative materials as first recalled materials;
extracting search information characteristics of the search information of the user by adopting a neural network to obtain the search information characteristics of the user, analyzing and searching from an alternative material set by adopting a distributed search analysis engine according to the search information characteristics of the user, and recalling the material aiming at the user as a second recalled material;
searching from an alternative material set by adopting a set search rule according to the search information of the user, and recalling the material for the user as a third recalled material;
fusing the first recall material, second recall material, and third recall material;
and sequencing the fused recall materials according to the second search intention probability value of the corresponding material type, and distributing the sequenced recall materials to the user.
Preferably, the second search intention probability value of the corresponding material type includes:
extracting search information characteristics of the search information of the user by adopting a neural network to obtain the search information characteristics of the user;
inputting the search information characteristics of the user and the search intention input characteristics of the user into a second intention classification model formed by a long-short term memory artificial neural network (LSTM) and an attention mechanism attention, and outputting to obtain a second search intention probability value of the user on different types of materials.
A computer product, comprising an electronic device, comprising:
a processor;
a memory storing a program configured to implement the method of any of the above recalls when executed by the processor.
A non-transitory computer readable storage medium storing instructions that, when executed by a processor, cause the processor to perform the steps of the method of material recall of any of the above.
As seen from the above, the present embodiment sets a plurality of time periods having different durations, wherein a long time period includes a long time period; for each time period, calculating user search portrait characteristics of the time period based on search information of a user and a user portrait of the user in the time period; according to the length sequence of the time periods with the duration, sequentially carrying out fusion calculation on the user search portrait features of the time periods to obtain search intention input features of the user, wherein the fusion calculation is carried out on the basis of the relevance of the user search portrait features of the time periods to the search intention input features of the user; inputting the search intention input characteristics of the user into an intention classification model to obtain the probability values of the search intentions of the user to different types of materials; and recalling the materials for the user from the alternative material set according to the materials with the probability values of the search intentions of the user on the different types of materials. In this way, at the stage of determining the search intention of the user in the material recalling process, the search intention of the user is determined not based on the user portrait in one time period but based on the user portraits in a plurality of time periods, so that the fine-grained user portrait can be obtained from the time dimension, the search intention of the user is fully understood based on the fine-grained user portrait, the determined search intention of the user is accurate, and the material is accurately recalled for the user based on the accurate search intention of the user.
Drawings
FIG. 1 is a flow chart of a method for distributing materials to a user according to the present embodiment;
FIG. 2 is a flow chart of a method for material recall according to an embodiment of the present application;
FIG. 3 is a block diagram of an implementation process for determining a search intention of a user according to an embodiment of the present disclosure;
FIG. 4 is a process diagram of an example of a search intent recognition phase of a user according to an embodiment of the present application;
FIG. 5 is a diagram illustrating an example process of a distribution sorting phase according to an embodiment of the present application;
fig. 6 is a schematic structural diagram of a material recall device according to an embodiment of the present application;
fig. 7 is a schematic diagram of an electronic device according to another embodiment of the present application.
Detailed Description
The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.
The terms "first," "second," "third," "fourth," and the like in the description and in the claims of the present application and in the drawings described above, if any, are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the application described herein are, for example, capable of operation in sequences other than those illustrated or otherwise described herein. Furthermore, the terms "comprising" and "having," as well as any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements explicitly listed, but may include other steps or elements not explicitly listed or inherent to such process, method, article, or apparatus.
The technical solution of the present application will be described in detail with specific examples. Several of the following embodiments may be combined with each other and some details of the same or similar concepts or processes may not be repeated in some embodiments.
It can be seen from the background art that, in the stage of determining the search intention of the user in the material recalling process, the determined search intention of the user is inaccurate only by combining the accumulated information of the set time period of the user image with the search intention of the user, so that the finally recalled materials for the user are inaccurate. Particularly, for a vertical domain network platform such as a house network platform, the network platform is a low-frequency use network platform for users, the user image, particularly the behavior characteristics of the user image are very sparse, after all information of the historical time period of the user image is accumulated, the user cannot fully understand the behavior of the user by using the whole image, and the behavior of the user changes in stages along with the time. Therefore, it is difficult to capture an accurate user search intention simply based on the accumulated information of the user figure in the history period.
Fig. 1 is a flowchart of a method for distributing materials to users in the present embodiment, which includes the following specific steps:
102, acquiring search information of a user;
103, splicing the user portrait with the search information to obtain the search splicing information characteristic of the user;
104, inputting the search splicing information characteristics of the user into an intention classification model, and outputting to obtain a search intention probability value of the user;
and 105, recalling the materials for the user from the alternative material set based on the search intention probability value of the user, distributing and sorting the materials, and pushing the materials to the user.
In the embodiment of the application, the user portrait mainly includes user behavior information and user tags, where the user behavior information includes information such as browsing or/and clicking behaviors of the user, and the user tags include material tags browsed by the user, for example, corresponding user tags are set for house sources browsed by the user.
As can be seen from fig. 1, after simply splicing the user image of the user in the historical time period with the search information of the user, the probability value of the search intention of the user is determined accordingly. For network platforms in the vertical domain, such simple splicing cannot fully consider the change of the user portrait in the time dimension. In fact, the user's user representation changes over time, for example, on a network platform in the housing area, the user searches for a house source in the same cell, and over time, the user's search intention is no longer just to buy the house source, but rather to decorate the house source, rent the house source, or sell a different search intention of the house source. Therefore, the search intention of the user changes with the lapse of time, and therefore, it is necessary to sufficiently take into account the change in the user figure in the time dimension.
As can be seen from the above analysis, although the user image of the user is introduced to understand the user's search intention when determining the user's search intention, the user image is not hierarchically understood by layering the user image in the time dimension, and the user image is not depicted in the time dimension on the basis of points in the understanding layer of the user's search intention. The fusion of user images and search information of a user in different time periods is not fully considered, understanding of search intentions of the user is different, and an effective multi-level user image fusion strategy is lacked.
In order to solve the above problem, in the embodiments of the present application, a plurality of time periods having different durations are set, where a time period having a long duration includes a time period having a long duration; for each time period, calculating user search portrait characteristics of the time period based on search information of a user and a user portrait of the user in the time period; according to the length sequence of the time periods with the duration, sequentially carrying out fusion calculation on the user search portrait features of the time periods to obtain search intention input features of the user, wherein the fusion calculation is carried out on the basis of the relevance of the user search portrait features of the time periods to the search intention input features of the user; inputting the search intention input characteristics of the user into an intention classification model to obtain the probability values of the search intentions of the user to different types of materials; and recalling the materials for the user from the alternative material set according to the materials with the probability values of the search intentions of the user on the different types of materials.
In this way, at the stage of determining the search intention of the user in the material recalling process, the search intention of the user is determined not based on the user portrait in one time period but based on the user portraits in a plurality of time periods, so that the fine-grained user portrait can be obtained from the time dimension, the search intention of the user is fully understood based on the fine-grained user portrait, the determined search intention of the user is accurate, and the material is accurately recalled for the user based on the accurate search intention of the user.
Furthermore, after the materials are accurately recalled for the users, the recalled materials of the users can be accurately sorted and distributed, so that the pushed materials are different by different users or the same user in different time periods aiming at the same search information. Therefore, the search intention of the user can be understood in a fine-grained and multi-level mode on the time dimension, and the search experience of the user is improved.
The set time periods with different durations are a long-term time period, a medium-term time period and a short-term time period, wherein the long-term time period can be set to be half a year, the medium-term time period can be set to be three months, the short-term time period can be set to be one week, and the time periods can be adjusted according to different applied network platforms during setting. Because the user portrait of different time periods can be fused, the user portrait is depicted in a fine granularity on a time dimension, and the search intention change of the user is fully understood.
In the embodiment of the application, when the user search portrait features of the time periods are sequentially subjected to fusion calculation according to the length sequence of the time periods with the duration, the set door structure is adopted, so that the fusion rate of the user search portrait features of different time periods is controlled, the rate of the user search portrait features of different time periods in the whole user search intention input feature is weakened or strengthened, and the current user search intention is fully understood.
Fig. 2 is a flowchart of a method for recalling a material according to an embodiment of the present application, which includes the following specific steps:
step 201, setting a plurality of time periods with different durations, wherein the time period with long duration comprises the time period with long duration;
step 203, sequentially performing fusion calculation on the user search portrait features of the time periods according to the length sequence of the time periods with duration, until all the user search portrait features of the time periods are fused, so as to obtain search intention input features of the user, wherein the fusion calculation is performed on the basis of the relevance of the user search portrait features of the time periods to the search intention input features of the user;
step 204, inputting the search intention input characteristics of the user into an intention classification model to obtain the probability values of the search intentions of the user to different types of materials;
and step 205, recalling the materials for the user from the alternative material set according to the materials with the probability values of the search intentions of the user to the materials of different types.
In the above method, the setting the plurality of periods having different durations to three periods includes:
the time-sharing system comprises a first time period, a second time period and a third time period, wherein the duration of the first time period is greater than that of the second time period, and the second time period is greater than that of the third time period. Specifically, the first time period is a long-term time period, the second time period is a medium-term time period, and the third time period is a short-term time period, where the long-term time period may be set to half a year, the medium-term time period may be set to three months, and the short-term time period may be set to one week, and may be adjusted according to different network platforms applied during setting.
Therefore, the user portrait in different time periods is obtained based on the time periods with different durations, and then subsequent fusion is carried out, so that the user portrait is depicted in a fine granularity on a time dimension, and the change of the search intention of the user is fully understood.
In the above method, the calculating the user search profile characteristic of the time period includes:
and inputting the user search information and the user image of the time period into a set feed-forward neural network, and calculating to obtain the user search image characteristics of the time period. Here, a feed-forward neural network is used to achieve fusion of the user search information and the user representation of the time segment within the time segment.
In the method, sequentially performing fusion calculation on the user search portrait features in the time period to obtain the search intention input features of the user comprises the following steps:
a. selecting a current time period according to the time period length sequence, multiplying the search information and the splicing vector of the user portrait of the current time period by a first weight vector, and calculating by adopting an activation function to obtain a first gating value;
b. multiplying the search information and the splicing vector of the user portrait in the current time period by a second weight vector, and calculating by adopting an activation function to obtain a second gating value;
c. adding the product of the residual probability value of the first gating value and the user search portrait characteristics of the current time period and the product of the second gating value and the search intention input characteristics of the time period in the sequence of the current time period to obtain a sum value, and taking the sum value as the search intention input characteristics of the current time period;
d. and b, determining whether the current time period is the time period with the shortest time length in the set time period, if so, taking the search intention input characteristic of the current time period as the search intention input characteristic of the user, if not, taking the next time period in the sequence of the current time period as the current time period, and returning to the step a to continue executing.
It can be seen that when the user search portrait features of the time period are sequentially subjected to fusion calculation, the user search portrait features of the time period with long duration are fused firstly, then the user search portrait features with long duration are fused, and during the fusion, according to the setting of the first gating value and the setting of the second gating value of the current time period, the correlation degree of the user search portrait features of the current time period to the finally obtained search intention input features and the correlation degree of the search intention input features of the previous time period to the finally obtained search intention input features of the user are respectively adjusted. Generally, the first gating value of the current time period is smaller than the second gating value, so that the degree of correlation between the search intention input features of the long and short time periods and the finally obtained search intention input features is higher than that of the search intention features of the long and long time periods.
In the above method, inputting the search intention input features of the user into an intention classification model, and obtaining the probability values of the search intention of the user for different types of materials includes:
extracting search information characteristics of the search information of the user by adopting a neural network to obtain the search information characteristics of the user;
inputting the search information characteristics of the user and the search intention input characteristics of the user into an intention classification model, and outputting to obtain the probability values of the search intention of the user on different types of materials.
In the above method, before obtaining the probability value of the search intention of the user for different types of materials, the method further includes:
and filtering the search intention probability values of the different types of materials output by the intention classification model according to a set filtering rule to obtain the search intention probability values of the different types of materials of the user. Here, the filtering rule is filtering by a set dictionary filtering or regular rule.
According to the process, the intention classification models with different structures are adopted for multi-layer classification, for example, two layers are adopted, the search intention probability values of the users with different types of materials, which are obtained by the first layer, are recalled for use, and the second search intention probability values of the users with fine granularity, which are obtained by the second layer, for the materials with different types are used in the subsequent sequencing and distribution of the materials.
In the above method, after the material recall is performed for the user, further sorting and distribution are performed, which specifically includes: recalling materials for the user from the set of alternative materials as first recalled materials;
extracting search information characteristics of the search information of the user by adopting a neural network to obtain the search information characteristics of the user, searching from an alternative material set by adopting a distributed search analysis engine according to the search information characteristics of the user, and recalling the material for the user as a second recalled material;
searching from an alternative material set by adopting a set search rule according to the search information of the user, and recalling the material for the user as a third recalled material;
fusing the first recall material, second recall material, and third recall material;
and sequencing the fused recall materials according to the set second search intention probability value corresponding to the material type, and distributing the sequenced recall materials to the user.
Wherein the second search intention probability value for the corresponding material type comprises:
extracting search information characteristics of the search information of the user by adopting a neural network to obtain the search information characteristics of the user;
inputting the search information characteristics of the user and the search intention input characteristics of the user into a second intention classification model formed by a long-short term memory artificial neural network (LSTM) and an attention mechanism (attention), and outputting to obtain a second search intention probability value of the user on different types of materials.
The embodiments of the present application will be described in detail below by taking a network platform as a network platform in the vertical field, particularly a house network platform as an example.
In this example, the set time periods of different durations are a long-term time period, a middle-term time period, and a short-term time period, where the long-term time period may be set to half a year, the middle-term time period may be set to three months, and the short-term time period may be set to one week. After user portraits in different time periods are obtained, discretization is carried out on the user portraits based on categories, then cross fusion of user vectors is carried out through a depth recommendation model (deepfm), so that the user portraits in a long time period, the user portraits in a medium time period and the user portraits in a short time period of the user are obtained, and the user portraits are expressed in a vector form. The obtained user portrait in different time periods can be directly output or stored, and subsequent use is facilitated.
Fig. 3 is a framework diagram of an implementation process of determining a search intention of a user according to an embodiment of the present application. In the present example, the user image is depicted in the time dimension by a long-term period, a medium-term period, and a short-term period, which are three different periods. In this example, in consideration of the network platform specificity in the real estate field, the long-term period is set to half a year, the middle-term period is set to three months, and the short-term period is set to one week. Depicting a user representation of the user at different time periods is a first advantage of this example, and how to fuse the user's search information with the user representation of the user at different time periods is a second advantage of this example. Considering that the relevance of the user images in different time periods to the search intention of the user is different, a fusion model with a gate structure is adopted in the fusion, and the ratio of the user search intention input feature in the previous time period fused to the user search image feature in the current time period is adjusted through the gate structure to obtain the user search intention feature in the current time period until the user search intention input feature in the short time period is obtained. In this way, in the finally obtained user search intention input features, the user search image features in the short-term time period are enhanced, the user search image features in the medium-term time period are weakened, and the user search image features in the long-term time period are weakened. Therefore, the convergence speed of the whole network of the model for generating the user search intention input features is accelerated, the phenomenon that the influence of the user image in a certain time period on the finally obtained user search intention input features is overlarge can be avoided, the overfitting risk of the network is reduced, and the model can fully learn to obtain the feature information of the user image in different time periods.
In this example, in the fusion calculation stage, the search information vectorization of the user is obtained through a set pre-training model, and based on the search information of the user and the user portrait of the user in the time period, the calculation of the user search portrait characteristics of the time period is performed by using a feedforward neural network, and the subsequent fusion is performed by using a fusion model with a gate structure.
The fusion model with the gate structure is specifically adopted as follows: sigma is a sigmoid function, and the vector can be converted into a value in the range of 0-1 through the function, so that the first gating value and the second gating value are set. The fusion model adopts the following formula, and the user search intention input characteristics of the long-term time period, the medium-term time period and the short-term time period are sequentially output and obtained.
Wherein, the fusion formula of the long-term time period is as follows:
the fusion formula for the middle period is:
the fusion formula for the short-term period is:
where u is the user representation for different time periods, with subscripts identifying the different time periods, q is the user's search information, w represents a weight vector,
and y is a query user representation characteristic for different time periods, with different subscripts identifying the different time periods.
The symbol indicates that elements in the operation vector matrix are multiplied correspondingly, and two multiplied vector matrixes are of the same type. H' is an initialization vector and is preset, r represents a first gating value obtained by calculation based on search information, a splicing vector of a user portrait and a set first weight vector by adopting a sigmoid function, and subscripts of r represent different time periods; z represents a second gating value calculated by adopting a sigmoid function based on the search information, a splicing vector of the user portrait and a set second weight vector, subscripts of z represent different time periods, h represents finally obtained search intention input features, and different subscripts of h represent different time periods.
Here, w represents that the weight vector is initially set according to an empirical value, and is gradually adjusted in the training process, and the weight vectors used in calculating the first gating value and calculating the second gating value are different.
In this example, in the stage of identifying the search intention of the user, in consideration of the specificity of the network platform in the vertical domain of the real estate domain, in order to better understand the search intention of the user, the identification may include a two-layer structure, and the structure is dynamic, and dynamic addition may be performed according to business development. The first layer structure identifies wide entity search intents such as large intents of second-hand houses, new houses, rents, house property information, decorations and the like, and the second layer structure identifies specific search intents of a certain type of entities such as specific intents of house types, positions, prices and the like. The method has the advantages that the new service can be dynamically accessed, so that the search intention identification process of the user does not need to be greatly changed when the platform accesses the new service, and the cost is low. The realization of each layer of search intention is identified by the set intention classification model and the set filtering rule.
As shown in fig. 4, fig. 4 is a process diagram of an example of a search intention identification phase of a user according to an embodiment of the present application. As shown, h is first generated by the fusion phaseShort termThe entity search intention probability value of the user, such as large intention probabilities of second-hand houses, new houses, leases, house property information, decorations and the like, is obtained through the set shallow Depth Neural Network (DNN), because the fine-grained intention of the user does not need to be specifically identified, the final user search intention probability is obtained after the filtering is carried out through the set filtering rule, and the accuracy of identification is high. After the probability value of the entity search intention of the user is determined, the specific search intention of the user can be identified, which needs text feature information with finer granularity, and the text feature information can be obtained by adopting a network model of LSTM + attribute.
As shown in fig. 4, the search information of the user is obtained by obtaining text features through LSTM or DNN, and a search information vector o is obtained, and is also used in a subsequent distribution and sorting stage. And calculating the probability value of the search intention of the user on different types of materials by adopting the following formula.
Wherein h is h of the fusion stageShort termAnd V is a vector formed by fusing two vectors. Wherein, ViRepresenting each element in the vector V, i representing the material type, the total material typeThe number is C. SiThe ratio of the index of the current element to the sum of the indices of all elements is shown. And (3) calculating to obtain the value of S through softmax, wherein the value of the vector represents the probability value of the search intention of the user on different types of materials, w represents a set weight value (set according to an empirical value), e is the index calculation of an element and represents a limit value which can be reached by continuous doubling and increasing in unit time.
And filtering the search intention probability values of the users outputting the model to the different types of materials according to a filtering rule to obtain the final search intention probability values of the users to the different types of materials. The process of fig. 4 may be performed in two layers to further improve recognition accuracy.
The distribution sorting stage in this example is a process of material recall, sorting, and distribution performed after the search intention recognition stage. Based on the recalled materials and the ordering of the recalled materials, the user experience is affected by pushing the applicable materials for the user. In the stage, the search intention determination result of the user is received, the materials are recalled through different recalling material ways of the first layer, and the recalled materials are sorted and distributed according to the set search intention probability value through the second layer, so that the final distribution sequence of the materials is obtained. As shown in fig. 5, fig. 5 is a schematic diagram of an example process of the distribution sorting phase provided in the embodiment of the present application. The material recalls under different recalling material paths are mainly determined by calculating the degree of relevance of the search intention of the user. And respectively recalling materials in different ways after model vectorization, word segmentation and ner entity identification of the search information of the user. Inputting the search intention input characteristics of the user into an intention classification model for vectorized search information to obtain search intention probability values of the user on different types of materials, and recalling the materials for the user from an alternative material set according to the materials with the search intention probability values of the user on the different types of materials; for the participles and the ner entities in the search information, searching from the alternative material set by adopting a distributed search analysis engine (elastic search) and recalling the materials; and for the participles and the ner entities in the search information, searching from the alternative material set by adopting the set search rule, and recalling the materials. The elastic search obtains the relevance between texts by calculating the BM25, the relevance can be obtained while recalling materials, the search rule recall is the material recall based on the prefix, the suffix, the text distance and the like, and the hit materials can be recalled by setting different search rules. And after the materials recalled by the three recalled material ways are fused, sorting the fused recalled materials according to the second search intention probability value of the corresponding material type, and distributing the sorted recalled materials to the user. The second search intention probability value may be calculated using a layer two network of the process of fig. 4. The process of sequencing the fused recall materials comprises the following steps: and setting a material type keyword corresponding to the search intention aiming at each search intention, and sequencing according to the material type keyword of the corresponding related search intention in the fused material hits and the determined probability value of the related search intention, thereby improving the accuracy of sequencing and distributing materials.
Fig. 6 is a schematic structural diagram of an apparatus for material recall according to an embodiment of the present application, where the apparatus includes: a setting unit, a feature extraction unit, a fusion unit, a search intention classification unit and a recall unit, wherein,
the device comprises a setting unit, a processing unit and a control unit, wherein the setting unit is used for setting a plurality of time periods with different durations, and the time period with the long duration comprises the time period with the long duration;
the characteristic extraction unit is used for calculating the user search portrait characteristics of the time periods based on the search information of the user and the user portrait of the user in the time periods for each time period;
the fusion unit is used for sequentially carrying out fusion calculation on the user search portrait features of the time periods according to the length sequence of the time periods with duration until all the user search portrait features of the time periods are fused to obtain the search intention input features of the user, and the fusion calculation is carried out on the correlation of the user search portrait features of the time periods to the search intention input features of the user;
the search intention classification unit is used for inputting the search intention input characteristics of the user into an intention classification model to obtain the probability value of the search intention of the user on different types of materials;
and the recalling unit is used for recalling the materials for the user from the alternative material set according to the materials with the probability values of the search intentions of the user on the different types of materials.
In the device, the setting unit is further configured to set three time periods, including a first time period, a second time period and a third time period, where a duration of the first time period is greater than the second time period, and the second time period is greater than the third time period.
In the device, the feature extraction unit is further configured to calculate the user search image feature of the time period by inputting the user search information and the user image of the time period into a feed-forward neural network.
In the apparatus, the fusion unit is further configured to obtain the search intention input feature of the user specifically includes: a. selecting a current time period according to the time period length sequence, multiplying the search information and the splicing vector of the user portrait of the current time period by a set first weight vector, and calculating by adopting an activation function to obtain a first gating value; b. multiplying the search information and the splicing vector of the user portrait in the current time period by a set second weight vector, and calculating by adopting an activation function to obtain a second gating value; c. adding the product of the residual probability value of the first gating value and the user search portrait characteristics of the current time period and the product of the second gating value and the search intention input characteristics of the time period in the sequence of the current time period to obtain a sum value, and taking the sum value as the search intention input characteristics of the current time period; d. and b, determining whether the current time period is the time period with the shortest time length in the set time period, if so, taking the search intention input characteristic of the current time period as the search intention input characteristic of the user, if not, taking the next time period in the sequence of the current time period as the current time period, and returning to the step a to continue executing.
In the apparatus, the search intention classification unit is further configured to input the search intention input features of the user into an intention classification model, and obtaining the probability values of the search intention of the user for different types of materials includes: extracting search information characteristics of the search information of the user by adopting a neural network to obtain the search information characteristics of the user; inputting the search information characteristics of the user and the search intention input characteristics of the user into a first intention classification model, and outputting to obtain the probability values of the search intentions of the user to different types of materials.
In the apparatus, the search intention classifying unit is further configured to, before obtaining the probability values of the search intentions of the user on different types of materials, further include: and filtering the search intention probability values of the different types of materials output by the intention classification model according to a set filtering rule to obtain the search intention probability values of the different types of materials of the user.
In the apparatus, the apparatus further comprises a sorting unit, configured to recall the item for the user from the set of alternative items as a first recalled item; extracting search information characteristics of the search information of the user by adopting a neural network to obtain the search information characteristics of the user, analyzing and searching from an alternative material set by adopting a distributed search analysis engine according to the search information characteristics of the user, and recalling the material aiming at the user as a second recalled material; searching from an alternative material set by adopting a set search rule according to the search information of the user, and recalling the material for the user as a third recalled material; fusing the first recall material, second recall material, and third recall material; and sequencing the fused recall materials according to the second search intention probability value of the corresponding material type, and distributing the sequenced recall materials to the user.
Specifically, the ranking unit, further for the second search intention probability value of the corresponding material type, includes: extracting search information characteristics of the search information of the user by adopting a neural network to obtain the search information characteristics of the user; inputting the search information characteristics of the user and the search intention input characteristics of the user into a second intention classification model formed by a long-short term memory artificial neural network (LSTM) and an attention mechanism attention, and outputting to obtain a second search intention probability value of the user on different types of materials.
In another embodiment of the present application, there is also provided an electronic device, including: a processor; a memory storing a program configured to implement a method of material recall as described above when executed by the processor.
In another embodiment of the present application, a non-transitory computer readable storage medium is provided that stores instructions that, when executed by a processor, cause the processor to perform a method of material recall in the preceding embodiments. Fig. 7 is a schematic diagram of an electronic device according to another embodiment of the present application. As shown in fig. 7, another embodiment of the present application further provides an electronic device, which may include a processor 701, wherein the processor 701 is configured to perform the steps of the method for material recall. As can also be seen from fig. 7, the electronic device provided by the above embodiment further includes a non-transitory computer readable storage medium 702, the non-transitory computer readable storage medium 702 having stored thereon a computer program, which when executed by the processor 701 performs the steps of one of the above methods for material recall.
In particular, the non-transitory computer readable storage medium 702 can be a general purpose storage medium such as a removable disk, a hard disk, a FLASH, a Read Only Memory (ROM), an erasable programmable read only memory (EPROM or FLASH memory), or a portable compact disc read only memory (CD-ROM), etc., and the computer program on the non-transitory computer readable storage medium 702, when executed by the processor 701, can cause the processor 701 to perform the steps of one of the methods of material recall described above.
In practical applications, the non-transitory computer readable storage medium 302 may be included in the device/apparatus/system described in the above embodiments, or may exist alone without being assembled into the device/apparatus/system. The computer readable storage medium carries one or more programs which, when executed, are capable of performing the steps of the above-mentioned method for evaluating the implementation effectiveness of the application-carried operation service.
Yet another embodiment of the present application further provides a computer program product, which includes a computer program or instructions, and when the computer program or instructions are executed by a processor, the computer program or instructions implement the steps in the above-mentioned method for evaluating the implementation effect of the operation service carried by the application.
The flowchart and block diagrams in the figures of the present application illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments disclosed herein. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams or flowchart illustration, and combinations of blocks in the block diagrams or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.
Those skilled in the art will appreciate that various combinations and/or combinations of features recited in the various embodiments and/or claims of the present disclosure can be made, even if such combinations or combinations are not explicitly recited in the present application. In particular, the features recited in the various embodiments and/or claims of the present application may be combined and/or coupled in various ways, all of which fall within the scope of the present disclosure, without departing from the spirit and teachings of the present application.
The principles and embodiments of the present application are explained herein using specific examples, which are provided only for the purpose of facilitating understanding of the method and the core idea of the present application and are not intended to limit the present application. It will be appreciated by those skilled in the art that changes may be made in this embodiment and its broader aspects and without departing from the principles, spirit and scope of the invention, and that all such modifications, equivalents, improvements and equivalents as may be included within the scope of the invention are intended to be protected by the claims.
Claims (10)
1. A method of material recall, the method comprising:
setting a plurality of time periods with different durations, wherein the long time period comprises a long time period;
for each time period, calculating user search portrait characteristics of the time period based on search information of a user and a user portrait of the user in the time period;
according to the length sequence of the time periods with the duration, sequentially carrying out fusion calculation on the user search portrait features of the time periods to obtain search intention input features of the user, wherein the fusion calculation is carried out on the basis of the relevance of the user search portrait features of the time periods to the search intention input features of the user;
inputting the search intention input characteristics of the user into an intention classification model to obtain the probability values of the search intentions of the user to different types of materials;
and recalling the materials for the user from the alternative material set according to the materials with the probability values of the search intentions of the user on the different types of materials.
2. The method of claim 1, wherein the setting the plurality of time periods having different durations to three time periods comprises:
the time-sharing system comprises a first time period, a second time period and a third time period, wherein the duration of the first time period is greater than that of the second time period, and the second time period is greater than that of the third time period.
3. The method of claim 1, wherein said calculating user search representation characteristics for said time period comprises:
and inputting the user search information and the user image of the time period into a set feed-forward neural network, and calculating to obtain the user search image characteristics of the time period.
4. The method of claim 1, wherein sequentially performing fusion calculation on the user search image features of the time period to obtain the search intention input features of the user comprises:
a. selecting a current time period according to the time period length sequence, multiplying the search information and the splicing vector of the user portrait of the current time period by a set first weight vector, and calculating by adopting an activation function to obtain a first gating value;
b. multiplying the search information and the splicing vector of the user portrait in the current time period by a set second weight vector, and calculating by adopting an activation function to obtain a second gating value;
c. adding the product of the residual probability value of the first gating value and the user search portrait characteristics of the current time period and the product of the second gating value and the search intention input characteristics of the time period in the sequence of the current time period to obtain a sum value, and taking the sum value as the search intention input characteristics of the current time period;
d. and b, determining whether the current time period is the time period with the shortest time length in the set time period, if so, taking the search intention input characteristic of the current time period as the search intention input characteristic of the user, if not, taking the next time period in the sequence of the current time period as the current time period, and returning to the step a to continue executing.
5. The method of claim 1, wherein inputting the user's search intent into a feature input intent classification model, obtaining search intent probability values for different types of material by the user comprises:
extracting search information characteristics of the search information of the user by adopting a neural network to obtain the search information characteristics of the user;
inputting the search information characteristics of the user and the search intention input characteristics of the user into a first intention classification model, and outputting to obtain the probability values of the search intentions of the user to different types of materials.
6. The method of claim 1 or 5, further comprising, prior to obtaining the user's search intention probability values for different types of material:
and filtering the search intention probability values of the different types of materials output by the intention classification model according to a set filtering rule to obtain the search intention probability values of the different types of materials of the user.
7. The method of claim 1, wherein the method further comprises:
recalling materials for the user from the set of alternative materials as first recalled materials;
extracting search information characteristics of the search information of the user by adopting a neural network to obtain the search information characteristics of the user, analyzing and searching from an alternative material set by adopting a distributed search analysis engine according to the search information characteristics of the user, and recalling the material aiming at the user as a second recalled material;
searching from an alternative material set by adopting a set search rule according to the search information of the user, and recalling the material for the user as a third recalled material;
fusing the first recall material, second recall material, and third recall material;
and sequencing the fused recall materials according to the second search intention probability value of the corresponding material type, and distributing the sequenced recall materials to the user.
8. The method of claim 7, wherein the second search intent probability value for the corresponding material type comprises:
extracting search information characteristics of the search information of the user by adopting a neural network to obtain the search information characteristics of the user;
inputting the search information characteristics of the user and the search intention input characteristics of the user into a second intention classification model formed by a long-short term memory artificial neural network (LSTM) and an attention mechanism attention, and outputting to obtain a second search intention probability value of the user on different types of materials.
9. A computer product, comprising an electronic device, comprising:
a processor;
a memory storing a program configured to implement the method of material recall of any of claims 1-8 when executed by the processor.
10. A non-transitory computer readable storage medium storing instructions that, when executed by a processor, cause the processor to perform the steps of the method of material recall of any of claims 1-8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111612538.4A CN114329199A (en) | 2021-12-27 | 2021-12-27 | Material recall method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111612538.4A CN114329199A (en) | 2021-12-27 | 2021-12-27 | Material recall method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114329199A true CN114329199A (en) | 2022-04-12 |
Family
ID=81012743
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111612538.4A Pending CN114329199A (en) | 2021-12-27 | 2021-12-27 | Material recall method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114329199A (en) |
-
2021
- 2021-12-27 CN CN202111612538.4A patent/CN114329199A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110162749B (en) | Information extraction method, information extraction device, computer equipment and computer readable storage medium | |
| US20220269707A1 (en) | Method and system for analyzing entities | |
| CN109544306B (en) | Cross-domain recommendation method and device based on user behavior sequence characteristics | |
| CN111061946B (en) | Method, device, electronic equipment and storage medium for recommending scenerized content | |
| CN110717098B (en) | Meta-path-based context-aware user modeling method and sequence recommendation method | |
| CN110032632A (en) | Intelligent customer service answering method, device and storage medium based on text similarity | |
| CN109766557B (en) | Emotion analysis method and device, storage medium and terminal equipment | |
| WO2016028446A1 (en) | Knowledge-graph biased classification for data | |
| CN111914067A (en) | Chinese text matching method and system | |
| CN111191092A (en) | Portrait data processing method and portrait model training method | |
| CN114169869B (en) | Attention mechanism-based post recommendation method and device | |
| CN111160000B (en) | Composition automatic scoring method, device terminal equipment and storage medium | |
| CN114358657B (en) | Post recommendation method and device based on model fusion | |
| CN111859967A (en) | Entity identification method and device and electronic equipment | |
| CN114168819B (en) | Post matching method and device based on graph neural network | |
| CN112380421A (en) | Resume searching method and device, electronic equipment and computer storage medium | |
| US20220366295A1 (en) | Pre-search content recommendations | |
| CN112765339B (en) | Personalized book recommendation method and system based on reinforcement learning | |
| CN116680481B (en) | Search ranking method, apparatus, device, storage medium and computer program product | |
| CN113609337A (en) | Pre-training method, device, equipment and medium of graph neural network | |
| CN112508177A (en) | Network structure searching method and device, electronic equipment and storage medium | |
| CN111062449A (en) | Prediction model training method, interestingness prediction device and storage medium | |
| CN115879508A (en) | Data processing method and related device | |
| CN115017413A (en) | Recommendation method and device, computing equipment and computer storage medium | |
| CN115186750A (en) | Model training method, device, equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |