A method and apparatus for material recall

Abstract

A method and apparatus for material recall includes setting multiple time periods with different durations, including shorter time periods; for each time period, calculating user search profile features based on user search information and user profiles within that time period; sequentially fusing the user search profile features of the time periods according to their duration to obtain user search intent input features; calculating the relevance of the user search profile features of the time periods to the user search intent input features; inputting the user search intent input features into an intent classification model to obtain probability values ​​of the user's search intent for different types of materials; and recalling materials from a candidate material set based on the probability values ​​of the user's search intent for different types of materials, thereby accurately determining the user's search intent and recalling the materials.

Description

Technical Field

[0001] This application relates to the field of computing network technology, and in particular to a method and apparatus for material recall. Background Technology

[0002] With the development of computer networks, more and more network platforms are hosted on these networks, providing various services to users. Currently, the most direct interaction between network platforms and users is that, based on the search information sent by the user, the search engine set up by the network platform searches the set of candidate materials and distributes materials that match the search information to the user. The core issue for network platforms in pushing materials to users is how to accurately understand the user's search intent. Only by fully understanding the user's search intent can materials be distributed to the user quickly and accurately.

[0003] With the emergence and development of user profiling technology, user profiles play a crucial role in understanding users' search intent. By combining user profiles with users' search information, search intent can be determined. This is especially true for vertical online platforms, where different users may have different search intents for the same search query, and the same user's search intent may also differ at different times. Therefore, combining user search information with user profiles to determine search intent is a commonly used technical approach.

[0004] However, in the stage of determining the user's search intent during the material recall process, the user profiles on which they are based will change over time. Simply combining the accumulated information of the user profile over a set time period with the user's search information will result in an inaccurate determination of the user's search intent. Based on inaccurate user search intent, the recalled materials will also be inaccurate. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method and apparatus for material recall, which can accurately determine the user's search intent and recall materials by combining user profiles from multiple different time periods with the user's search information.

[0006] The embodiments of this application are implemented as follows:

[0007] A method for recalling materials, the method comprising:

[0008] Set multiple time periods with different durations, where longer time periods include shorter time periods;

[0009] For each time period, based on the user's search information and the user profile of the user in that time period, calculate the user search profile features for that time period;

[0010] Based on the length order of the time period, the user search profile features of the time period are sequentially fused and calculated to obtain the user's search intent input features. The fusion calculation is based on the relevance of the user's search intent input features to the user search profile features of the time period.

[0011] The user's search intent is input into the intent classification model to obtain the probability value of the user's search intent for different types of materials;

[0012] Based on the probability values ​​of the user's search intent for different types of materials, materials for the user are recalled from the candidate material set.

[0013] Preferably, the setting of multiple time periods with different durations is three time periods, including:

[0014] The time period is divided into three time periods: the first time period, the second time period, and the third time period. The first time period is longer than the second time period, and the second time period is longer than the third time period.

[0015] Preferably, the calculation of user search profile features for the time period includes:

[0016] The user's search information and the user profile for the time period are input into a feedforward neural network to calculate the user search profile features for the time period.

[0017] Preferably, the user search profile features for the time period are sequentially fused and calculated to obtain the user's search intent input features, including:

[0018] a. Based on the length order of the time periods, select the current time period, multiply the concatenated vector of the search information and the user profile of the current time period with the set first weight vector, and then use the activation function to calculate to obtain the first gate value;

[0019] b. Multiply the concatenated vector of the search information and the user profile of the current time period with the set second weight vector, and then use the activation function to calculate the second gate value;

[0020] c. The product of the remaining probability value of the first gate value and the user search profile feature of the current time period, and the product of the second gate value and the search intent input feature of the previous time period in the current time period are added together to obtain the sum value, which is used as the search intent input feature of the current time period.

[0021] d. Determine whether the current time period is the shortest time period within the set time period. If yes, use the search intent input feature of the current time period as the user's search intent input feature. If no, use the next time period in sequence as the current time period and return to step a to continue execution.

[0022] Preferably, the user's search intent is input into the intent classification model to obtain the probability values ​​of the user's search intent for different types of materials, including:

[0023] The user's search information features are obtained by using a neural network to extract search information features.

[0024] The user's search information features and search intent input features are input into the first intent classification model, and the output is the probability value of the user's search intent for different types of materials.

[0025] Preferably, before obtaining the probability values ​​of the user's search intent for different types of materials, the method further includes:

[0026] The search intent probability values ​​for different types of materials output by the intent classification model are filtered according to the set filtering rules to obtain the user's search intent probability values ​​for different types of materials.

[0027] Preferably, the method further includes:

[0028] The materials for the user will be recalled from the alternative material set as the first recalled materials;

[0029] The search information features of the user are extracted using a neural network. Based on the search information features of the user, a distributed search analysis engine is used to analyze and search from the candidate material set, and recall materials for the user as the second recall materials.

[0030] Based on the user's search information, the set search rules are used to search from the candidate material set and recall the material for the user as the third recalled material;

[0031] The first recalled material, the second recalled material, and the third recalled material are merged;

[0032] Based on the second search intent probability value of the corresponding material type, the merged recall materials are sorted, and the sorted recall materials are distributed to the user.

[0033] Preferably, the second search intent probability value for the corresponding material type includes:

[0034] The user's search information features are obtained by using a neural network to extract search information features.

[0035] The user's search information features and search intent input features are input into a second intent classification model composed of a long short-term memory artificial neural network (LSTM) and an attention mechanism, and the output is the probability value of the user's second search intent for different types of materials.

[0036] A computer product, including an electronic device, comprising:

[0037] processor;

[0038] A memory storing a program configured to implement the material recall method described in any of the preceding claims when executed by the processor.

[0039] A non-transitory computer-readable storage medium storing instructions, characterized in that, when executed by a processor, the instructions cause the processor to perform the steps of the material recall method described in any of the preceding claims.

[0040] As seen above, this embodiment sets multiple time periods with different durations, including shorter time periods. For each time period, user search profile features are calculated based on the user's search information and the user profile within that time period. The user search profile features are then sequentially fused according to the length of the time periods to obtain the user's search intent input features. This fusion calculation assesses the relevance of the user search profile features to the user's search intent input features. These features are then input into an intent classification model to obtain the user's search intent probability values ​​for different types of materials. Based on these probability values, materials for the user are recalled from a candidate material set. Thus, because the user's search intent is determined not based on a single time period's user profile but on multiple time period profiles, a fine-grained user profile can be obtained from the time dimension. This allows for a thorough understanding of the user's search intent, ensuring accuracy in the determined search intent. Based on this accurate search intent, materials can be accurately recalled for the user. Attached Figure Description

[0041] Figure 1 A flowchart illustrating a current method for distributing materials to users, provided as an embodiment of this application;

[0042] Figure 2A flowchart illustrating a material recall method provided in this application embodiment;

[0043] Figure 3 A framework diagram of the implementation process for determining a user's search intent provided in the embodiments of this application;

[0044] Figure 4 Example process diagram of the user's search intent identification stage provided in the embodiments of this application;

[0045] Figure 5 A schematic diagram illustrating a process example of the distribution and sorting stage provided in an embodiment of this application;

[0046] Figure 6 A schematic diagram of a material recall device provided in an embodiment of this application;

[0047] Figure 7 This is a schematic diagram of an electronic device provided for another embodiment of this application. Detailed Implementation

[0048] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0049] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0050] The technical solution of this application will be described in detail below with reference to specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0051] As the background technology shows, in the material recall process, determining the user's search intent by simply combining the accumulated information of the user profile over a set time period with the user's search intent results in inaccurate information, leading to inaccurate material recall. This is especially true for vertical online platforms such as housing networks, which are infrequently used by users. User profiles, particularly behavioral characteristics, are very sparse. Accumulating all historical information from user profiles does not provide a sufficient understanding of user behavior, which changes over time. Therefore, relying solely on the accumulated information of user profiles over historical time periods is insufficient to accurately capture user search intent.

[0052] Figure 1 The flowchart of the method for distributing materials to users provided in this application embodiment includes the following specific steps:

[0053] Step 101: Obtain user profiles for the user within a historical time period;

[0054] Step 102: Obtain the user's search information;

[0055] Step 103: Combine the user profile with the search information to obtain the user's search composite information features;

[0056] Step 104: Input the user's search concatenation information features into the intent classification model, and output the user's search intent probability value;

[0057] Step 105: Based on the user's search intent probability value, recall materials for the user from the candidate material set, sort and distribute them, and then push them to the user.

[0058] In this embodiment of the application, the user profile mainly consists of user behavior information and user tags. The user behavior information includes information such as the user's browsing and / or clicking behavior, and the user tags include the tags of the materials browsed by the user, such as setting corresponding user tags for the properties that the user has browsed.

[0059] from Figure 1It's evident that simply concatenating a user's historical profile with their search information to determine their search intent probability is insufficient, especially for vertical online platforms. This approach fails to adequately account for the temporal changes in user profiles. In reality, user profiles evolve over time. For instance, on a real estate platform, a user searching for listings in the same neighborhood might, over time, express different intentions: renovating, renting, or selling the property. Therefore, user search intent changes over time, necessitating a comprehensive consideration of these temporal shifts in user profiles.

[0060] The above analysis shows that although user profiles are currently used to understand user search intent, they are not layered over time to provide a multi-layered understanding of users. The understanding of user search intent remains point-based, failing to characterize user profiles over time. Furthermore, the integration of user profiles with search information across different time periods is not adequately considered, resulting in varying understandings of user search intent and a lack of effective multi-layered user profile fusion strategies.

[0061] To address the aforementioned issues, this application embodiment sets multiple time periods with varying durations, including shorter time periods. For each time period, user search profile features are calculated based on the user's search information and the user profile within that time period. The user search profile features are then sequentially fused according to the length of the time periods to obtain the user's search intent input features. This fusion calculation assesses the relevance of the user's search intent input features to the user search profile features within the time periods. These user search intent input features are then input into an intent classification model to obtain the user's search intent probability values ​​for different types of materials. Based on the user's search intent probability values ​​for different types of materials, materials relevant to the user are recalled from a candidate material set.

[0062] In this way, since the user search intent is determined not based on a user profile of a single time period, but on user profiles of multiple time periods, fine-grained user profiles can be obtained from the time dimension. Based on this, the user search intent is fully understood, making the determined user search intent accurate. Based on the accurate user search intent, materials can be accurately recalled for the user.

[0063] Furthermore, after accurately recalling materials for a user, the platform can also accurately sort and distribute the recalled materials to that user. This means that the materials pushed to different users or the same user at different times will differ for the same search query. Thus, this embodiment of the application can understand the user's search intent at a fine-grained, multi-layered time dimension, improving the user's search experience.

[0064] The different time periods that can be set are long-term, medium-term, and short-term. Long-term periods can be set to six months, medium-term periods to three months, and short-term periods to one week. These settings can be adjusted for different network platforms. Because user profiles from different time periods can be merged, user profiles can be depicted with fine granularity over time, allowing for a full understanding of changes in user search intent.

[0065] In this embodiment of the application, when the user search profile features of the time period are sequentially fused according to the length of the time period, a gate structure is used to control the fusion ratio of user search profile features in different time periods, thereby weakening or strengthening the ratio of user search profile features in different time periods to the overall user's search intent input features, and fully understanding the current user's search intent.

[0066] Figure 2 A flowchart of a material recall method provided in this application embodiment is shown, and its specific steps are as follows:

[0067] Step 201: Set multiple time periods with different durations, where longer time periods include shorter time periods;

[0068] Step 202: For each time period, calculate the user search profile features for that time period based on the user's search information and the user profile of the user in that time period;

[0069] Step 203: According to the length order of the time periods, the user search profile features of the time periods are sequentially fused and calculated until all the user search profile features of the time periods are fused to obtain the user's search intent input features. The fusion calculation is based on the relevance of the user's search intent input features to the user search profile features of the time periods.

[0070] Step 204: Input the user's search intent into the intent classification model to obtain the probability values ​​of the user's search intent for different types of materials;

[0071] Step 205: Based on the probability values ​​of the user's search intent for different types of materials, recall materials for the user from the candidate material set.

[0072] In the above method, the setting of multiple time periods with different durations is divided into three time periods, including:

[0073] The system is divided into three time periods: a first time period, a second time period, and a third time period. The first time period is longer than the second time period, and the second time period is longer than the third time period. Specifically, the first time period is a long-term period, the second time period is a medium-term period, and the third time period is a short-term period. The long-term period can be set to six months, the medium-term period can be set to three months, and the short-term period can be set to one week. These settings can be adjusted according to the different network platforms used.

[0074] In this way, user profiles for different time periods are obtained based on different durations, and then further fusion is performed to create a fine-grained user profile over time, thereby fully understanding changes in users' search intent.

[0075] In the above method, calculating the user search profile features for the time period includes:

[0076] The user's search information and the user profile for the specified time period are input into a feedforward neural network to calculate the user search profile features for that time period. Here, a feedforward neural network is used to fuse the user's search information and the user profile for that time period within that time period.

[0077] In the above method, the user search profile features of the time period are sequentially fused and calculated to obtain the user's search intent input features, including:

[0078] a. Based on the length order of the time periods, select the current time period, multiply the concatenated vector of the search information and the user profile of the current time period with the first weight vector, and then use the activation function to calculate to obtain the first gate value;

[0079] b. Multiply the concatenated vector of the search information and the user profile of the current time period with the second weight vector, and then use the activation function to calculate the second gating value;

[0080] c. The product of the remaining probability value of the first gate value and the user search profile feature of the current time period is added to the product of the second gate value and the search intent input feature of the previous time period in the current time period, and the sum is used as the search intent input feature of the current time period.

[0081] d. Determine whether the current time period is the shortest time period within the set time period. If yes, use the search intent input feature of the current time period as the user's search intent input feature. If no, use the next time period in sequence as the current time period and return to step a to continue execution.

[0082] It can be seen that when the user search profile features of the time periods are sequentially fused, the user search profile features of longer time periods are fused first, followed by the user search profile features of shorter time periods. During fusion, the relevance of the user search profile features of the current time period to the final search intent input features is adjusted according to the settings of the first and second gate values ​​for the current time period, as well as the relevance of the search intent input features of the previous time period to the final user search intent input features. Typically, the first gate value for the current time period is smaller than the second gate value, resulting in a higher relevance between the search intent input features of shorter time periods and the final search intent input features compared to those of longer time periods.

[0083] In the above method, the user's search intent is input into the intent classification model to obtain the probability values ​​of the user's search intent for different types of materials, including:

[0084] The user's search information features are obtained by using a neural network to extract search information features.

[0085] The user's search information features and search intent are input into the intent classification model, and the output is the probability value of the user's search intent for different types of materials.

[0086] In the above method, before obtaining the probability values ​​of the user's search intent for different types of materials, the method further includes:

[0087] The search intent probability values ​​for different types of materials output by the intent classification model are filtered according to the set filtering rules to obtain the user's search intent probability values ​​for different types of materials. Here, the filtering rules are set through dictionary filtering or regular expression rules.

[0088] Following the above process, different intent classification models can be used for multi-level classification, such as two layers. The probability values ​​of user search intent for different types of materials obtained from the first layer can be used for material recall, while the fine-grained probability values ​​of user search intent for different types of materials obtained from the second layer can be used in subsequent sorting and distribution of materials.

[0089] In the above method, after recalling materials for users, further sorting and distribution are carried out, specifically including: recalling materials for users from the candidate material set as the first recalled materials;

[0090] The search information features of the user are extracted using a neural network. Based on the search information features of the user, a distributed search analysis engine is used to search from the candidate material set and recall materials for the user as the second recalled materials.

[0091] Based on the user's search information, the set search rules are used to search from the candidate material set and recall the material for the user as the third recalled material;

[0092] The first recalled material, the second recalled material, and the third recalled material are merged;

[0093] Based on the second search intent probability value set for the corresponding material type, the merged recall materials are sorted, and the sorted recall materials are distributed to the user.

[0094] The second search intent probability value for the corresponding material type includes:

[0095] The user's search information features are obtained by using a neural network to extract search information features.

[0096] The user's search information features and search intent input features are input into a second intent classification model composed of a long short-term memory artificial neural network (LSTM) and an attention mechanism, and the output is the probability value of the user's second search intent for different types of materials.

[0097] The following describes the embodiments of this application in detail, taking network platforms in the vertical field, especially housing network platforms, as an example.

[0098] In this example, the different time periods are defined as long-term, medium-term, and short-term. Long-term periods can be set to six months, medium-term periods to three months, and short-term periods to one week. After obtaining user profiles for different time periods, these profiles are discretized based on categories. Then, a deep recommendation model (deepfm) is used to cross-fuse the user vectors, resulting in user profiles for the long-term, medium-term, and short-term periods, represented as vectors. The obtained user profiles for different time periods can be directly output or stored for later use.

[0099] Figure 3 This document presents a framework diagram for determining a user's search intent, as provided in an embodiment of this application. In this example, user profiles are depicted over time using three different time periods: long-term, medium-term, and short-term. Considering the unique characteristics of online platforms in the real estate sector, the long-term period is set to six months, the medium-term period to three months, and the short-term period to one week. The first advantage of this example is its ability to depict user profiles across different time periods; the second advantage is its ability to integrate user search information with these profiles. Given that the relevance of user profiles to user search intent varies across time periods, a fusion model with a gate structure is employed. This gate structure adjusts the ratio of user search intent input features from the previous time period to the user search profile features of the current time period, thus obtaining the user search intent features for the current time period, until the user search intent input features for the short-term time period are obtained. This results in a final user search intent input feature set that strengthens the user search profile features of the short-term time period, weakens the user search profile features of the medium-term time period, and further weakens the user search profile features of the long-term time period. In this way, not only can the convergence speed of the network of the model that generates user search intent input features be accelerated, but the influence of user profiles in a certain time period on the final user search intent input features can also be avoided, reducing the risk of overfitting of the network and enabling the model to fully learn the feature information of user profiles in different time periods.

[0100] In this example, during the fusion calculation stage, the user's search information is vectorized using a pre-trained model. Based on the user's search information and the user profile of the user during the time period, the user search profile features for the time period are calculated using a feedforward neural network. The subsequent fusion is performed using a fusion model with a gate structure.

[0101] The fusion model employing a gated structure is as follows: σ is the sigmoid function, which transforms a vector into a value within the range of 0 to 1, thereby setting the first and second gate values. This fusion model uses the following formula to sequentially output the user search intent input features for long-term, medium-term, and short-term time periods.

[0102] The fusion formula for the long-term time period is as follows:

[0103]

[0104]

[0105]

[0106] The fusion formula for the intermediate time period is:

[0107]

[0108]

[0109]

[0110] The fusion formula for short-term time periods is:

[0111]

[0112]

[0113]

[0114] Where u represents the user profile at different time periods, with its subscript indicating the different time periods, q represents the user's search information, and w represents the weight vector. This represents the combined information of user profiles and user search information at different time periods. y represents the query user profile features at different time periods, with different subscripts indicating different time periods. The symbol indicates that the elements in the operation vector matrix are multiplied correspondingly, and the two multiplied vector matrices are of the same type. Here, h′ is the initialization vector, which is pre-set; r represents the first gate value calculated using the sigmoid function based on the concatenated vector of search information and user profile and the set first weight vector; the subscript of r indicates different time periods. z represents the second gate value calculated using the sigmoid function based on the concatenated vector of search information and user profile and the set second weight vector; the subscript of z indicates different time periods. h represents the final search intent input feature; different subscripts of h represent different time periods.

[0115] Here, w indicates that the weight vector is initially set based on empirical values ​​and is gradually adjusted during training. The weight vectors used to calculate the first gating value and the second gating value are different.

[0116] In this example, considering the unique characteristics of a real estate-focused online platform during the user search intent identification phase, a two-layer structure can be incorporated for better understanding user search intent. This structure is dynamic and can be added dynamically based on business development. The first layer identifies broad entity search intents, such as "used homes," "new homes," "rentals," "real estate information," and "renovation." The second layer identifies specific search intents for a particular type of entity, such as "house type," "location," and "price." This approach allows for dynamic integration of new services without significantly altering the user search intent identification process, resulting in low costs. Each layer of search intent is identified through a combination of an established intent classification model and filtering rules.

[0117] like Figure 4 As shown, Figure 4 This is an example process diagram of the user search intent recognition stage provided in an embodiment of this application. As shown in the figure, h is first generated through the fusion stage. 短期 A shallow deep neural network (DNN) is used to obtain the probability values ​​of the user's entity search intent, such as broad intent probabilities like "used houses," "new houses," "rentals," "real estate information," and "renovation." Since it's not necessary to specifically identify the user's fine-grained intent, filtering by set rules yields the final user search intent probability, ensuring high accuracy. After determining the user's entity search intent probability values, the specific search intent can be identified. This requires more fine-grained text feature information, which can be obtained using an LSTM+attention network model.

[0118] like Figure 4 As shown, the user's search information is obtained by acquiring text features through LSTM or DNN, resulting in a search information vector o. This vector will also be used in the subsequent distribution and ranking stage. The probability values ​​of the user's search intent for different types of materials are calculated using the following formula.

[0119]

[0120]

[0121] Where h is the h in the fusion stage. 短期 V is the vector obtained by fusing the two vectors. i Let S represent each element in vector V, where i represents the material type, and the total number of material types is C. iThis represents the ratio of the current element's exponent to the sum of the exponents of all elements. The value of S is calculated using the softmax algorithm. This vector represents the probability of a user's search intent for different types of materials. w represents the set weight value (set based on experience), and e is the element's exponent calculation, signifying the maximum value achievable through continuous doubling within a unit of time.

[0122] After filtering the model output of user search intent probability values ​​for different types of materials using filtering rules, the final user search intent probability values ​​for different types of materials are obtained. Figure 4 The process can be performed in two layers to further improve recognition accuracy.

[0123] In this example, the distribution and sorting stage follows the search intent identification stage, involving the recall, sorting, and distribution of materials. Pushing suitable materials to users based on the recalled materials and their sorting impacts the user experience. This stage takes over from the user's search intent determination result, recalling materials through different recall methods in the first layer, and then sorting and distributing each recalled material according to the set search intent probability value in the second layer, resulting in the final distribution order of the materials. For example... Figure 5 As shown, Figure 5 This is a schematic diagram illustrating an example of the distribution and sorting stage provided in this application embodiment. Material recall under different recall paths is primarily determined by calculating the relevance to the user's search intent. After the user's search information is vectorized, segmented, and identified as a NER entity, different material recall paths are applied. For vectorized search information, the user's search intent is input into an intent classification model to obtain the probability value of the user's search intent for different types of materials. Based on the probability value of the user's search intent for different types of materials, materials targeting the user are recalled from a candidate material set. For segmented words and NER entities in the search information, a distributed search analysis engine (Elasticsearch) is used to search from the candidate material set for material recall. Furthermore, for segmented words and NER entities in the search information, set search rules are also used to search from the candidate material set for material recall. Elasticsearch obtains the relevance between texts by calculating BM25, thus obtaining relevance while recalling materials. Search rule recall uses methods such as prefixes, suffixes, and text distance for material recall; different set search rules can also be used to recall matched materials. After merging the materials recalled through the three recall channels, the merged recalled materials are sorted according to the second search intent probability value of the corresponding material type, and then the sorted recalled materials are distributed to the users. This second search intent probability value can be obtained by using... Figure 4The second layer of the network is used for calculation. The process of sorting the merged recall materials is as follows: for each search intent, set the material type keywords corresponding to the search intent, and sort them according to the material type keywords of the merged materials that match the relevant search intent and the determined probability value of the relevant search intent, thereby improving the accuracy of sorting and distributing materials.

[0124] Figure 6 This is a schematic diagram of a material recall device provided in an embodiment of this application. The device includes: a setting unit, a feature extraction unit, a fusion unit, a search intent classification unit, and a recall unit.

[0125] The setting unit is used to set multiple time periods with different durations, wherein the time periods with longer durations include the time periods with shorter durations;

[0126] The feature extraction unit is used to calculate the user search profile features for each time period based on the user's search information and the user profile of the user in the time period.

[0127] The fusion unit is used to sequentially perform fusion calculations on the user search profile features of the time periods according to the length of the time periods, until all the user search profile features of the time periods are fused to obtain the user's search intent input features. The fusion calculation is based on the relevance of the user's search intent input features to the user search profile features of the time periods.

[0128] The search intent classification unit is used to input the user's search intent input features into the intent classification model to obtain the probability value of the user's search intent for different types of materials;

[0129] The recall unit is used to recall materials for the user from a set of candidate materials based on the probability values ​​of the user's search intent for different types of materials.

[0130] In this device, the setting unit is also used to set three time periods, including a first time period, a second time period, and a third time period, wherein the duration of the first time period is longer than that of the second time period, and the second time period is longer than that of the third time period.

[0131] In this device, the feature extraction unit is also used to input the user's search information and the user profile of the time period into a set feedforward neural network to calculate the user search profile features of the time period.

[0132] In this device, the fusion unit is further configured to obtain the user's search intent input features, specifically including: a) selecting the current time period according to the length order of the time periods, multiplying the concatenated vector of the search information and the user profile of the current time period with a set first weight vector, and then calculating using an activation function to obtain a first gate value; b) multiplying the concatenated vector of the search information and the user profile of the current time period with a set second weight vector, and then calculating using an activation function to obtain a second gate value; c) adding the product of the remaining probability value of the first gate value and the user search profile features of the current time period, and the product of the second gate value and the search intent input features of the previous time period in sequence of the current time period, and obtaining a sum value as the search intent input feature of the current time period; d) determining whether the current time period is the shortest time period within the set time period. If so, the search intent input feature of the current time period is used as the user's search intent input feature; if not, the next time period in sequence of the current time period is used as the current time period, and the process returns to step a to continue execution.

[0133] In this device, the search intent classification unit is further configured to input the user's search intent input features into the intent classification model to obtain the probability values ​​of the user's search intent for different types of materials, including: extracting search information features from the user's search information using a neural network to obtain the user's search information features; inputting the user's search information features and the user's search intent input features into the first intent classification model, and outputting the probability values ​​of the user's search intent for different types of materials.

[0134] In this device, the search intent classification unit is further configured to, before obtaining the search intent probability values ​​of the user for different types of materials, filter the search intent probability values ​​of different types of materials output by the intent classification model according to the set filtering rules, and then obtain the search intent probability values ​​of the user for different types of materials.

[0135] The device also includes a sorting unit, used to recall materials for the user from a set of candidate materials as first recalled materials; extract search information features from the user's search information using a neural network to obtain the user's search information features; analyze and search from the set of candidate materials using a distributed search analysis engine based on the user's search information features to recall materials for the user as second recalled materials; search from the set of candidate materials using set search rules based on the user's search information to recall materials for the user as third recalled materials; merge the first, second, and third recalled materials; sort the merged recalled materials according to the second search intent probability value corresponding to the material type; and distribute the sorted recalled materials to the user.

[0136] Specifically, the sorting unit is further configured to use the second search intent probability value for the corresponding material type as follows: extracting search information features from the user's search information using a neural network to obtain the user's search information features; inputting the user's search information features and the user's search intent input features into a second intent classification model composed of a long short-term memory artificial neural network LSTM and an attention mechanism, and outputting the second search intent probability value for the user for different types of materials.

[0137] In another embodiment of this application, an electronic device is also provided, characterized in that it includes an electronic device comprising: a processor; and a memory storing a program configured to implement a material recall method as described above when executed by the processor.

[0138] In another embodiment of this application, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing instructions, which, when executed by a processor, cause the processor to perform a material recall method as described in the foregoing embodiments. Figure 7 This is a schematic diagram of an electronic device provided for another embodiment of this application. (See diagram below.) Figure 7 As shown, another embodiment of this application also provides an electronic device, which may include a processor 701, wherein the processor 701 is used to perform the steps of the above-described method for material recall. From Figure 7 It can also be seen that the electronic device provided in the above embodiments further includes a non-transitory computer-readable storage medium 702, on which a computer program is stored, and the computer program is executed by the processor 701 to perform the steps of the above-described material recall method.

[0139] Specifically, the non-transitory computer-readable storage medium 702 can be a general-purpose storage medium, such as a removable disk, hard disk, FLASH, read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), or portable compact disk read-only memory (CD-ROM), etc. When the computer program on the non-transitory computer-readable storage medium 702 is run by the processor 701, it can trigger the processor 701 to execute the various steps of the above-described material recall method.

[0140] 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 it may exist independently without being assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, can perform the various steps in the above-described method for evaluating the operational effectiveness of an application-borne service.

[0141] Another embodiment of this application provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the various steps in the above-described method for evaluating the effectiveness of operational services carried by an application.

[0142] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments disclosed in this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings. For example, two blocks shown connectedly may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0143] Those skilled in the art will understand that the features described in the various embodiments and / or claims of this disclosure can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, without departing from the spirit and teachings of this application, the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways, and all such combinations and / or combinations fall within the scope of this application.

[0144] This document uses specific embodiments to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and core ideas of this application, and are not intended to limit this application. For those skilled in the art, changes can be made to the specific implementation methods and application scope based on the ideas, spirit and principles of this application. Any modifications, equivalent substitutions, improvements, etc., made should be included within the scope of protection of this application.

Claims

1. A method for material recall, characterized in that, The method includes: Set multiple time periods with different durations, where longer time periods include shorter time periods; For each time period, based on the user's search information and the user profile of the user in that time period, the user search profile features for that time period are calculated. Based on the length order of the time period, the user search profile features of the time period are sequentially fused and calculated to obtain the user's search intent input features. The fusion calculation is based on the relevance of the user's search intent input features to the user search profile features of the time period. The user's search intent is input into the intent classification model to obtain the probability value of the user's search intent for different types of materials; Based on the probability values ​​of the user's search intent for different types of materials, recall materials for the user from the candidate material set; The step of sequentially fusing and calculating the user search profile features within the time period to obtain the user's search intent input features includes: a. Based on the length order of the time periods, select the current time period, multiply the concatenated vector of the search information and the user profile of the current time period with the set first weight vector, and then use the activation function to calculate to obtain the first gate value; b. Multiply the concatenated vector of the search information and the user profile of the current time period with the set second weight vector, and then use the activation function to calculate the second gate value; c. The product of the remaining probability value of the first gate value and the user search profile feature of the current time period, and the product of the second gate value and the search intent input feature of the previous time period in the current time period are added together to obtain the sum value, which is used as the search intent input feature of the current time period. d. Determine whether the current time period is the shortest time period within the set time period. If yes, use the search intent input feature of the current time period as the user's search intent input feature. If no, use the next time period in sequence as the current time period and return to step a to continue execution.

2. The method as described in claim 1, characterized in that, The setting includes three time periods with different durations: The time period is divided into three time periods: the first time period, the second time period, and the third time period. The first time period is longer than the second time period, and the second time period is longer than the third time period.

3. The method as described in claim 1, characterized in that, The user search profile features calculated for the time period include: The user's search information and the user profile for the time period are input into a feedforward neural network to calculate the user search profile features for the time period.

4. The method as described in claim 1, characterized in that, The user's search intent is input into the intent classification model to obtain the probability values ​​of the user's search intent for different types of materials, including: The user's search information features are obtained by using a neural network to extract search information features. The user's search information features and search intent input features are input into the first intent classification model, and the output is the probability value of the user's search intent for different types of materials.

5. The method as described in claim 1 or 4, characterized in that, Before obtaining the probability values ​​of the user's search intent for different types of materials, the process also includes: The search intent probability values ​​for different types of materials output by the intent classification model are filtered according to the set filtering rules to obtain the user's search intent probability values ​​for different types of materials.

6. The method as described in claim 1, characterized in that, The method further includes: The materials for the user will be recalled from the alternative material set as the first recalled materials; The search information features of the user are extracted using a neural network. Based on the search information features of the user, a distributed search analysis engine is used to analyze and search from the candidate material set, and recall materials for the user as the second recall materials. Based on the user's search information, the set search rules are used to search from the candidate material set and recall the material for the user as the third recalled material; The first recalled material, the second recalled material, and the third recalled material are merged; Based on the second search intent probability value of the corresponding material type, the merged recall materials are sorted, and the sorted recall materials are distributed to the user.

7. The method as described in claim 6, characterized in that, The second search intent probability value for the corresponding material type includes: The user's search information features are obtained by using a neural network to extract search information features. The user's search information features and search intent input features are input into a second intent classification model composed of a long short-term memory artificial neural network (LSTM) and an attention mechanism, and the output is the probability value of the user's second search intent for different types of materials.

8. A computer product, characterized in that, An electronic device includes: processor; A memory storing a program configured to implement the method as described in any one of claims 1 to 7 when executed by the processor.

9. A non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores instructions, characterized in that, When executed by a processor, the instructions cause the processor to perform the method as described in any one of claims 1 to 7.

Images