A method and device for mining recommended information

Abstract

The present invention discloses a method and device for mining recommendation information, and relates to the field of computer technology. A specific implementation of the method includes: extracting attribute elements of a set attribute type from text description information of a target object to obtain scene attribute information of the target object; combining the attribute elements of the scene attribute information according to a set combination strategy to generate scene combination information including event attribute elements; generating recommendation information of the target object according to the attribute elements of the scene combination information and a set recommendation template. This implementation combines the attributes of the target object itself with the user's intention, which can not only highlight the target object, but also meet the user's needs, improve the recommendation effect, and enhance the user experience.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method and device for mining recommendation information. Background Art

[0002] In the recommendation scenario, recommendation information is important information to assist users in making decisions. Users can decide whether to click, collect, or purchase related target objects based on the recommendation information provided on the website page. In the prior art, there are mainly the following ways to generate recommendation information: Method 1: Operation personnel manually write recommendation information, such as product highlights and promotion information; Method 2: Professionals design several templates for different categories of products, and obtain recommendation information based on the template splicing; Method 3: Extract product review information as recommendation information.

[0003] In the process of implementing the present invention, there are at least the following problems in the prior art:

[0004] The existing recommendation information generation scheme does not take into account the differences in recommendation scenarios, resulting in the same recommendation information generated in different recommendation scenarios, which makes it impossible to achieve personalized recommendations and leads to poor recommendation effects. Summary of the invention

[0005] In view of this, an embodiment of the present invention provides a method and device for mining recommendation information, which obtains scene attribute information of a target object, combines attribute elements of the scene attribute information to obtain scene combination information that takes into account user intent, and then generates recommendation information based on the scene combination information. The above method combines the attributes of the target object itself with the user intent, which can not only highlight the target object, but also meet user needs, improve recommendation effects, and enhance user experience.

[0006] To achieve the above objective, according to one aspect of an embodiment of the present invention, a method for mining recommendation information is provided.

[0007] A method for mining recommendation information in an embodiment of the present invention includes: extracting attribute elements of a set attribute type from text description information of a target object to obtain scene attribute information of the target object; combining the attribute elements of the scene attribute information according to a set combination strategy to generate scene combination information including event attribute elements; and generating recommendation information of the target object according to the attribute elements of the scene combination information and a set recommendation template.

[0008] Optionally, the combination strategy defines the attribute elements included in the generated scene combination information, as well as the number of combinations and the order of combinations of the attribute elements; the generation of scene combination information including event attribute elements includes: screening the initial combination information for semantic rationality, and using the screening result as the scene combination information; wherein the initial combination information is the result obtained by combining the attribute elements of the scene attribute information.

[0009] Optionally, the semantic rationality screening includes perplexity screening and / or similarity screening; wherein, the perplexity screening includes: calculating the perplexity of the initial combination information or the similarity screening result; wherein, the perplexity represents the probability that the attribute elements in the initial combination information constitute a sentence; comparing the perplexity with a set perplexity threshold, and filtering out the initial combination information or similarity screening results whose perplexity is greater than the perplexity threshold; the similarity screening includes: vectorizing the attribute elements of the initial combination information or the perplexity screening result to obtain the corresponding attribute vector; calculating the similarity between the attribute vectors, and comparing the similarity with a set similarity threshold; filtering out the initial combination information or the perplexity screening results whose similarity is less than or equal to the similarity threshold.

[0010] Optionally, there are multiple target objects; after the step of generating scene combination information including event attribute elements, the method further includes: determining the target object having the same scene combination information, and constructing a candidate object set according to the determined target object; calculating the scene combination feature data corresponding to the scene combination information, and calculating the candidate object feature data corresponding to the candidate object set; weightedly fusing the scene combination feature data and the corresponding candidate object feature data to obtain fused feature data; inputting the fused feature data and the corresponding candidate object feature data into a dual-tower model for training, and using the trained dual-tower model to predict the scene combination information of the object to be tested; wherein the object to be tested includes the target object for which the scene combination information fails to be generated.

[0011] Optionally, the calculating the scene combination feature data corresponding to the scene combination information includes: vectorizing the attribute elements in the scene combination information to obtain the corresponding attribute vector; and performing weighted summation on the obtained attribute vector to obtain the scene combination feature data corresponding to the scene combination information.

[0012] Optionally, the calculation of the candidate object feature data corresponding to the candidate object set includes: performing word segmentation vectorization on the text description information of the target object in the candidate object set to obtain the word segmentation vector corresponding to the word segmentation; performing weighted summation on the obtained word segmentation vectors to obtain the target object vector corresponding to the target object; performing weighted summation on the target object vectors corresponding to the target objects in the candidate object set to obtain the candidate object feature data corresponding to the candidate object set.

[0013] Optionally, the use of the trained dual-tower model to predict the scene combination information of the object to be tested includes: performing word segmentation and vectorization on the text description information of the object to be tested to obtain word segmentation vectors corresponding to the word segmentation; performing weighted summation on the obtained word segmentation vectors to obtain the object vector to be tested corresponding to the object to be tested; calculating the similarity between the object vector to be tested and the fused feature data corresponding to the scene combination information, and according to the similarity calculation result, selecting the corresponding scene combination information as the scene combination information of the object to be tested.

[0014] Optionally, the recommendation information includes multiple items; after the step of generating the recommendation information of the target object, the method further includes: obtaining scene combination information generated for the user's historical operation objects, and constructing a scene combination information set; calculating the first scores of attribute elements of different attribute types in the scene combination information set according to the frequency of occurrence of attribute elements in the scene combination information set and the frequency of occurrence of attribute types corresponding to the attribute elements; calculating the second scores of the multiple recommendation information of the target object respectively according to the first scores; and sorting the multiple recommendation information according to the size of the second scores.

[0015] Optionally, there are multiple target objects; obtaining the scene attribute information of the target objects includes: selecting a current target object from the multiple target objects, selecting a current attribute type from the attribute types, and repeating the following steps until the last attribute type of the last target object: counting the frequency of occurrence of attribute elements corresponding to the current attribute type of the current target object in an attribute element set; wherein the attribute element set includes attribute elements corresponding to the current attribute types of multiple target objects; dividing the attribute element set according to the frequency to obtain a high-frequency attribute element set and a low-frequency attribute element set; calculating the correlation between the low-frequency attribute elements of the low-frequency attribute element set and the high-frequency attribute elements of the high-frequency attribute element set; and selecting high-frequency attribute elements from the high-frequency attribute element set according to the correlation, and rewriting the low-frequency attribute elements into the selected high-frequency attribute elements.

[0016] Optionally, extracting attribute elements of a set attribute type from text description information of the target object includes: performing word segmentation on the text description information of the target object to obtain a word segmentation result; performing part-of-speech tagging on the word segmentation result, and obtaining attribute elements of a corresponding attribute type according to a part-of-speech tag of the part-of-speech tagging result.

[0017] To achieve the above objective, according to another aspect of an embodiment of the present invention, a device for mining recommendation information is provided.

[0018] A device for mining recommendation information according to an embodiment of the present invention comprises: an acquisition module for extracting attribute elements of a set attribute type from text description information of a target object to obtain scene attribute information of the target object; a combination module for combining the attribute elements of the scene attribute information according to a set combination strategy to generate scene combination information including event attribute elements; and a generation module for generating recommendation information of the target object according to the attribute elements of the scene combination information and a set recommendation template.

[0019] Optionally, the combination strategy defines the attribute elements included in the generated scene combination information, as well as the number of combinations and the order of combinations of the attribute elements; the combination module is also used to: perform semantic rationality screening on the initial combination information, and use the screening result as the scene combination information; wherein the initial combination information is the result obtained by combining the attribute elements of the scene attribute information.

[0020] Optionally, the semantic rationality screening includes perplexity screening and / or similarity screening; wherein, the perplexity screening includes: calculating the perplexity of the initial combination information or the similarity screening result; wherein, the perplexity represents the probability that the attribute elements in the initial combination information constitute a sentence; comparing the perplexity with a set perplexity threshold, and filtering out the initial combination information or similarity screening results whose perplexity is greater than the perplexity threshold; the similarity screening includes: vectorizing the attribute elements of the initial combination information or the perplexity screening result to obtain the corresponding attribute vector; calculating the similarity between the attribute vectors, and comparing the similarity with a set similarity threshold; filtering out the initial combination information or the perplexity screening results whose similarity is less than or equal to the similarity threshold.

[0021] Optionally, there are multiple target objects; the device also includes: a model prediction module, used to determine the target object with the same scene combination information, and construct a candidate object set according to the determined target object; calculate the scene combination feature data corresponding to the scene combination information, and calculate the candidate object feature data corresponding to the candidate object set; weightedly fuse the scene combination feature data and the corresponding candidate object feature data to obtain fused feature data; input the fused feature data and the corresponding candidate object feature data into a dual-tower model for training, and use the trained dual-tower model to predict the scene combination information of the object to be tested; wherein the object to be tested includes a target object that fails to generate the scene combination information.

[0022] Optionally, the model prediction module is further used to vectorize the attribute elements in the scene combination information to obtain corresponding attribute vectors; and perform weighted summation on the obtained attribute vectors to obtain scene combination feature data corresponding to the scene combination information.

[0023] Optionally, the model prediction module is also used to calculate the candidate object feature data corresponding to the candidate object set, including: performing word segmentation vectorization on the text description information of the target object in the candidate object set to obtain the word segmentation vector corresponding to the word segmentation; performing weighted summation on the obtained word segmentation vectors to obtain the target object vector corresponding to the target object; performing weighted summation on the target object vectors corresponding to the target objects in the candidate object set to obtain the candidate object feature data corresponding to the candidate object set.

[0024] Optionally, the model prediction module is also used to perform word segmentation and vectorization on the text description information of the object to be tested to obtain a word segmentation vector corresponding to the word segmentation; perform weighted summation on the obtained word segmentation vector to obtain an object vector corresponding to the object to be tested; calculate the similarity between the object vector to be tested and the fused feature data corresponding to the scene combination information, and select the corresponding scene combination information as the scene combination information of the object to be tested based on the similarity calculation result.

[0025] Optionally, the recommendation information includes multiple items; the device also includes: a sorting module, used to obtain scene combination information generated for the user's historical operation objects and construct a scene combination information set; according to the frequency of occurrence of attribute elements in the scene combination information set and the frequency of occurrence of attribute types corresponding to the attribute elements, calculate the first score of attribute elements of different attribute types in the scene combination information set; according to the first score, respectively calculate the second score of multiple items of the recommendation information of the target object; according to the size of the second score, sort the multiple items of the recommendation information.

[0026] Optionally, there are multiple target objects; the acquisition module is further used to select a current target object from the multiple target objects, select a current attribute type from the attribute types, and repeat the following steps until the last attribute type of the last target object: counting the frequency of occurrence of attribute elements corresponding to the current attribute type of the current target object in an attribute element set; wherein the attribute element set includes attribute elements corresponding to the current attribute types of multiple target objects; dividing the attribute element set according to the frequency to obtain a high-frequency attribute element set and a low-frequency attribute element set; calculating the correlation between the low-frequency attribute elements of the low-frequency attribute element set and the high-frequency attribute elements of the high-frequency attribute element set; according to the correlation, selecting high-frequency attribute elements from the high-frequency attribute element set, and rewriting the low-frequency attribute elements into the selected high-frequency attribute elements.

[0027] Optionally, the acquisition module is further used to segment the text description information of the target object to obtain a segmentation result; perform part-of-speech tagging on the segmentation result, and obtain attribute elements of the corresponding attribute type according to the part-of-speech tag of the part-of-speech tagging result.

[0028] To achieve the above objective, according to another aspect of an embodiment of the present invention, an electronic device is provided.

[0029] An electronic device according to an embodiment of the present invention includes: one or more processors; a storage device for storing one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement a method for mining recommendation information according to an embodiment of the present invention.

[0030] To achieve the above objective, according to another aspect of the embodiments of the present invention, a computer readable medium is provided.

[0031] A computer-readable medium according to an embodiment of the present invention stores a computer program, and when the program is executed by a processor, a method for mining recommendation information according to an embodiment of the present invention is implemented.

[0032] An embodiment of the above invention has the following advantages or beneficial effects: by acquiring the scene attribute information of the target object and combining the attribute elements of the scene attribute information to obtain scene combination information that takes user intention into consideration, and then generating recommendation information based on the scene combination information, the above method combines the attributes of the target object itself with the user intention, which can not only highlight the target object, but also meet user needs, thereby improving recommendation effects and user experience.

[0033] The initial combination information obtained by combining attribute elements according to the combination strategy is screened to ensure the semantic rationality of the final scene combination information. By calculating the perplexity, from the perspective of the sentence, the initial combination information with a low probability of constituting a sentence is screened out to ensure the semantic rationality of the obtained scene combination information. By calculating the similarity, from the perspective of the word, the initial combination information with a distance between words less than or equal to the similarity threshold is screened out, or the perplexity screening results are further screened to further ensure the semantic rationality of the obtained scene combination information.

[0034] In the case where the scene information cannot be extracted from the text description information, the existing scene combination information and the candidate object set built based on the scene combination information can be used to train the dual-tower model, and then combined with the text description information of the object to be tested to predict its scene combination information, which reduces the generation cost of the recommended information, has high accuracy, and improves the object recall rate. During the model training process, vector weighting is performed to ensure the model training effect.

[0035] By calculating the similarity between the vector of the object to be tested and the fused feature data corresponding to the existing scene combination information, the scene combination information corresponding to the fused feature data with high similarity is selected as the scene combination information of the object to be tested for the vector of the object to be tested, thereby ensuring the accuracy of the generated scene combination information. Based on the scene combination information generated for the user's historical operation objects, multiple recommended information are sorted so that the recommended information that the user is more interested in and is more likely to guide the user's click behavior can be ranked first, further improving the recommendation effect and user experience.

[0036] The attribute element set is divided according to the frequency of occurrence to obtain high-frequency attribute elements and low-frequency attribute elements, and then the low-frequency attribute elements are rewritten to ensure the rationality of the expression of the attribute elements themselves and further improve the user experience. Through word segmentation and part-of-speech tagging, attribute elements can be obtained efficiently and accurately from text description information.

[0037] The further effects of the above-mentioned non-conventional optional manner will be described below in conjunction with the specific implementation manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings are used to better understand the present invention and do not constitute an improper limitation of the present invention.

[0039] Figure 1 is a schematic diagram of the main steps of a method for mining recommendation information according to an embodiment of the present invention;

[0040] Figure 2 is a schematic diagram of the main flow of a method for mining recommendation information according to another embodiment of the present invention;

[0041] Figure 3This is a schematic diagram of the main process of merging basic scene attribute information according to an embodiment of the present invention;

[0042] Figure 4 is a schematic diagram of the main process of sorting recommendation information according to an embodiment of the present invention;

[0043] Figure 5 is a schematic diagram of the main flow of a method for mining recommendation information according to yet another embodiment of the present invention;

[0044] Figure 6 It is a structural schematic diagram of a double-tower model according to an embodiment of the present invention;

[0045] Figure 7 is a schematic diagram of main modules of a device for mining recommendation information according to an embodiment of the present invention;

[0046] Figure 8 is an exemplary system architecture diagram to which embodiments of the present invention may be applied;

[0047] Fig. 9 It is a schematic diagram of the structure of a computer device suitable for implementing the electronic device of the embodiment of the present invention. DETAILED DESCRIPTION

[0048] The following is a description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0049] The technical terms involved in this embodiment are explained below.

[0050] LTP: The full name is Language Technology Platform, which is a Chinese natural language processing platform developed by the Social Computing and Information Retrieval Research Center of Harbin Institute of Technology.

[0051] BiLSTM: Bidirectional long short-term memory network.

[0052] CRF: The full name is Conditional Random Field Algorithm, conditional random field.

[0053] CNN: The full name is Convolutional Neural Networks.

[0054] RNN: The full name is Recurrent Neural Networks.

[0055] LSTM: The full name is Long Short-Term Memory, long short-term memory network

[0056] The main application scenario of the embodiment of the present invention is to generate recommendation reasons (i.e., recommendation information) for a target object described using text. In the embodiment, the target object may be, for example, a product, news, etc. Taking the generation of recommendation reasons for a product as an example, in order to improve the product click-through rate, click-through conversion rate, and user experience, the product is usually described on the website to generate recommendation reasons for the product. As described in the background technology, the existing recommendation reasons are mainly generated by considering the product's own highlights, promotional information, user review data, user search data, etc., and do not combine the product's own attributes with the user's intent.

[0057] Since users usually have certain intentions when browsing the website, this embodiment uses the link of the scene to connect the product with the user's intention, which can not only highlight the product, but also meet the needs of the user, and ensure that the generated recommendation reasons are novel and effective. It should be noted that the scene of this embodiment focuses on the situation where the user uses the target object to perform an activity or behavior (ie, an event). The scene must include events, and optionally, it can also include time, people, places, etc. For example, winter wear, women's wear, winter women's wear, wear, etc. This embodiment is described in detail below.

[0058] Embodiment 1

[0059] Figure 1 FIG. 1 is a schematic diagram of the main steps of a method for mining recommendation information according to an embodiment of the present invention. Figure 1 As shown, the method for mining recommendation information in the embodiment of the present invention mainly includes the following steps:

[0060] Step S101: extracting attribute elements of set attribute types from the text description information of the target object to obtain scene attribute information of the target object. This step is used to obtain scene attribute information required for the scene based on the text description information of the target object. The scene attribute information here refers to elements that can describe a certain activity or behavior, such as time attribute element-winter, location attribute element-outdoor, event attribute element-sports. Attribute elements have attribute types, for example, the attribute type of "winter" is time, and the attribute type of "outdoor" is location.

[0061] In order to extract attribute elements, it is necessary to first segment the text description information of the target object to obtain a segmentation result; then perform part-of-speech tagging on the segmentation result to obtain a part-of-speech tagging result; then obtain the attribute elements of the corresponding attribute type according to the part-of-speech tag in the part-of-speech tagging result. Among them, the word segmentation tool can be jieba word segmentation, Harbin Institute of Technology's LTP, etc. Part-of-speech tagging can be implemented using a part-of-speech tagging model, such as BiLSTM+CRF, BiLSTM+CNN+CRF, TENER and other machine learning models. In an embodiment, the attribute type and the corresponding attribute element can be used as the scene attribute information of the target object, and the scene attribute information includes the attribute elements of the event attribute type.

[0062] Step S102: According to the set combination strategy, the attribute elements of the scene attribute information are combined to generate scene combination information including event attribute elements. The combination strategy is pre-configured, and the combination strategy defines the attribute elements included in the generated scene combination information, as well as the combination quantity and combination order of the attribute elements. According to the combination strategy, the attribute elements in the scene attribute information are reasonably combined in consideration of the user's intention, and the scene combination information can be obtained.

[0063] The number of combinations can be set according to the specific business. For example, for short copy recommendation information, 2-3 scene attribute information can usually be selected; for long copy recommendation information, more scene attribute information can be selected. When combining, in order to avoid the irrationality caused by the position of attribute elements, it is necessary to define the combination order in the combination strategy. The rationality here refers to the rationality of the expression. Taking the Chinese expression as an example, what to do at what time, that is, the time attribute comes first and the event attribute comes later. Otherwise, it is unreasonable.

[0064] At the same time, since the user's actions such as clicking and purchasing the target object have the intention of doing something, and the user's intention needs to be combined when finally generating the recommendation information, if the scene combination information does not contain event attribute elements, such as winter office, winter lady, etc., then the link between the target object's own attributes and the user's intention is not established. Therefore, in order to ensure that the generated scene combination information has the integrity of the scene narrative, each scene combination information must contain event attribute elements. Through this limitation, the link between the target object's own attributes and the user's intention can be established.

[0065] For example, the number of combinations is defined as 2-3, and the order of combinations is "crowd-event", "time-event", and "time-crowd-event". The scene attribute information "crowd-men", "time-autumn and winter", and "event-sports" can be combined into the scene combination information of [men-sports], [autumn and winter-sports], and [autumn and winter-men-sports].

[0066] Step S103: Generate recommendation information for the target object according to the attribute elements of the scene combination information and the set recommendation template. The recommendation template is a pre-set template for generating recommendation information, including a variable part. The variable part is used to fill in the attribute elements of the scene combination information. Preferably, the recommendation template may also include a fixed part. In an embodiment, the fixed part is at least one recommendation word, and the recommendation word can be customized according to the specific business.

[0067] For example, the recommendation template can be: [crowd][event], [crowd][event] Select / Choose it, [time][crowd][event] Good things, etc. The content in [] is the variable part, and "Select / Choose it" and "Good things" are the fixed parts. In this step, the attribute elements of the scene combination information are filled into the variable part of the recommendation template, and the recommendation information of the target object can be automatically generated based on the filled attribute elements and recommendation words. For example, if the scene combination information is [Men-Sports] and the recommendation template is [Crowd][Event] Select / Choose it, then the recommended information is Men's Sports Select / Choose it.

[0068] Embodiment 2

[0069] Figure 2 FIG. 1 is a schematic diagram of the main flow of a method for mining recommendation information according to another embodiment of the present invention. Figure 2 As shown, the method for mining recommendation information in the embodiment of the present invention, which concretizes the target object as a product, mainly includes the following steps:

[0070] Step S201: extract attribute elements of set attribute types from the text description information of multiple products respectively to obtain the scene attribute information of the products. The text description information may be product titles, expert copywriting, user comments, text in product detail page images, various product attributes filled in by merchants, etc. The specific implementation of extracting attribute elements from text description information includes: segmenting the text description information, performing part-of-speech tagging on the segmentation results, and obtaining attribute elements of set attribute types.

[0071] Before part-of-speech tagging, it is necessary to determine the part-of-speech tags that the segmentation results may belong to based on the characteristics of the information to be tagged. For example, for product titles, the possible part-of-speech tags are: brand words, model words, product words, applicable events, applicable time, applicable places, applicable people, other applicable objects, functional attributes, material attributes, style attributes, style attributes, origin attributes, color attributes, taste attributes, specification attributes, etc. Then, all product titles are tagged with part-of-speech based on the part-of-speech tags. In the embodiment, attribute words (i.e., attribute elements) of various attribute types such as people, time, place, origin, event, style, etc. are extracted as basic scene attribute information.

[0072] In a preferred embodiment, since the text description information such as product titles is mostly input manually, there may be typos. In order to ensure the user experience and the accuracy of the word segmentation results, in the embodiment, the typos of the text description information such as product titles can be corrected before word segmentation. Specific typos correction methods include but are not limited to RNN_CRF models, seq2seq models (sequence to sequence models), etc. Specific examples are shown in Table 1:

[0073] Table 1 Example of basic scene attribute information acquisition results

[0074]

[0075]

[0076] After the above processing, a set of attribute elements of different attribute types can be extracted. Because the attribute elements corresponding to an attribute type may contain uncommon colloquial expressions, long sentences, ambiguous attribute elements with low frequency of occurrence, etc. For example, in the crowd attribute elements, "1-2-3-4-5-6 grades", "junior high school students", etc., usually have low frequency of occurrence and are not user-friendly. Among them, the frequency of occurrence of the current attribute element F t The calculation can be used as follows:

[0077] F t =N t / N total , t=1,2,…,n Formula 1

[0078] Where N t is the number of times the current attribute element appears in all the extracted attribute elements, N total is the total number of times all attribute elements with the same attribute type as the current attribute element appear in all attribute elements extracted, and n is the number of attribute elements with the same attribute type as the current attribute element. For example, if the current attribute element is the crowd attribute element "primary school student", its frequency of occurrence is the number of times "primary school student" appears in all attribute elements / the total number of times the crowd attribute element appears in all attribute elements.

[0079] In order to improve the user experience, in the embodiment, similar basic scene attribute information is merged based on the association rules (i.e., the attribute elements with lower frequency of occurrence are rewritten) to obtain the final scene attribute information of the product. Figure 3 Description.

[0080] Step S202: According to the set combination strategy, the attribute elements of the scene attribute information are combined to obtain initial combination information including event attribute elements. The specific implementation of this step is shown in step S102, which will not be repeated here. It can be understood that the scene combination information of Example 1 is the initial combination information here. However, since the scene combination information obtained in Example 1 may have unreasonable combinations, this embodiment further screens it using step S203.

[0081] Step S203: Perform semantic rationality screening on the initial combination information, and use the screening results as the scene combination information. Among them, semantic rationality screening includes perplexity screening and / or similarity screening. Perplexity (Perplexity) represents the probability that the attribute elements in the initial combination information constitute a sentence. The greater the perplexity, the lower the probability that the attribute elements in the initial combination information constitute a sentence, which means that the combination of the scene combination information is more unreasonable, and the scene combination information with a perplexity greater than the perplexity threshold is screened out. The size of the perplexity threshold depends on the balance between the specific business's demand for the accuracy of the scene combination information and the product coverage. The accuracy calculation formula is as follows:

[0082] Correct rate = the number of scene combination information determined to be reasonable / the number of all scene combination information Formula 2

[0083] Product coverage rate = the number of products corresponding to the scenario combination information / the total number of products Formula 3

[0084] Assume that the scene combination information is C = (attribute element 1, attribute element 2, ..., attribute element k ), then the calculation formula of perplexity PP(C) is as follows:

[0085]

[0086] In the formula, P(C) is the probability of scene combination information, k is the number of attribute elements contained in the scene combination information, P(attribute element 1) is the frequency of attribute element 1 appearing in the text description information / the total frequency of the attribute type to which attribute element 1 belongs appearing in the text description information,

[0087] P(attribute element 2|attribute element 1) is the frequency of attribute element 1 and attribute element 2 appearing simultaneously in the text description information / the frequency of attribute element 1 appearing in the text description information. k |Attribute element 1,…,Attribute element k-1 ) is attribute element 1, attribute element 2, ..., attribute element k-1 , attribute elements k Frequency of co-occurrence / attribute element 1, attribute element 2, ..., attribute element k-1The frequency of co-occurrence in the text description information. The frequency here refers to the number of times, and the total frequency refers to the total number of times.

[0088] In the embodiment, the implementation process of perplexity screening includes: firstly calculating the perplexity of the initial combination information according to Formula 4, then comparing the perplexity with a set perplexity threshold, and finally filtering out the initial combination information whose perplexity is greater than the perplexity threshold to obtain a perplexity screening result.

[0089] Similarity is the semantic distance between the attribute elements in the scene combination information. In an embodiment, an unsupervised word vector tool can be used to vectorize the attribute elements in the scene combination information, and then the distance between each vector is calculated to obtain the similarity. Unsupervised word vector tools include but are not limited to tools such as word2vec and fasttext. Vector distances include but are not limited to cosine distance, Euclidean distance, etc. If cosine distance is selected, scene combination information in which the cosine distances between each attribute element are greater than the similarity threshold is screened out. The size of the similarity threshold depends on the balance between the specific business requirements for the accuracy of the scene combination information and the product coverage. The calculation formula for the cosine distance cosθ is as follows:

[0090]

[0091] Where n is the vector dimension of the attribute elements vectorized.

[0092] In the embodiment, the implementation process of similarity screening includes: firstly, vectorizing the attribute elements of the perplexity screening results to obtain the corresponding attribute vectors; then calculating the similarity between the attribute vectors, and comparing the similarity with the set similarity threshold; finally, filtering out the perplexity screening results whose similarity is less than or equal to the similarity threshold, retaining the perplexity screening results whose similarity is greater than the similarity threshold, and obtaining the similarity screening results. The following is an example.

[0093] If the vector dimension is set to 50, the vectorized representation of the scene attribute information "men", "sports", and "pregnancy" is shown in Table 2. After calculation, the semantic distance of the scene combination information [men, sports] is 0.5244, and the semantic distance of the scene combination information [men, pregnancy] is 0.2964. The semantic distance shows that the rationality of the scene combination information [men, sports] is significantly better than that of the scene combination information [men, pregnancy].

[0094] Table 2 Results of vectorized representation of scene attribute information

[0095]

[0096] It can be understood that perplexity filtering is to filter the combination results from the perspective of sentences, and similarity filtering is to filter the combination results from the perspective of words. The two are parallel and independent processes. Therefore, you can perform perplexity filtering first, and then perform similarity filtering (as shown in the above example, the basic data for similarity filtering is the perplexity filtering result), and the similarity filtering result is the final scene combination information. You can also perform similarity filtering first, and then perform perplexity filtering (at this time, the basic data for similarity filtering is the initial combination information, and the basic data for perplexity filtering is the similarity filtering result), and the perplexity filtering result is the final scene combination information. You can also perform only one of perplexity filtering and similarity filtering to obtain the final scene combination information.

[0097] Step S204: Generate product recommendation information based on the attribute elements of the scene combination information and the set recommendation template. This step generates recommendation information of a set granularity (such as a short sentence granularity) for the scene combination information based on the recommendation template, as shown in Table 3. In a preferred embodiment, the recommendation information can also be personalized and rewritten in combination with the user portrait to enhance the user experience. For example, in the second row of Table 3, if the user portrait indicates that the user is 18-25 years old, the word "lady" can be rewritten as an attribute word that is more in line with the temperament of the crowd, such as girl or fairy. In addition, the number of words in the recommendation information can be limited according to the application scenario of the recommendation information.

[0098] Table 3 Example of generating scenario-based recommendation information

[0099]

[0100] Each product will have a variety of different combinations of recommended information. For example, in the third row of Table 3, the generated recommended information can be "good winter outfits for petite people", "good winter outfits", or "good outfits for petite people", and the recommended words "good things" can also change. Moreover, an attribute type of a product can have multiple attribute elements. For example, the event attribute element in Table 4 can be "sports", "running", or "rewritten outfits", etc. In a preferred embodiment, the recommended information that the user is more interested in and is more likely to lead the user to click can be placed in front of the recommendation sequence, see step S205.

[0101] Step S205: Calculate the first scores of the attribute elements of different attribute types in the recommended information, and sort the recommended information according to the calculated first scores. The basis for sorting includes but is not limited to the user's historical browsing, clicking or searching information and other behavioral data. For example, based on the list of products that the user has recently clicked, the corresponding scene combination information is mined for the listed products in turn through steps S201 to S203, and then the scores of each attribute element of each attribute type are calculated based on the frequency of occurrence, and then the attribute elements of the recommended information are averaged and weighted and summed, or the weights of each attribute element of each attribute type are determined according to business needs, and then weighted and summed. For specific implementation, see about Figure 4 Description.

[0102] Figure 3 FIG. 2 is a schematic diagram of the main process of merging basic scene attribute information in an embodiment of the present invention. Figure 3 As shown, the main process of merging basic scene attribute information in an embodiment of the present invention includes the following steps:

[0103] Step S301: Select a current target object from multiple target objects, and select a current attribute type from set attribute types. The target object in the embodiment is a product.

[0104] Step S302: Count the frequency of the attribute elements corresponding to the current attribute type of the current target object in the attribute element set. The attribute element set includes the attribute elements corresponding to the current attribute types of multiple target objects. The frequency here is the number of times the attribute elements corresponding to the current attribute type appear in the attribute element set.

[0105] Step S303: Divide the attribute element set according to the frequency of occurrence to obtain a high-frequency attribute element set and a low-frequency attribute element set. In an embodiment, the number of occurrences of each attribute element in the attribute element set can be sorted, and the attribute elements with a higher ranking (such as the first 10%, 15%) are selected to be added to the high-frequency attribute element set, and the attribute elements with a lower ranking (such as the last 10%, 15%) are selected to be added to the low-frequency attribute element set.

[0106] Step S304: Calculate the correlation between the low-frequency attribute elements of the low-frequency attribute element set and the high-frequency attribute elements of the high-frequency attribute element set. The correlation is calculated as follows:

[0107] score (correlation degree) = P (high-frequency attribute element i | low-frequency attribute element j) * P (low-frequency attribute element j | high-frequency attribute element i) Formula 6

[0108]

[0109]

[0110] In the formula, F(high-frequency attribute element i, low-frequency attribute element j) is the frequency of simultaneous occurrence of high-frequency attribute element i and low-frequency attribute element j, F(low-frequency attribute element j) is the total frequency of occurrence of low-frequency attribute element j, and F(high-frequency attribute element i) is the total frequency of occurrence of high-frequency attribute element i.

[0111] Step S305: According to the correlation, a high-frequency attribute element is selected from the high-frequency attribute element set, and the low-frequency attribute element is rewritten with the selected high-frequency attribute element. The greater the correlation, the more similar the high-frequency attribute element and the low-frequency attribute element are, so the high-frequency attribute element with the greatest correlation with the low-frequency attribute element can be selected for rewriting. That is, the low-frequency attribute element is replaced with the high-frequency attribute element with the greatest correlation with the low-frequency attribute element.

[0112] Step S306: Repeat steps S301 to S305 until the last attribute type of the last target object, that is, determine whether the current attribute type is the last attribute type of the last target object. If not, repeat steps S301 to S305, otherwise end this process.

[0113] Figure 4 FIG. 1 is a schematic diagram of the main process of sorting recommendation information in an embodiment of the present invention. Figure 4 As shown, the main process of sorting recommendation information in the embodiment of the present invention includes the following steps:

[0114] Step S401: Obtain the scene combination information generated for the user's historical operation object and construct a scene combination information set. The historical operation object is the product that the user browsed, clicked, searched, etc. on the website within a set time period. This step requires using steps S201 to S203, or using steps S501 to S507 to mine the scene combination information corresponding to the product to obtain a scene combination information set.

[0115] Step S402: Calculate the first score of the attribute elements of different attribute types in the scene combination information set according to the frequency of occurrence of the attribute elements in the scene combination information set and the frequency of occurrence of the attribute types corresponding to the attribute elements. The calculation formula of the first score of the current attribute element is as follows:

[0116]

[0117] Step S403: Calculate the second scores of multiple recommendation information of the target object according to the first scores. The target object here is the product to be recommended, which is provided by the demander who needs to conduct the recommendation business. The second score of the current recommendation information of the product to be recommended is the sum of the first scores of the attribute elements belonging to the current recommendation information. The specific calculation formula is as follows:

[0118]

[0119] Step S404: Sort the multiple recommendation information according to the second score. Sort the multiple recommendation information of the product to be recommended according to the second score from high to low. The recommendation information sorting result of the embodiment is shown in Table 4.

[0120] Table 4 Example of recommended information ranking results

[0121]

[0122]

[0123] As shown in Table 4, for the recommendation information "Spring Fairy Sports Selection", in the recommendation information in the second column, the frequency of "Spring" is 1, the attribute type of "Spring" is "time", and the total frequency of occurrence is 1, so the first score of "Spring" is 1. The frequency of "Fairy" is 2, the attribute type of "Fairy" is crowd, and the total frequency of occurrence is 4, so the first score of "Fairy" is 2 / 4. The frequency of "Sports" is 2, the attribute type of "Sports" is event, and the total frequency of occurrence is 6, so the first score of "Sports" is 2 / 6, and the second score of the recommendation information "Spring Fairy Sports Selection" is 1+2 / 4+2 / 6=1.83.

[0124] Embodiment 3

[0125] Since the text description information of the product may not contain the concept of scene, the scene combination information cannot be obtained through Example 1 and Example 2. In order to recall more products, in this embodiment, a machine learning model is trained based on the existing scene combination information to predict the scene combination information of the product. In the embodiment, the machine learning model includes but is not limited to the deep semantic matching model DSSM (Deep Structured Semantic Models), DSSM model variants such as CNN-DSSM, LSTM-DSSM, or other machine learning models. The following is a detailed description.

[0126] Figure 5 FIG. 1 is a schematic diagram of the main flow of a method for mining recommendation information according to another embodiment of the present invention. Figure 5 As shown, the method for mining recommendation information in the embodiment of the present invention, which concretizes the target object as a product, mainly includes the following steps:

[0127] Step S501: extracting attribute elements of set attribute types from text description information of multiple products respectively to obtain scene attribute information of the products.

[0128] Step S502: combining the attribute elements of the scene attribute information according to the set combination strategy to obtain initial combination information including event attribute elements.

[0129] Step S503: screening the initial combination information for semantic rationality, and using the screening result as the scene combination information.

[0130] The specific implementation of the above steps S501 to S503, and the following steps S508 and S509 refer to the corresponding steps S201 to S205, which will not be described in detail in this embodiment.

[0131] Step S504: Determine products with the same scene combination information, and construct a candidate object set based on the determined products. This step uses the scene combination information as a keyword to perform aggregation and construct a candidate object set for products. The aggregation is to treat products with the same scene combination information as a set. For example, if the scene combination information of product 1 and product 2 is both "winter outfit", then product 1 and product 2 are aggregated into the set corresponding to winter outfit.

[0132] In the embodiment, a scene combination information and its corresponding candidate object set constitute a seed, and the seed can be expressed in the following manner: scene combination information\t product 1, product 2, ... For example: winter wear\t product 1Id, product 2Id, product 3Id, .... Id is an identifier.

[0133] In addition, in order to avoid the product scene in the candidate object set being single or mixed, the initial object set obtained by aggregation needs to be screened. Among them, single scene means that the products in the initial object set are single, such as all from the same store. Mixed scene means that the products in the initial object set are mixed, such as from multi-level classification categories.

[0134] Specifically, to avoid a single scene, you can filter based on the number of stores and products contained in the initial object set. For example, filter out stores with less than 2 and less than 20 products. To avoid mixed scenes, you can filter based on product words (such as mobile phones and refrigerators) or the number of categories at all levels. For example, keep products with less than 5 product words in the same first-level category.

[0135] Step S505: Calculate the scene combination feature data corresponding to the scene combination information, and calculate the candidate object feature data corresponding to the candidate object set. The calculation process of the scene combination feature data corresponding to the scene combination information is as follows: vectorize the attribute elements in the scene combination information to obtain the corresponding attribute vector; perform weighted summation on the obtained attribute vector to obtain the scene combination feature data corresponding to the scene combination information. The specific calculation formula of the scene combination feature data is as follows:

[0136]

[0137] Where V 场景组合信息 That is, the scene combination feature data corresponding to the scene combination information, V 属性要素 is the attribute vector corresponding to the attribute element, and n(attribute element|seed S) is the total number of attribute elements under seed S.

[0138] The calculation process of the candidate object feature data corresponding to the candidate object set is as follows: perform word segmentation vectorization on the text description information of the products in the candidate object set to obtain the word segmentation vector corresponding to the word segmentation; perform weighted summation on the obtained word segmentation vector to obtain the product vector corresponding to the product; perform weighted summation on the product vectors corresponding to the products in the candidate object set to obtain the candidate object feature data corresponding to the candidate object set. The specific calculation formula for the candidate object feature data is as follows:

[0139]

[0140] Where V 候选对象集合 That is, the candidate object feature data corresponding to the candidate object set, V 分词 That is, the word segmentation vector corresponding to the word segmentation, n(word segmentation|product Pro) is the total number of word segmentations in the text description information of a product Pro in the candidate object set, and n(product|seed S) is the total number of products under seed S.

[0141] In the above implementation process, the word segmentation method and unsupervised word vector tool of step S201 can be used to perform word segmentation and vectorization (including word segmentation and vectorization) on the text description information. The vectorization can also be obtained by encoding by one-hot, word2vec or fasttext method and then performing feature splicing.

[0142] Step S506: weighted fusion of the scene combination feature data and the corresponding candidate object feature data to obtain fused feature data. In this step, the fused feature data is obtained by fusing the scene combination feature data and the candidate object feature data, so that the fused feature data can contain both user intention level and product level information at the same time, and the expression is more comprehensive. The specific calculation formula of the fused feature data is as follows:

[0143]

[0144] Where V 种子S That is, the fusion feature data corresponding to the current scene combination information, 1 / 2 is the weight, it can be understood that the weight value is only an example, V 场景组合信息 and V 候选对象集合 The weights can be different.

[0145] Step S507: input the fused feature data and the corresponding candidate object feature data into the dual-tower model for training, and use the trained dual-tower model to predict the scene combination information of the object to be tested. Figure 6 Schematic diagram of the structure of the double tower model of the embodiment of the present invention. Figure 6 As shown in the figure, the dual-tower model uses two neural networks to train the training data separately, which is divided into input layer, representation layer and matching layer.

[0146] Among them, the two inputs of the input layer are fusion feature data and candidate object feature data. The two different input information are trained according to the corresponding deep learning networks. The number of network layers and the number of neurons in the two deep learning networks may be different, but the final output dimension must be consistent. The deep learning network includes but is not limited to neural network (NN), convolutional neural network (CNN), recurrent neural network (RNN) and other deep learning networks. After model training, the final scene combination information embedding (vector) and product embedding can be obtained and stored in the memory database.

[0147] When generating scene combination information for the object to be tested, firstly, the text description information of the object to be tested is vectorized by word segmentation to obtain the word segmentation vector corresponding to the word segmentation; then, the obtained word segmentation vector is weighted summed to obtain the object vector to be tested corresponding to the object to be tested; then, the similarity (such as cosine distance) between the object vector to be tested and the fused feature data corresponding to all the scene combination information is calculated, and according to the similarity calculation result, the corresponding scene combination information is selected as the scene combination information of the object to be tested. In the embodiment, the scene combination information corresponding to the fused feature data with the greatest similarity is selected as the scene combination information of the object to be tested.

[0148] In a preferred embodiment, if the scene combination information predicted by the twin-tower model is different from the scene combination information extracted from the text description information of the product, the extracted scene combination information is selected as the final scene combination information.

[0149] Step S508: Generate product recommendation information based on the attribute elements of the scene combination information and the set recommendation template.

[0150] Step S509: Calculate the first scores of the attribute elements of different attribute types in the recommended information, and sort the recommended information according to the calculated first scores.

[0151] It can be seen from the mining method of recommendation information in the embodiment of the present invention that the embodiment can customize the recommendation information suitable for the product according to different product characteristics, with low cost and good diversity, and can provide users with product scenario application information, and can also hit the user's current scenario application needs to impress the user. At the same time, based on the scenario recommendation information, the user can compare his or her needs with the actual product characteristics provided, change and strengthen the user's positive view of the product, make him or her accept the recommendation results and click, collect or purchase, etc., so that the recommendation results look more friendly, which greatly improves the user experience.

[0152] Figure 7 FIG. 1 is a schematic diagram of the main modules of the device for mining recommendation information according to an embodiment of the present invention. Figure 7 As shown, the recommended information mining device 700 of the embodiment of the present invention mainly includes:

[0153] The acquisition module 701 is used to extract the attribute elements of the set attribute type from the text description information of the target object to obtain the scene attribute information of the target object. This module is used to obtain the scene attribute information required for the scene based on the text description information of the target object. The scene attribute information here refers to the elements that can describe a certain activity or behavior, such as the time attribute element-winter, the location attribute element-outdoors, and the event attribute element-sports. The attribute element has an attribute type, for example, the attribute type of "winter" is time, and the attribute type of "outdoors" is location.

[0154] In order to extract attribute elements, it is necessary to first segment the text description information of the target object to obtain a segmentation result; then perform part-of-speech tagging on the segmentation result to obtain a part-of-speech tagging result; and then obtain the attribute elements of the corresponding attribute type according to the part-of-speech tags in the part-of-speech tagging result. In an embodiment, the attribute type and the corresponding attribute elements can be used as the scene attribute information of the target object, and the scene attribute information includes the attribute elements of the event attribute type.

[0155] The combination module 702 is used to combine the attribute elements of the scene attribute information according to the set combination strategy to generate scene combination information including event attribute elements. The combination strategy is pre-configured, and the combination strategy defines the attribute elements included in the generated scene combination information, as well as the combination quantity and combination order of the attribute elements. According to the combination strategy, the attribute elements in the scene attribute information are reasonably combined in consideration of the user's intention, so as to obtain the scene combination information.

[0156] The number of combinations can be set according to the specific business. For example, for short copy recommendation information, 2-3 scene attribute information can usually be selected; for long copy recommendation information, more scene attribute information can be selected. When combining, in order to avoid the irrationality caused by the position of attribute elements, it is necessary to define the combination order in the combination strategy. The rationality here refers to the rationality of the expression. Taking the Chinese expression as an example, what to do at what time, that is, the time attribute comes first and the event attribute comes later. Otherwise, it is unreasonable.

[0157] At the same time, since the user's actions such as clicking and purchasing the target object have the intention of doing something, and the user's intention needs to be combined when finally generating the recommendation information, if the scene combination information does not contain event attribute elements, such as winter office, winter lady, etc., then the link between the target object's own attributes and the user's intention is not established. Therefore, in order to ensure that the generated scene combination information has the integrity of the scene narrative, each scene combination information must contain event attribute elements. Through this limitation, the link between the target object's own attributes and the user's intention can be established.

[0158] The generation module 703 is used to generate the recommendation information of the target object according to the attribute elements of the scene combination information and the set recommendation template. The recommendation template is a pre-set template for generating recommendation information, including a variable part. Among them, the variable part is used to fill in the attribute elements of the scene combination information. Preferably, the recommendation template may also include a fixed part. In an embodiment, the fixed part is at least one recommendation word, and the recommendation word can be customized according to the specific business. The attribute elements of the scene combination information are filled into the variable part of the recommendation template, and the recommendation information of the target object can be automatically generated based on the filled attribute elements and recommendation words.

[0159] In addition, the recommendation information mining device 700 of the embodiment of the present invention may also include: a model prediction module and a sorting module ( Figure 7 ). The model prediction module is used to determine the target object having the same scene combination information, and to construct a candidate object set according to the determined target object; to calculate the scene combination feature data corresponding to the scene combination information, and to calculate the candidate object feature data corresponding to the candidate object set; to perform weighted fusion on the scene combination feature data and the corresponding candidate object feature data to obtain fused feature data; to input the fused feature data and the corresponding candidate object feature data into a twin-tower model for training, and to use the trained twin-tower model to predict the scene combination information of the object to be tested; wherein the object to be tested includes the target object that fails to generate the scene combination information.

[0160] A sorting module is used to obtain scene combination information generated for the user's historical operation objects and construct a scene combination information set; calculate the first score of the attribute elements of different attribute types in the scene combination information set according to the frequency of occurrence of the attribute elements in the scene combination information set and the frequency of occurrence of the attribute types corresponding to the attribute elements; calculate the second score of the multiple recommended information of the target object according to the first score; and sort the multiple recommended information according to the size of the second score.

[0161] From the above description, it can be seen that by obtaining the scene attribute information of the target object and combining the attribute elements of the scene attribute information to obtain scene combination information that takes user intention into consideration, and then generating recommendation information based on the scene combination information, the above method combines the attributes of the target object itself with user intention, which can not only highlight the target object, but also meet user needs, thereby improving recommendation effects and enhancing user experience.

[0162] Figure 8 An exemplary system architecture 800 is shown to which the method for mining recommendation information or the apparatus for mining recommendation information according to the embodiments of the present invention can be applied.

[0163] like Figure 8 As shown, system architecture 800 may include terminal devices 801, 802, 803, a network 804, and a server 805. Network 804 is used to provide a medium for communication links between terminal devices 801, 802, 803 and server 805. Network 804 may include various connection types, such as wired, wireless communication links, or optical fiber cables, etc.

[0164] Users can use terminal devices 801, 802, and 803 to interact with server 805 through network 804 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 801, 802, and 803, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0165] The terminal devices 801 , 802 , and 803 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0166] The server 805 may be a server that provides various services, such as a background management server that processes the text description information sent by the administrator using the terminal devices 801, 802, and 803. The background management server may extract attribute elements from the text description information, generate scene combination information, generate recommendation information, and feed back the processing results (such as the generated recommendation information) to the terminal device.

[0167] It should be noted that the method for mining recommendation information provided in the embodiment of the present application is generally executed by the server 805 , and accordingly, the device for mining recommendation information is generally set in the server 805 .

[0168] It should be understood that Figure 8 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0169] According to an embodiment of the present invention, the present invention also provides an electronic device and a computer-readable medium.

[0170] The electronic device of the present invention comprises: one or more processors; a storage device for storing one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement a method for mining recommendation information according to an embodiment of the present invention.

[0171] The computer-readable medium of the present invention stores a computer program, and when the program is executed by a processor, a method for mining recommendation information according to an embodiment of the present invention is implemented.

[0172] Reference below Fig. 9 , which shows a schematic diagram of the structure of a computer system 900 suitable for implementing an electronic device of an embodiment of the present invention. Fig. 9 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0173] like Fig. 9 As shown, the computer system 900 includes a central processing unit (CPU) 901, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 902 or a program loaded from a storage part 908 into a random access memory (RAM) 903. In the RAM 903, various programs and data required for the operation of the computer system 900 are also stored. The CPU 901, the ROM 902, and the RAM 903 are connected to each other via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0174] The following components are connected to the I / O interface 905: an input section 906 including a keyboard, a mouse, etc.; an output section 907 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a LAN card, a modem, etc. The communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to the I / O interface 905 as needed. A removable medium 911, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 910 as needed, so that a computer program read therefrom is installed into the storage section 908 as needed.

[0175] In particular, according to the embodiments disclosed in the present invention, the process described in the main step diagram above can be implemented as a computer software program. For example, the embodiments disclosed in the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the main step diagram. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 909, and / or installed from the removable medium 911. When the computer program is executed by the central processing unit (CPU) 901, the above-mentioned functions defined in the system of the present invention are executed.

[0176] It should be noted that the computer-readable medium shown in the present invention may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present invention, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0177] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0178] The modules involved in the embodiments of the present invention may be implemented by software or hardware. The modules described may also be set in a processor, for example, they may be described as: a processor includes an acquisition module, a combination module and a generation module. The names of these modules do not constitute a limitation on the modules themselves in certain cases. For example, the acquisition module may also be described as "a module for extracting attribute elements of a set attribute type from the text description information of the target object to obtain the scene attribute information of the target object".

[0179] As another aspect, the present invention also provides a computer-readable medium, which may be included in the device described in the above embodiment; or may exist independently without being assembled into the device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by a device, the device includes: extracting attribute elements of set attribute types from the text description information of the target object to obtain the scene attribute information of the target object; combining the attribute elements of the scene attribute information according to the set combination strategy to generate scene combination information including event attribute elements; generating recommendation information of the target object according to the attribute elements of the scene combination information and the set recommendation template.

[0180] According to the technical solution of the embodiment of the present invention, by acquiring the scene attribute information of the target object and combining the attribute elements of the scene attribute information to obtain the scene combination information taking the user intention into consideration, and then generating the recommendation information based on the scene combination information, the above method combines the attributes of the target object itself with the user intention, which can not only highlight the target object, but also meet the user needs, thereby improving the recommendation effect and the user experience.

[0181] The above product can execute the method provided by the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided by the embodiment of the present invention.

[0182] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for mining recommendation information, characterized in that: include: Extracting attribute elements of set attribute types from text description information of a target object to obtain scene attribute information of the target object; wherein there are multiple target objects; Combining the attribute elements of the scene attribute information according to the set combination strategy to generate scene combination information including event attribute elements; Determine a target object having the same scene combination information, and construct a candidate object set according to the determined target object; calculate scene combination feature data corresponding to the scene combination information, and calculate candidate object feature data corresponding to the candidate object set; perform weighted fusion on the scene combination feature data and the corresponding candidate object feature data to obtain fused feature data, so as to predict the scene combination information of the object to be tested by using a preset dual-tower model; The recommendation information of the target object is generated according to the attribute elements of the scene combination information and the set recommendation template.

2. The method according to claim 1, characterized in that The attribute elements included in the scene combination information generated by the combination strategy definition, and the combination quantity and combination order of the attribute elements; The generating of scene combination information including event attribute elements includes: The initial combination information is screened for semantic rationality, and the screening result is used as the scene combination information; wherein the initial combination information is a result obtained by combining the attribute elements of the scene attribute information.

3. The method according to claim 2, characterized in that The semantic rationality screening includes perplexity screening and / or similarity screening; wherein, The perplexity screening comprises: Calculating the perplexity of the initial combination information or the similarity screening result; wherein the perplexity represents the probability that the attribute elements in the initial combination information constitute a sentence; Comparing the perplexity with a set perplexity threshold, and filtering out initial combination information or similarity screening results whose perplexity is greater than the perplexity threshold; The similarity screening includes: Vectorizing the attribute elements of the initial combination information or the perplexity screening result to obtain a corresponding attribute vector; Calculating the similarity between the attribute vectors, and comparing the similarity with a set similarity threshold; The initial combination information or the perplexity screening results whose similarity is less than or equal to the similarity threshold are filtered out.

4. The method according to claim 1, characterized in that: The fused feature data and the corresponding candidate object feature data are input into a dual-tower model for training, and the trained dual-tower model is used to predict scene combination information of the object to be tested; wherein the object to be tested includes a target object that fails to generate the scene combination information.

5. The method according to claim 4, characterized in that The calculating the scene combination feature data corresponding to the scene combination information includes: Vectorizing the attribute elements in the scene combination information to obtain a corresponding attribute vector; A weighted sum is performed on the obtained attribute vectors to obtain scene combination feature data corresponding to the scene combination information.

6. The method according to claim 4, characterized in that The calculating the candidate object feature data corresponding to the candidate object set includes: Performing word segmentation vectorization on the text description information of the target object in the candidate object set to obtain a word segmentation vector corresponding to the word segmentation; Performing weighted summation on the obtained word segmentation vectors to obtain a target object vector corresponding to the target object; The target object vectors corresponding to the target objects in the candidate object set are weightedly summed to obtain the candidate object feature data corresponding to the candidate object set.

7. The method according to claim 4, characterized in that The method of using the trained dual-tower model to predict scene combination information of the object to be tested includes: Vectorize the text description information of the object to be tested to obtain a word segmentation vector corresponding to the word segmentation; Performing weighted summation on the obtained word segmentation vectors to obtain a test object vector corresponding to the test object; The similarity between the vector of the object to be detected and the fused feature data corresponding to the scene combination information is calculated, and according to the similarity calculation result, the corresponding scene combination information is selected as the scene combination information of the object to be detected.

8. The method according to claim 1, characterized in that The recommendation information includes multiple; After the step of generating the recommendation information of the target object, the method further includes: Obtaining scene combination information generated for the user's historical operation objects, and constructing a scene combination information set; Calculating first scores of attribute elements of different attribute types in the scene combination information set according to the frequency of occurrence of attribute elements in the scene combination information set and the frequency of occurrence of attribute types corresponding to the attribute elements; Calculating second scores of the plurality of recommendation information of the target object respectively according to the first scores; The plurality of recommendation information are sorted according to the size of the second score.

9. The method according to claim 1, characterized in that: There are multiple target objects; The step of obtaining scene attribute information of the target object includes: Select a current target object from the multiple target objects, select a current attribute type from the attribute types, and repeat the following steps until the last attribute type of the last target object: Counting the frequency of occurrence of the attribute element corresponding to the current attribute type of the current target object in the attribute element set; wherein the attribute element set includes a plurality of attribute elements corresponding to the current attribute type of the target object; According to the frequency, the attribute element set is divided to obtain a high-frequency attribute element set and a low-frequency attribute element set; Calculating the correlation between the low-frequency attribute elements of the low-frequency attribute element set and the high-frequency attribute elements of the high-frequency attribute element set; According to the association degree, a high-frequency attribute element is selected from the high-frequency attribute element set, and the low-frequency attribute element is rewritten as the selected high-frequency attribute element.

10. The method according to any one of claims 1 to 9, characterized in that The step of extracting attribute elements of a set attribute type from the text description information of the target object includes: Segment the text description information of the target object to obtain the segmentation result; Part-of-speech tagging is performed on the word segmentation result, and attribute elements of corresponding attribute types are obtained according to the part-of-speech tags of the part-of-speech tagging result.

11. A device for mining recommended information, characterized in that: include: An acquisition module, used to extract attribute elements of a set attribute type from text description information of a target object, and obtain scene attribute information of the target object; wherein there are multiple target objects; A combination module, used to combine the attribute elements of the scene attribute information according to a set combination strategy to generate scene combination information including event attribute elements; A model prediction module is used to determine a target object having the same scene combination information, and to construct a candidate object set according to the determined target object; to calculate scene combination feature data corresponding to the scene combination information, and to calculate candidate object feature data corresponding to the candidate object set; to perform weighted fusion on the scene combination feature data and the corresponding candidate object feature data to obtain fused feature data, so as to predict the scene combination information of the object to be tested by using a preset dual-tower model; A generating module is used to generate the recommendation information of the target object according to the attribute elements of the scene combination information and the set recommendation template.

12. An electronic device, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 10.

13. A computer readable medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 10 is implemented.

Images