Product recommendation method and apparatus, and computer device and storage medium

Abstract

A product recommendation method, which is executed by a computer device. The method comprises: a server separately acquiring comment information that has been posted by a user entity for at least one piece of information, and separately acquiring attribute information sets respectively corresponding to the at least one piece of information (S201); for each piece of information, the server extracting a corresponding description target set on the basis of at least one of an attribute information set and comment information that correspond to the information (S202); for each description target set, the server performing semantic analysis on comment information associated with the description target set, in order to obtain a sentiment tendency result corresponding to the comment information, and separately determining, on the basis of the sentiment tendency result, the degree of preference of the user entity for at least one description target comprised in the description target set (S203); and on the basis of the obtained degree of preference for each description target, selecting, from an associated product set corresponding to the description target, a recommended product for the user entity (S204).

Description

Product recommendation method and device, computer device and storage medium

[0001] Related applications

[0002] The present application claims priority to the Chinese patent application No. 202411059849.6, filed on August 3, 2024, and entitled "A product recommendation method, device, computer device and storage medium", the contents of which are hereby incorporated by reference in its entirety. TECHNICAL FIELD

[0003] The present application relates to the technical field of computer, and in particular, to a product recommendation method, device, computer device and storage medium. BACKGROUND

[0004] Currently, in the product recommendation scenario, in order to realize personalized recommendation, it is usually necessary to first understand the interests and hobbies of the object, determine which products the object is interested in, and then set a targeted recommendation scheme for the object.

[0005] In related technologies, the emotional tendency of the object to a product can only be captured when the object expresses a clear attitude towards the product. For example, after the object purchases a product, the object's interest in the product can be determined from the object's use evaluation of the product, or the object's likes and dislikes of the product, etc.

[0006] However, the amount of such information that can be obtained is small, is limited by privacy requirements, and is difficult to obtain. Moreover, the object will not generate such information for products that the object has not actually contacted, so many product categories cannot be effectively covered, resulting in a smaller number of product categories involved in personalized recommendation, greater limitations of the recommendation, and lower accuracy. SUMMARY

[0007] The present application provides a product recommendation method, device, computer device and storage medium.

[0008] In one aspect, a product recommendation method is provided, comprising:

[0009] Respectively obtaining comment information of a user object for at least one information information published, and respectively obtaining attribute information sets corresponding to the at least one information information;

[0010] For each information information, based on at least one of the attribute information set and the comment information corresponding to the information information, a corresponding description target set is extracted, each description target representing a content subject capable of triggering the user object to generate an emotional tendency;

[0011] respectively determine a preference degree of the user object for at least one description target included in the one description target set based on the sentiment tendency result; and

[0012] based on the obtained preference degree for each description target, select a recommended product for the user object in an associated product set corresponding to each description target.

[0013] In an aspect, a product recommendation apparatus is provided, comprising:

[0014] an obtaining unit configured to respectively obtain comment information published by a user object for at least one information information, and respectively obtain an attribute information set corresponding to each of the at least one information information;

[0015] an extracting unit configured to, for each of the information information, extract a corresponding description target set based on at least one of the attribute information set and the comment information corresponding to the information information, each description target representing a content subject capable of triggering a sentiment tendency of the user object;

[0016] a determining unit configured to, for each of the description target set, respectively perform semantic analysis on comment information associated with the description target set, obtain a sentiment tendency result corresponding to the comment information, and respectively determine a preference degree of the user object for at least one description target included in the one description target set based on the sentiment tendency result; and

[0017] a selecting unit configured to, based on the obtained preference degree for each description target, select a recommended product for the user object in an associated product set corresponding to each description target.

[0018] In an aspect, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the above methods when executing the computer program.

[0019] In an aspect, a computer storage medium is provided, which stores a computer program, wherein the computer program is executed by a processor to implement the steps of any of the above methods.

[0020] In an aspect, a computer program product is provided, which comprises a computer program stored in a computer readable storage medium. A processor of a computer device reads the computer program from the computer readable storage medium, and the processor executes the computer program, so that the computer device executes the steps of any of the above methods.

[0021] Other features and advantages of the present application will be set forth in the following specification, and in part will be apparent from the description, or can be learned by practice of the application. The purposes and other advantages of the present application will be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of the disclosed drawings.

[0023] FIG. 1 is a schematic diagram of an application scenario of a product recommendation method provided by an embodiment of the present application;

[0024] FIG. 2 is a whole flowchart of a product recommendation method provided by an embodiment of the present application;

[0025] FIG. 3 is a schematic diagram of video meta information provided by an embodiment of the present application;

[0026] FIG. 4 is a schematic diagram of comment information provided by an embodiment of the present application;

[0027] FIG. 5 is a schematic diagram of a window provided by an embodiment of the present application;

[0028] FIG. 6 is a schematic diagram of a connection relationship provided by an embodiment of the present application;

[0029] FIG. 7 is a schematic diagram of logic for determining a sentiment tendency result based on a BERT model provided by an embodiment of the present application;

[0030] FIG. 8 is a schematic diagram of comment information preprocessing provided by an embodiment of the present application;

[0031] FIG. 9 is a schematic diagram of a self-attention mechanism provided by an embodiment of the present application;

[0032] FIG. 10 is a schematic diagram of determining a description target for a sentiment tendency provided by an embodiment of the present application;

[0033] FIG. 11 is another schematic diagram of determining a description target for a sentiment tendency provided by an embodiment of the present application;

[0034] FIG. 12 is a schematic diagram of logic for product recommendation based on a neural network provided by an embodiment of the present application;

[0035] FIG. 13 is a schematic diagram of a product recommendation system provided by an embodiment of the present application;

[0036] Fig. 14 is a schematic diagram of a structural composition of a product recommendation device according to an embodiment of the present application;

[0037] Fig. 15 is a schematic diagram of a hardware composition of a computer device according to an embodiment of the present application;

[0038] Fig. 16 is a schematic diagram of a hardware composition of another computer device according to an embodiment of the present application. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0040] In order to facilitate understanding of the technical solutions provided by the embodiments of the present application, some key terms used in the embodiments of the present application are explained first:

[0041] Meta information: information used to describe information, which can reflect the source, structure, characteristics, etc. of the information, for example, the meta information of a video can include title, description, subtitle, partition, etc., which are used to describe the content and attributes of the video; the meta information of a post can include title, topic label, etc., which are used to describe the content and attributes of the post.

[0042] Softmax: an activation function commonly used in machine learning and deep learning, mainly used in multi-classification scenarios. It can map a set of input values to a set of output values, which can be regarded as the probability of each category, and the sum of the probabilities of all categories is 1.

[0043] Description target: extracted from the attribute information set or the comment information, described by the corresponding information information or the corresponding comment information, and capable of triggering a content subject that can trigger the user object to generate an emotional tendency; the description targets extracted from the attribute information set and the comment information corresponding to the same information information can be the same or different.

[0044] Word segmentation refers to the units obtained by performing word segmentation (i.e. splitting) on a paragraph, a sentence, etc. according to a preset rule. The preset rule can be set according to actual needs, for example, an English text can be segmented in a manner of taking a word as a unit, or segmented in a manner of taking a Chinese word as a unit, etc. In the present application, it is mainly divided into attribute word segmentation and comment word segmentation. The attribute word segmentation is obtained by performing word segmentation on the attribute information set, and the comment word segmentation is obtained by performing word segmentation on the comment information.

[0045] In the scenario of the present application, for the information of the English text, the preset rule can be to split in a manner of taking one word as a unit, while taking into account the abbreviations, hyphen, and the like in English, such as "don't" as a whole word, and "mother-in-law" as a word. For the information of the Chinese text, the preset rule can split in a manner of taking a Chinese word as a unit, and a mature Chinese word segmentation tool, such as the Jieba word segmentation tool, can be used to split the Chinese sentence into word units according to the default word segmentation strategy. For the information containing multiple languages, the information can be first distinguished according to the language type, and then the word segmentation processing is performed according to the corresponding preset rule.

[0046] Comment sub-word segmentation: the comment sub-word segmentation obtained by performing secondary word segmentation processing on the comment word segmentation that meets the preset composition condition, for example, taking an English word as an example, if the preset composition condition is a word composed of a root and an affix, then the word "loving" can be secondary word segmented into "lov" and "ing", and "lov" and "ing" are two comment sub-words.

[0047] Next, the technical idea of the embodiments of the present application is briefly described.

[0048] The premise of realizing personalized recommendation is to understand the preferences of the object, and the related technology is often to capture and analyze the emotional tendency of the object to the product through the information that the object expresses a clear attitude to the product after using the product. For example, in the evaluation "I have tried this toothbrush, and it is very useful", it can be captured that the object has a strong positive emotion to the toothbrush product.

[0049] However, since the amount of the above information that can be obtained is small, the preferences of the object that can be analyzed therefrom are very limited, and the object often does not produce the above information for the products that the object has not actually contacted, so many product categories cannot be effectively covered, so that the product categories involved in the personalized recommendation are less, the recommendation is limited, and the accuracy is low.

[0050] Based on this, the embodiments of the present application provide a product recommendation method and device, computer equipment and storage medium, which can obtain a large amount of information and comment information published by a user object under various information, for each information, according to the attribute information corresponding to the information and the comment information obtained from the information, extract the content subject described by the information and the comment information, that is, the description target, and analyze the emotional tendency of the user object in the comment information, further determine which description target the emotional tendency of the user object is directed to, so as to determine the preference degree of the user object to the description target, and further analyze the preference degree of the user object to other description targets.

[0051] The application can improve the accuracy of the judgment of the user's preference by analyzing a large amount of comment information; and even if the user has not actually contacted the description target in the information, the user can still express the evaluation and emotion for the above-mentioned uncontacted description target, and finally the application can recommend the product associated with the description target interested by the user based on the acquired emotional tendency of the user to each description target, so that the preference of the user acquired by the application can widely cover multiple product categories.

[0052] The preferred embodiments of the application are described below in conjunction with the accompanying drawings of the specification, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the application, and are not used to limit the application, and the embodiments in the application and the features in the embodiments can be combined with each other without conflict.

[0053] As shown in FIG. 1, it is a schematic diagram of the application scenario of the embodiment of the application. The application scenario diagram includes two terminal devices 101 and one server 102.

[0054] In the embodiment of the application, the terminal device 101 includes but is not limited to mobile phones, tablet computers, notebook computers, desktop computers, e-book readers and other devices; the server 102 can be a standalone physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content distribution networks (CDN), big data and artificial intelligence platforms and other basic cloud computing services.

[0055] It should be noted that the method in each of the embodiments of the present application can be executed by a computer device, which can be the terminal device 101 or the server 102, i.e., the method can be executed by the terminal device 101 or the server 102 alone, or can be executed by the terminal device 101 and the server 102 together. For example, when the server 102 is the main execution subject, the terminal device 101 acquires the comment information published by the user object under each information information, and sends the attribute information set corresponding to each information information and each comment information to the server 102; the server 102 extracts the corresponding description target set from the attribute information set corresponding to each information information and the comment information; and based on the sentiment tendency result obtained by performing semantic analysis on the comment information, the preference degree of the user object for each description target is determined respectively; then, the server 102 selects a recommended product for the user object in the associated product set corresponding to each description target based on the obtained preference degree for each description target, and sends the recommended product to the terminal device 101; in addition, in the above process, the process of extracting the description target set, determining the preference degree of the user object for each description target, and selecting a recommended product for the user object based on the preference degree of the user object for each description target can also be completed by the terminal device 101, or part of the process is implemented by the terminal device 101 and the other part is implemented by the server 102. In actual application, specific configuration can be made according to the situation, which is not limited in the present application.

[0056] The server 102 and the terminal device 101 can each include one or more processors, memories, and I / O interfaces, etc. In addition, the server 102 can also be configured with a database, which can be used to store attribute information sets, comment information, etc. The memories of the server 102 and the terminal device 101 can also store program instructions required for the respective execution of the product recommendation method provided by the embodiments of the present application, which can be used to implement the product recommendation process provided by the embodiments of the present application when executed by the processor.

[0057] It should be noted that when the product recommendation method provided by the embodiments of the present application is executed by the server 102 or the terminal device 101 alone, the application scenario described above can only include a single device of the server 102 or the terminal device 101, or the server 102 and the terminal device 101 can be considered as the same device. Of course, in actual application, when the product recommendation method provided by the embodiments of the present application is executed by the server 102 and the terminal device 101 together, the server 102 and the terminal device 101 can also be the same device, i.e., the server 102 and the terminal device 101 can be different functional modules of the same device, or virtual devices virtualized by the same physical device.

[0058] In a possible implementation, the user object can provide input content, such as comment information for certain information information, through the terminal device 101, so that the server 102 can receive the comment information of the user object, and then determine the preference of the user object according to the comment information by using the method of the embodiment of the application, and recommend corresponding products to the user object according to the preference, and finally return the products to be recommended to the terminal device 101 for presentation.

[0059] In the embodiment of the application, the terminal device 101 and the server 102 can be directly or indirectly connected in communication through one or more networks 103. The network 103 can be a wired network or a wireless network. For example, the wireless network can be a mobile cellular network or a wireless fidelity (WIFI) network. Of course, the network can also be other possible networks, and the embodiment of the application does not make any limitation in this regard. It should be noted that the number of terminal devices and servers shown in FIG. 1 is only for illustration, and the number of terminal devices and servers is not limited in the embodiment of the application.

[0060] The product recommendation method provided by the exemplary embodiment of the application will be described below in combination with the application scenarios described above and with reference to the accompanying drawings. It should be noted that the above-mentioned application scenarios are only for facilitating the understanding of the spirit and principles of the application, and the embodiments of the application are not limited in this regard.

[0061] Referring to FIG. 2, an implementation flowchart of a product recommendation method provided by the embodiment of the application is shown. The method is executed by a computer device, which can be a terminal or a server. The specific implementation flow of the method is as follows:

[0062] S201: The server respectively acquires comment information published by a user object for at least one information information, and respectively acquires attribute information sets corresponding to the at least one information information.

[0063] The information information refers to an information carrier that can be acquired and spread, has a certain content value, and can reflect the situation of a specific field. The information information can include videos, audios, pictures, texts, etc., and can cover sports, painting, music, etc. For example, the video can be a movie, a TV series, an animation, a short video, etc.; the audio can be music, a radio program, etc.; the picture can be a comic, a photography work, etc.; and the text can be an electronic article, an electronic book, a post, etc.

[0064] Since the information information is various in type and various in content, it can cover various fields such as sports, painting, music, games, handicraft, and technology. By acquiring the public information information and the public comment information published by the user object under the information information, the preference of the user object can be more comprehensively and accurately determined.

[0065] In an embodiment, for the information information of the video type, the server can obtain the meta information of the video, such as the title, the introduction, the subtitles, the tags, the partitions, etc. as the attribute information set by calling the open interface of the video platform. For the information information of the audio type, if it is music, the server can obtain the name of the song, the singer, the album, the genre, etc. from the music platform as the attribute information set; if it is a radio program, the program name, the host, the theme tag, etc. can be obtained. For the information information of the picture type, such as a comic, the server can obtain the title of the comic, the author, the theme tag, etc. from the comic website; for the photographic work, the shooting location, the shooting time, the theme tag, etc. can be obtained. For the information information of the text type, such as an electronic article, the server can obtain the title of the article, the author, the keywords, the abstract, etc. from the article publishing platform; for an electronic book, the book name, the author, the publisher, the classification tag, etc. can be obtained; for a post, the title, the topic tag, etc. can be obtained as the attribute information set.

[0066] It can be understood that in the specific embodiments of the present application, if the data such as the comment information related to the user object is involved, the use of the data needs the user object to know and further needs the active authorization of the user object; when the above data is applied to specific products or technologies in the present application, the permission or consent of the user object needs to be obtained, and the collection, use and processing of the relevant data need to comply with the relevant laws, regulations and standards of the country and region.

[0067] An attribute information set corresponding to one information information contains at least one attribute information. The attribute information set is a set associated with a certain information information, which is composed of attribute information used to describe the characteristics of the information information. The attribute information is often in the form of meta information, which can reflect the source, structure and characteristics of the information. The specific form of the attribute information can be meta information, which is information used to describe the information, and can reflect the source, structure, characteristics, etc. of the information. For example, the meta information of a video can include the title, the introduction, the subtitles, the tags, the partitions, etc. for describing the content and attributes of the video, the meta information of a post can include the title, the topic tag, etc. for describing the content and attributes of the post, and so on. Taking a video as an example, as shown in FIG. 3, it is a video meta information diagram provided by an embodiment of the present application. The introduction of the video shown in the figure contains the title “What is the experience of a person going on a trip”, the introduction “Prepare for a long time, make a lot of strategies, finally get on the plane to Z city last week! This video records the whole process of my travel, I hope you like it”, the partition “life area”, and the tags “travel”, “travel shooting”, “Z city”. The content corresponding to the above title, introduction, partition, tag is the meta information of the video, and the above meta information can be stored as four attribute information in the attribute information set corresponding to the video.

[0068] The comment information can be a comment published by the user object under a video, music, article, a discussion on a post, or a video pop-up, etc. Taking a video as an example, as shown in FIG. 4, a comment information diagram provided by an embodiment of the present application is shown. The comment information published by the user object a under the comment of the video, "I am very envious. I also want to experience it when I have time", and the pop-up "It is too beautiful" published in the video can be regarded as the comment information published by the user object a on the video.

[0069] That is, in a possible implementation, each type of information about the comment or the pop-up published by the user object a can be acquired. For each information, each meta information of the information is taken as each attribute information, and at least one of the comment or the pop-up published by the user object a on the information is taken as the comment information.

[0070] It can be understood that, when the data such as the comment information related to the user object, the attribute information of the information published by others, etc. is used, the user object and the publisher of the information need to be known, and further, the active authorization of the user object and the publisher is needed, and the data is deleted in time after being used and is not saved.

[0071] S202: The server extracts a corresponding description target set based on at least one of the attribute information set corresponding to the information and the comment information for each information.

[0072] In the above, each description target represents a content subject capable of triggering the user object to generate an emotional tendency.

[0073] In an embodiment of the present application, since the comment information published by the user object and the content described by the information do not necessarily completely coincide, in order to determine the emotional tendency expressed in the comment information is on what content, the main content, that is, the content subject, described in the information and the corresponding comment information needs to be determined for each information. For example, the content subject of the article "Spending three days to evaluate seven kinds of notebook computers" is "notebook computer evaluation"; the content subject of the video "Professional trainer teaches you how to learn basketball from zero" is "basketball teaching", etc. Similarly, the content subject of the comment information "This snack looks very delicious" is "snack", and the content subject of the comment information "I was really excited when I watched a basketball game live" is "basketball game". These content subjects are also the corresponding description targets.

[0074] In addition, the description target can also be directly a product or a product category. If the description target is a product or a product category, the preference degree of the user object on the product or the product category can be directly determined when the emotional tendency is analyzed based on the comment information.

[0075] In a possible implementation, the main content described by the information information and the comment information, i.e., the description target, can be analyzed by keyword extraction. For example, the server can process the attribute information set corresponding to the information information to obtain the description target corresponding to the information information, and process the comment information to obtain the description target corresponding to the comment information.

[0076] Taking an information information as an example, the specific implementation process of keyword extraction can be as follows: the server performs word segmentation on the attribute information set corresponding to the information information and the comment information respectively to obtain each word segment; for each word segment, based on the reference evaluation value of the word segment and the reference evaluation value of at least one other word segment that has a connection relationship with the word segment, an overlapping evaluation value of the word segment is obtained; the overlapping evaluation value is taken as a new reference evaluation value of a word segment, and the operation of obtaining the overlapping evaluation value is repeatedly performed until a preset iteration stop condition is reached; the overlapping evaluation value after iteration is stopped is taken as a target evaluation value corresponding to a word segment; at least one keyword that satisfies a keyword selection condition is determined from the word segments; based on the word meaning of the at least one keyword, the at least one keyword is classified, and each keyword that meets a preset classification is taken as a description target.

[0077] The reference evaluation value is an initial evaluation value preset for each word segment during keyword extraction, and is the basis for subsequent calculation of the overlapping evaluation value. The overlapping evaluation value is a new evaluation value calculated according to the reference evaluation value of a word segment and the reference evaluation values of other word segments that have a connection relationship with the word segment during keyword extraction, and will be used as a new reference evaluation value for subsequent iteration. The target evaluation value is the final evaluation value determined by the overlapping evaluation value after the iteration calculation of keyword extraction is stopped, and is used to determine whether a word segment is a keyword.

[0078] The keyword selection condition is a basis for screening keywords from word segments. A threshold value can be set according to the size of the target evaluation value, or the first k word segments can be selected as keywords according to the order of the target evaluation values. The keyword selection condition can be set according to the size of the target evaluation value. A threshold value of the target evaluation value can be preset, and when the target evaluation value of a word segment is greater than the threshold value, the word segment is determined as a keyword. In addition, the word segments can also be sorted according to the target evaluation values, and the first k word segments are selected as keywords. The value of k can be adjusted according to actual conditions, for example, the value of k can be appropriately increased for longer texts, and the value of k can be reduced for shorter texts.

[0079] In one embodiment, for each word, the positional distance between the word and each other word can be determined respectively, the superposition evaluation value of the word can be obtained based on the reference evaluation value of the word and the reference evaluation value of at least one other word which has a connection relationship with the word, the superposition evaluation value is taken as the new reference evaluation value of the word, the operation of obtaining the superposition evaluation value is repeated until a preset iteration stopping condition is reached (the reference evaluation value of the word and the reference evaluation value of each other word are preset in the first iteration), and the superposition evaluation value after iteration is stopped is taken as the target evaluation value corresponding to the word. The connection relationship refers to a relationship between two words of the same word type, the positional distance of which satisfies a preset distance condition, and is used to calculate the superposition evaluation value of the word in the keyword extraction process.

[0080] In the keyword extraction process of the attribute information set and the comment information, the word processing is performed respectively, the word type of the word corresponding to the attribute information set is an attribute word, and the word type of the word corresponding to the comment information is a comment word. Correspondingly, the selected keyword in the attribute word is an attribute keyword, and the selected keyword in the comment word is a comment keyword.

[0081] The word processing is to divide a piece of text into multiple units based on a preset rule, each unit can be a word, a word, a phrase, etc., and the specific rule can be set according to actual needs, for example, an English text can be divided in a way of taking a word as a unit, etc., which is not limited in the present application.

[0082] Then, the server extracts keywords from the preprocessed words. For example, assuming that a comment information is processed by word, the comment words f1, f2, f3, …, fg arranged in order according to the position in the comment information can be obtained, each comment word is regarded as a node, and the server needs to determine whether there is a connection relationship between the nodes according to the positional distance between the nodes. Specifically, if the positional distance between two comment words satisfies a preset distance condition, it is determined that there is a connection relationship between the two comment words.

[0083] For example, a window size can be set, as shown in FIG. 5, which is a window diagram provided by an embodiment of the present application. Assuming that the window size is h, and 2

[0084] An initial reference evaluation value is preset for each comment word, for example, the initial reference evaluation value of each comment word is 1; for each comment word, the superimposed evaluation value of the comment word is calculated according to the reference evaluation value of the comment word itself and the reference evaluation values of other comment words connected with the comment word.

[0085] For example, the superimposed evaluation value can be the sum of the reference evaluation values.

[0086] For another example, a weight can also be given to each connection relationship, and the superimposed evaluation value is obtained according to the reference evaluation value and the weight. Specifically, a weight can be given to each connection relationship. Assuming that a word w and other n words w1, w2, …, wn connected with the word w are respectively, the reference evaluation values of the n words are v1, v2, …, vn, and the weights of the connection relationships are k1, k2, …, kn, then the calculation formula of the superimposed evaluation value V of the word w is n The corresponding reference evaluation values are v1, v2, …, vn. n The weights of the connection relationships are k1, k2, …, kn. n The calculation formula of the superimposed evaluation value V of the word w is Wherein v is the reference evaluation value of the word w itself.

[0087] After obtaining the superimposed evaluation value of each comment word, the superimposed evaluation value is taken as the new reference evaluation value of the corresponding comment word, or the superimposed evaluation value is normalized based on the out-degree of the comment word, and the normalized value is taken as the new reference evaluation value. The out-degree is the number of other comment words connected with a comment word. The calculation formula of the normalization can be Wherein V norm is the normalized superimposed evaluation value, V is the superimposed evaluation value, and d is the out-degree of the comment word.

[0088] Based on the new reference evaluation value, the above steps are repeated until a preset iteration number is reached, or when the difference between the superimposed evaluation value and the reference evaluation value obtained in the current iteration is within a preset difference range, the iteration is stopped. The superimposed evaluation value of each comment word obtained in the last iteration, that is, the target evaluation value corresponding to each comment word; according to the target evaluation value of each comment word, the comment words are sorted, and the first k comment words can be selected as the comment keywords.

[0089] The method of extracting attribute keywords in the attribute information set is the same as the method of extracting comment keywords, which will not be repeated here; each attribute information in the attribute information set can be extracted as a keyword individually, or can be integrated into a target attribute information after being processed (such as splicing, removing repeated parts, etc.), and then the keyword is extracted, which is not limited in the present application.

[0090] It can be seen that the more the connection relationship of a segmented word is, the more other segmented words it is associated with, the more critical it is in the text, and the higher the superposition evaluation value is. As shown in FIG. 6, it is a connection relationship diagram provided by an embodiment of the present application, in which the connection relationship of "linear", "system", "solution", and "minimum" is more, that is, the possibility of these segmented words being finally selected as keywords is greater.

[0091] In addition, considering that some comment information may not contain a description target, such as "this is too good-looking", "beautiful ah", and the like, the above extraction method can extract keywords regardless of whether the comment information contains a content subject, for example, "this is too good-looking" can finally extract the keywords "good-looking", and "beautiful ah" can finally extract the keyword "beautiful", therefore, the extracted comment keywords need to be screened; in addition, the attribute keywords can also be screened to further ensure the accuracy of keyword extraction.

[0092] The specific method and standard of screening can be set from multiple aspects. From the aspect of word type, as described above, the keywords under the noun and noun phrase are taken as the description target, and the keywords of other word types such as adjectives and verbs are filtered out. From the aspect of semantics, some keywords with too broad meaning and lack of clear directionality, such as "things" and "matter", can be filtered out. Some common meaningless words (such as "of", "is", "in", and the like) can also be removed from the keywords in combination with a stop word list. In addition, keywords with too low frequency of occurrence, which may be accidental and not representative, can also be screened out. The stop word list is a list containing common meaningless words (such as "of", "is", "in", and the like), which is used to remove these words from the keywords in the keyword extraction process to improve the accuracy of keyword extraction.

[0093] In the aspect of word type screening, only nouns and noun phrases are specified to be reserved as the description target. For some verbs or adjectives with special meanings, if they can be clearly directed to a specific content subject in a specific context, they can also be reserved. For example, "running" in "running shoes" is a verb, but it has a clear direction after being combined with "shoes", and can be used as a description target. In the aspect of semantic screening, a semantic range threshold is set for keywords with too broad meanings. The semantic range threshold is a numerical value set in the semantic screening link of keyword screening. By calculating the cosine similarity between the semantic vector of the keyword and a set of reference vectors with clear direction, if the cosine similarity is lower than the semantic range threshold, it is considered that the keyword has too broad a meaning and is filtered out to ensure that the extracted keywords have clear direction. The cosine similarity between the semantic vector of the keyword and a set of reference vectors with clear direction can be calculated. If the cosine similarity is lower than 0.3, it is considered that the keyword has too broad a meaning and is filtered out. For keywords with too low frequency, the number of occurrences of the keyword in all information and comment information is counted. If the number of occurrences is lower than the preset low proportion threshold of the total information, such as 0.1%, it is considered that the keyword is accidentally appeared and has no representativeness, and is screened out. The preset low proportion threshold of the total information is a proportion numerical value set when screening keywords with too low frequency. The number of occurrences of the keyword in all information and comment information is counted. When the number of occurrences is lower than the preset low proportion threshold of the total information, it is considered that the keyword is accidentally appeared and has no representativeness, and is screened out.

[0094] An optional implementation is that the server can classify at least one keyword according to the word meaning of the keyword, such as adjectives, nouns, verbs, etc., and take each keyword meeting a preset classification as a description target, for example, take the keywords under nouns and noun phrases as description targets.

[0095] In addition to keyword extraction, other methods can also be used to obtain the description target corresponding to the information, for example, if the information is a video, the content subject corresponding to the information can also be obtained by using deep learning image recognition, speech recognition and semantic understanding technology, that is, the audio content of the video is subjected to speech recognition and semantic understanding, and the description target (i.e. the content subject) corresponding to the video is extracted, and multiple image frames are obtained by extracting multiple frames of the video, and the content in the image frames is recognized by image recognition technology, that is, a multi-modal method is used to extract the description target, which can comprehensively analyze the video content from multiple angles.

[0096] In terms of image recognition, a target detection algorithm based on convolutional neural network (CNN) can be selected, such as Faster R-CNN or YOLO series algorithm. First, multiple frames are extracted from the video to obtain multiple image frames, and the image frames are input into the trained CNN model, which can identify objects, scenes, and other information in the image. To improve the accuracy of recognition, a large amount of image data can be used to pre-train and fine-tune the model. In terms of speech recognition, an end-to-end speech recognition model such as DeepSpeech or Wav2Vec 2.0 can be used. The audio content of the video is extracted and pre-processed, and then input into the speech recognition model, which converts the audio signal into text information. To ensure the accuracy of speech recognition, a large-scale speech dataset can be used to train the model, and a language model can be used to post-process the recognition results. In terms of semantic understanding, a pre-trained language model based on the Transformer architecture, such as BERT or GPT, can be used. The text information obtained by speech recognition is input into the language model, which performs semantic analysis on the text and extracts the description target (i.e., the content subject) corresponding to the video. Through multi-modal methods, the results of image recognition, speech recognition, and semantic understanding are fused to comprehensively analyze the video content from multiple perspectives.

[0097] In addition, after the word segmentation processing, before calculating the target evaluation value, the obtained words can be filtered and pre-processed according to actual needs to remove invalid words, etc. For example, some conjunctions (such as and in English, and in Chinese), prepositions (such as in, at in English, in, from in Chinese), auxiliary words (such as am in English, in Chinese) and other words that are used for grammatical structure or logical connection but express weak meaning can be removed.

[0098] Whether a word is an invalid word can be considered from multiple aspects. From the aspect of word type, some conjunctions (such as and in English, and in Chinese), prepositions (such as in, at in English, in, from in Chinese), auxiliary words (such as am in English, in Chinese) and other words that are used for grammatical structure or logical connection but express weak meaning can be considered as invalid words. From the aspect of semantics, some words with too broad meaning and lack of clear directionality, such as "things" and "things", can also be considered as invalid words. The specific algorithm for filtering can use the method of stop word table matching, and a stop word table is pre-established to include common invalid words. After word segmentation, each word is matched with the stop word table, and if the word is in the stop word table, it is filtered out. In addition, the frequency of the word can also be filtered, and words with too high or too low frequency can be meaningless filler words or accidental words.

[0099] S203: The server respectively performs semantic analysis on the comment information associated with each set of description targets, obtains a sentiment tendency result corresponding to the comment information, and respectively determines the preference degree of the user object for at least one description target included in one set of description targets based on the sentiment tendency result.

[0100] In the embodiments of the present application, in addition to determining the content subject described by the information information and the comment information, the sentiment tendency of the user object also needs to be determined.

[0101] An optional implementation is that the server can obtain each comment subword obtained after the comment information is processed by the word segmentation in S202, and respectively perform secondary word segmentation processing on each comment subword whose subword structure meets a preset composition condition to obtain a corresponding comment subword.

[0102] In the above, the preset composition condition can be set based on actual needs. For example, for compound words or words composed of roots and affixes, secondary word segmentation processing can be performed. For example, the word "loving" can be secondary segmented into "lov" and "ing", and "lov" and "ing" are two comment subwords. The word "airport" can be secondary segmented into "air" and "port", and "air" and "port" are also two comment subwords.

[0103] Compared with word segmentation, the subwords obtained by further splitting the word segmentation can capture more fine-grained language features and structures, and can better generalize and respond to changes in different text scenarios. If a word that has not appeared before is encountered, word segmentation usually cannot be processed, but subword segmentation can solve this problem by combining existing subwords, thereby improving processing capability and effect.

[0104] After obtaining the comment subwords and the comment subwords, each comment subword and each comment subword that has not been processed by secondary word segmentation is used as a reprocessing subword. The sentiment tendency result corresponding to the comment information is obtained based on the first association degree of each reprocessing subword and each other reprocessing subword.

[0105] The first association degree is the association degree between each reprocessing subword and other reprocessing subwords output by the self-attention mechanism when determining the sentiment tendency result of the comment information, which reflects the context features of the reprocessing subword.

[0106] The above process can be completed by a model, for example, a Bidirectional Encoder Representation from Transformers (BERT) model from a transformer, as shown in FIG. 7, which is a logic diagram for determining a sentiment tendency result based on a BERT model according to an embodiment of the present application. The overall steps are as follows: the review information is preprocessed and input into the BERT model, in the output of the BERT model, the output value (i.e., the output vector) corresponding to [CLS] is input into a fully connected layer neural network, thereby classifying the sentiment to obtain the sentiment tendency result corresponding to the review information.

[0107] Specifically, the fully connected layer neural network maps the output value corresponding to the [CLS] symbol to a lower-dimensional space. For the output of the fully connected neural network, normalization calculation can be performed to convert the output into a probability distribution, thereby obtaining a vector representing the probabilities of positive, neutral, and negative sentiments. Finally, according to the probability distribution, the class with the highest probability is selected as the prediction result, and finally the sentiment tendency result corresponding to the review information is obtained as positive or negative or neutral sentiment.

[0108] In one embodiment, the sentiment tendency result corresponding to the review information is obtained based on the first correlation degree of each reprocessed word to each other reprocessed word. The specific process is as follows: the review information is preprocessed and input into the BERT model, the BERT model encodes the input text through a multi-layer Transformer encoder to capture the dependency between words and generate context-dependent word vectors. The core of the BERT model is a multi-layer Transformer encoder, each layer of which contains a self-attention mechanism and a feedforward neural network. The self-attention mechanism enables the model to calculate the first correlation degree of each unit in the input sequence with other units in the sequence, thereby capturing long-distance dependencies in the text. The feedforward neural network further nonlinearly transforms the output of the self-attention. In the output of the BERT model, the output value (i.e., the output vector) corresponding to [CLS] is input into a fully connected layer neural network, thereby classifying the sentiment to obtain the sentiment tendency result corresponding to the review information. Specifically, the fully connected layer neural network maps the output value corresponding to the [CLS] symbol to a lower-dimensional space. For the output of the fully connected neural network, normalization calculation can be performed to convert the output into a probability distribution, thereby obtaining a vector representing the probabilities of positive, neutral, and negative sentiments. Finally, according to the probability distribution, the class with the highest probability is selected as the prediction result, and finally the sentiment tendency result corresponding to the review information is obtained as positive or negative or neutral sentiment.

[0109] The above process will be described in detail below.

[0110] First, the pre-processing process of adding special symbols and converting formats to the comment information, the segmentation processing and the secondary segmentation processing mentioned in the foregoing can be regarded as a part of the pre-processing of the comment information, or the comment information can be segmented in advance to obtain the comment segmentation and the comment sub-segmentation, and then the following steps are executed.

[0111] The connector is added to the comment sub-segmentation originally belonging to a comment segmentation, for example, after "language" is divided into "lang" and "uage", ## is added to "uage" to obtain "##uage"; a special [CLS] symbol is added at the beginning of the comment information, the [CLS] symbol is a pre-defined fixed symbol, which does not have any meaning itself, and is used to mark the starting position of a group of texts, and a special [SEP] symbol is added at the end of each sentence of the text, indicating the boundary between each sentence.

[0112] In addition to the comment segmentation and the comment sub-segmentation, [CLS] and [SEP] can also be regarded as re-processing segmentation.

[0113] In the present application, one comment information can be regarded as a group of texts; in other application scenarios, if multiple groups of texts need to be input into the model at the same time, [CLS] is a separator between groups, and the model processes each group of texts as a unit, that is, a group of texts marked by [CLS] can obtain one output result; in the present application, one or more comment information can be input at a time.

[0114] Then, the comment segmentation and the comment sub-segmentation are replaced by the corresponding indexes in the vocabulary table to meet the input format of BERT, and the replaced comment segmentation, comment sub-segmentation and added special symbols are combined into texts, and the position information and segmentation information of each comment segmentation, comment sub-segmentation and special symbol in the comment information are combined as the input of the BERT model, wherein the segmentation information represents that the comment segmentation, comment sub-segmentation or special symbol is located in which sentence of the comment information, and the position information represents the ordering position of the comment segmentation, comment sub-segmentation or special symbol in the comment information.

[0115] As shown in FIG. 8, an opinion information preprocessing schematic diagram provided by an embodiment of the present application is shown, assuming that the opinion information is my dog is cute, he likes playing, after the word segmentation processing, the secondary word segmentation processing and the adding symbol operation, [CLS], my, dog, is, cute, [SEP], he, likes, play, ##ing, [SEP] are obtained. Among them, the part before "he" belongs to the first sentence in the opinion information, "he" itself and the part after "he" belong to the second sentence in the opinion information, therefore, the opinion segmentation before "he" and the segmentation information corresponding to the special symbol are consistent, recorded as E A , the opinion segmentation, the opinion sub-segmentation and the segmentation information corresponding to the special symbol after "he" are consistent, recorded as E B ; according to the position of each opinion segmentation, opinion sub-segmentation and special symbol in the opinion information, the position information corresponding to [CLS], my, dog, is, cute, [SEP], he, likes, play, ##ing, [SEP] is determined respectively, recorded as E1, E2, E3, E4, E5, E6, E7, E8, E9, E 10 , E 11 .

[0116] Then, the mapping E [CLS] , E my , E dog , E is , E cute , E [SEP] , E he , E likes , E play , E ##ing , E [SEP] of [CLS], my, dog, is, cute, [SEP], he, likes, play, ##ing, [SEP] in the vocabulary table is determined respectively, which is combined with the corresponding position information and segmentation information to be the model input input after each other; specifically: Input i = Token i + Seg i + Pos i

[0117] Among them, Input i is the input corresponding to the i-th unit (that is, the i-th reprocessed segmentation) in the opinion information, which can be an opinion segmentation, an opinion sub-segmentation or a special symbol, Token i is the mapping of the i-th unit in the opinion information in the vocabulary table, Seg i is the segmentation information of the i-th unit in the opinion information, Posi Position information of the i-th unit in the review information.

[0118] After inputting the above content into the BERT model, the BERT model encodes the input text through a multi-layer Transformer encoder, captures the dependency between words and generates context-related word vectors.

[0119] The core of the BERT model is a multi-layer Transformer encoder. Each layer contains a self-attention mechanism (English: Self-Attention) and a feed-forward neural network (English: Feed Forward Neural Network, FNN). The self-attention mechanism enables the model to calculate the first correlation of each unit in the input sequence with other units in the sequence, thereby capturing long-distance dependencies in the text; the feed-forward neural network further transforms the output of the self-attention.

[0120] As shown in FIG. 9, it is a schematic diagram of a self-attention mechanism provided by an embodiment of the present application. Following the assumption in FIG. 8, Input i is the input corresponding to the i-th unit in the review information, i∈[1,11]; taking the input Input1 corresponding to the first unit in the review information as an example, the calculation of the self-attention mechanism involves three types of weight matrices (query matrix Qi, key matrix Ki and value matrix Vi). The result of multiplying Input1 and the query matrix Q1, and the result of multiplying Input1 and the key matrix K1 can be combined to obtain the first attention score a 1,1 corresponding to Input1; further, the result of multiplying Input1 and the query matrix Q1, and the result of multiplying Input1 and the key matrix K2 can be combined to obtain the second attention score a 1,2 corresponding to Input1; …; the result of multiplying Input1 and the query matrix Q1, and the result of multiplying Input1 and the key matrix K11 can be combined to obtain the eleventh attention score a 1,11 .

[0121] Then, the respective attention scores a 1,i corresponding to Input1 are respectively multiplied by the corresponding value matrices Vi and then added to obtain the output b1 corresponding to Input1, that is:

[0122] b1 can be regarded as the first correlation degree of the reprocessed word segmentation corresponding to Input1 and other reprocessed word segmentations, b1 is a vector, since Input1 is a text content code, b1 is a corresponding text feature vector; based on the same calculation method as described above, the outputs b of other units in the comment information can be obtained i .

[0123] In the above example, the input of the self-attention mechanism contains 11 units, that is, 11 input sequences, and the output is also 11 output sequences. The BERT model is usually composed of multiple hidden layers in structure, and each hidden layer has an independent self-attention mechanism. If the number of hidden layers is M, then the input and output are: BERT([1x11]) = Mx[1x11]

[0124] That is, after inputting the input sequence of [1x11] into the BERT model of M layers, the BERT model outputs M output vectors of [1x11].

[0125] As can be seen, the self-attention mechanism can output the first correlation degree of each unit and other units. The first correlation degree is essentially the context feature of the unit, which contains the context feature of [CLS]. Since [CLS] itself does not have any meaning, and it is located before the comment information, the corresponding context feature is obtained based on the comment word segmentation and the comment sub-word segmentation in the comment information. Therefore, the output value corresponding to [CLS] contains the global information of the entire comment information.

[0126] The output values corresponding to the M [CLS] symbols are combined into a vector of [Mx1] (such as arranging and combining the output values corresponding to the M [CLS] symbols in order to form a vector), which is input into a fully connected neural network to map the output values corresponding to the [CLS] symbols to a lower-dimensional space, facilitating subsequent sentiment classification. For the output of the fully connected neural network, normalization calculation can be performed to convert the output into a probability distribution, obtaining a vector representing the probabilities of positive, neutral, and negative sentiments. Finally, according to the probability distribution, the class with the highest probability is selected as the prediction result, and the final sentiment tendency result corresponding to the comment information is positive or negative or neutral sentiment.

[0127] Optionally, in the step of S201, after obtaining the comment information, the comment information can also be filtered. The comment information that does not express a sentiment tendency and the comment information with a neutral sentiment tendency are filtered. In this case, the subsequent sentiment classification step becomes sentiment binary classification, that is, the sentiment tendency result only has positive sentiment and negative sentiment.

[0128] The BERT model can be trained before use, for example, using a pre-training + fine-tuning training process. In the pre-training stage, a large number of sample samples without sample labels are used to train the model to improve the feature learning ability and generalization ability of the model. The samples in this process can not come from real reviews, and the data can use open source natural language processing data sets. After pre-training, the BERT model has basic language processing ability, and at this time, real reviews are used as samples, and the real sentiment tendency corresponding to the real reviews is used as sample labels, and the two are combined into a fine-tuning data set to fine-tune the model, so that the model can adapt to a specific application scenario. A large number of sample samples without sample labels can be thousands or even tens of thousands of text data without sample labels, and the number of real reviews can be determined according to actual conditions, for example, several hundred to several thousand.

[0129] In the pre-training stage, a large number of sample samples without sample labels are used to train the model, and the sample can use an open source natural language processing data set. In terms of pre-training parameter settings, the learning rate can be set to 2e-5, the batch size is set to 16, and the training round number is set to 4. The training duration is determined by the size of the data set and the computing resources, and generally needs several hours to several days. In the training process, the data is randomly shuffled, padded and other preprocessing operations are performed to improve the generalization ability of the model. After pre-training, real reviews are used as samples, and the real sentiment tendency corresponding to the real reviews is used as sample labels, and the two are combined into a fine-tuning data set to fine-tune the model. In terms of fine-tuning parameter settings, the learning rate can be adjusted to 1e-5, the batch size is set to 8, and the training round number is set to 2. The fine-tuning training time is relatively short, generally a few hours. In the fine-tuning training process, the real review data is cleaned, labeled and processed to ensure the quality and accuracy of the data.

[0130] For an information information, the description target is extracted from the information information and the comment information, and the sentiment tendency of the comment information is analyzed. Considering that the content described by the comment information of the user object may not be completely consistent with the content described by the information information, it is necessary to determine which description target the sentiment tendency of the user object is directed to. On this basis, further generalization can be performed to predict the sentiment tendency of the user object to related content, so as to obtain more associated products. The related content can be content that has similarity or relevance in semantics, function, use, etc. with the description target, and the associated products can be determined by analyzing the semantic mapping relationship between the description target and the product, historical purchase data, user action data, etc.

[0131] In a possible implementation, if the description targets included in the description target set are all extracted from the attribute information set corresponding to the information information or all extracted from the comment information, the preference degree of the user object for the description targets can be directly determined based on the sentiment tendency result; if the description targets included in the description target set are extracted based on both the attribute information set and the comment information, the preference degree of the user object for a first description target extracted based on the comment information is determined based on the sentiment tendency result; and the preference degree of the user object for a second description target extracted based on the attribute information set is determined based on the correlation degree between the first description target and the second description target and the sentiment tendency result.

[0132] The correlation degree is an index for measuring the close correlation between the first description target extracted based on the comment information and the second description target extracted based on the attribute information set, and can be determined by calculating semantic similarity, co-occurrence frequency, etc., and is used to determine the preference degree of the user for the second description target.

[0133] In actual scenarios, the information information can extract at least one content subject, that is, at least one description target, and the comment information can have a case where no description target is included, therefore, the description targets included in the description target set mainly include the following two cases:

[0134] Case 1: The description targets included in the description target set are all extracted from the attribute information set.

[0135] Case 2: Part of the description targets included in the description target set are extracted from the attribute information set, and the other part is extracted from the comment information.

[0136] For case 1, that is, only the sentiment tendency result can be determined in the comment information, but the description target cannot be extracted from the comment information, the sentiment tendency of the user object can be regarded as the sentiment tendency for the content subject described by the information information, that is, the sentiment tendency for the description target extracted from the attribute information set.

[0137] As shown in FIG. 10, it is a schematic diagram for determining the description target to which the sentiment tendency is directed, provided by an embodiment of the present application. It is assumed that the information information is a video about family cleaning, the description target extracted from the attribute information set corresponding to the video is “cleaning”, and the comment information of the user object α is “This is too amazing, it can be cleaned so clean”. It can be analyzed that the sentiment tendency result of the user object α is positive sentiment, but the effective description target cannot be extracted from the comment information, therefore, the description target set corresponding to the video only has one description target “cleaning”, finally, the server can regard the positive sentiment as the positive sentiment for “cleaning”, that is, the preference degree of the user object α for “cleaning” is positive preference.

[0138] For case 2, for the first description target extracted from the comment information in the description target set, since the first description target and the sentiment tendency are from the same comment information, the sentiment tendency of the user object can be directly determined to be for the first description target; further, the content described by the comment information can not be completely consistent with the content described by the information information, in this case, the preference degree of the user object for the second description target extracted from the attribute information set can be determined by the correlation degree between the description targets, if the second description target extracted based on the attribute information set has a high correlation degree with the first description target, it can be predicted that the sentiment tendency of the user object for the second description target is consistent with the sentiment tendency of the first description target.

[0139] As shown in FIG. 11, another schematic diagram for determining the description target to which the sentiment tendency is directed is provided by the embodiment of the application, still assuming that the information information is a video about family cleaning, according to the attribute information set corresponding to the video, the description target "cleaning" is extracted, the comment information published by the user object β is "that cleaner looks very good", it can be analyzed that the sentiment tendency result of the user object β is positive emotion, and the description target "cleaner" can be extracted from the comment information, therefore, the description target set corresponding to the video contains two description targets respectively from the attribute information set and the comment information; first, the server can determine that the sentiment tendency result of the user object β for the first description target "cleaner" is positive emotion, that is, the preference degree of the user object β for "cleaner" is positive preference, further, the correlation degree between the second description target "cleaning" and the first description target "cleaner" is high, therefore, it can also be predicted that the sentiment tendency result of the user object β for the second description target "cleaning" is positive emotion.

[0140] Further, if the correlation degree of the second description target and the first description target is low, the sentiment tendency of the user object for the second description target is not predicted.

[0141] In addition, a minimum threshold can also be set for the correlation degree, if the correlation degree of the second description target and the first description target is lower than the minimum threshold, it indicates that the comment information and the information information are almost irrelevant, then the comment information is discarded, the sentiment tendency result and the description target of the comment information are not analyzed, based on this, the abnormal comments such as false comments, robot comments can be effectively excluded, to ensure the accuracy of the final preference analysis result.

[0142] The correlation degree can be determined by calculating the semantic similarity, co-occurrence frequency and the like between the two description targets.

[0143] In one embodiment, the semantic similarity can be calculated by calculating the cosine similarity between the semantic vectors of the two description targets, assuming that the semantic vectors of the two description targets are and The formula for calculating the cosine similarity C between them is The co-occurrence frequency can be calculated by counting the number of times N co two description targets appear simultaneously in a certain number of information and comment information, and the total number of times N total one of the description targets appears, that is, the co-occurrence frequency F

[0144] In one embodiment, the semantic similarity and the co-occurrence frequency can be weighted and summed to determine the correlation degree, and the formula is R = αC + (1-α)F, where R is the correlation degree, C is the semantic similarity, F is the co-occurrence frequency, and α is the weight coefficient, which can be set according to the actual situation, and the value range is 0≤α≤1. The semantic similarity can be calculated by calculating the cosine similarity between the semantic vectors of the two description targets, assuming that the semantic vectors of the two description targets are and The formula for calculating the cosine similarity C between them is The co-occurrence frequency can be calculated by counting the number of times N co two description targets appear simultaneously in a certain number of information and comment information, and the total number of times N total one of the description targets appears, that is, the co-occurrence frequency F

[0145] The minimum threshold value can be set according to the actual situation, for example, it can be set to a value between 0.2 and 0.5.

[0146] For the description target mentioned in S202, the description target can be directly a product or a product category. Assuming that the description target set only contains this one description target and the description target is extracted from the attribute information set, after obtaining the sentiment tendency result corresponding to the comment information, the preference degree of the user object to the product or the product category can be directly determined according to the sentiment tendency. If the preference is positive, a favorite product set corresponding to the user object can be directly established, and each product corresponding to the product or the product category is added to the set. Further, generalization can be performed based on the product or the product category, and other products having a strong association with the product are also added to the favorite product set, or for example, products corresponding to other product categories having a strong association with the product category are also added to the favorite product set. The favorite product set is a set corresponding to the user object. When it is determined that the user object has a positive preference for a certain product or product category, each product corresponding to the product or the product category is added to the set, and further generalization can be performed based on the product or the product category, and products corresponding to other products or product categories having a strong association with the product or the product category are also added to the set.

[0147] For example, if it is determined that the user object is interested in balls (such as footballs, basketballs, etc.), it can be considered that the user object is also interested in sports shoes for playing ball games. If it is determined that the user object is interested in painting pigments, it can be considered that the user object is also interested in brushes.

[0148] After the word segmentation processing, the keyword extraction process of S202 and the process of determining the sentiment tendency result of S203 can be executed in parallel or sequentially. The specific order can be adjusted according to actual conditions, and the present application does not make specific limitations.

[0149] S204: The server selects a recommended product for the user object in the associated product set corresponding to each description target based on the obtained preference degree for each description target.

[0150] An optional embodiment is that for each preset product, a semantic mapping relationship between a description target and the preset product is obtained; based on the semantic mapping relationship, a second association degree between a description target and the preset product is determined, and a corresponding relationship between a preset product and a description target whose second association degree reaches a set threshold is stored in an associated product set corresponding to the description target; based on the obtained preference degree for each description target, at least one recommended product satisfying a preset recommendation rule is selected for the user object in the associated product set.

[0151] The semantic mapping relationship is a corresponding relationship between a description target and a preset product at a semantic level, which is usually embodied by calculating the cosine similarity of semantic vectors of the two, and is used to determine the second association degree between them.

[0152] The second correlation degree is a correlation degree between the target and the preset product, which can be determined by calculating the cosine similarity of the semantic vectors of the two, and is used for screening the product associated with the description target. The associated product set is a set corresponding to the description target, which stores the corresponding relationship between the preset product whose second correlation degree reaches a set threshold and the description target, and is used for selecting the recommended product for the user.

[0153] The semantic mapping relationship can be represented by calculating the cosine similarity between the semantic vectors of the description target and the preset product. Assuming that the semantic vector of the description target is the semantic vector of the preset product is the calculation formula of the cosine similarity S between them is The cosine similarity S is taken as a second correlation degree between the description target and the preset product.

[0154] The source of the preset product can be multiple e-commerce platforms, product databases, etc. The server can obtain the related information of the preset product, such as the product name, description, category, etc., by calling the interface of these platforms or databases. The following method can be used to obtain the semantic mapping relationship between the description target and the preset product: first, perform word segmentation on the text information of the description target and the preset product, and then convert the segmented words into word vectors, which can use a pre-trained word vector model such as Word2Vec or GloVe. Next, the word vectors of the description target and the preset product are spliced or averaged, etc., to obtain the semantic vectors of the description target and the preset product. Finally, the semantic mapping relationship is represented by calculating the cosine similarity between the semantic vectors of the description target and the preset product.

[0155] The preset recommendation rule is a criterion for selecting a recommended product for the user from the associated product set, which can be based on the second correlation degree threshold, sorting selection, or combined with product price, sales, evaluation, and user historical purchase records. The preset recommendation rule can be that the second correlation degree reaches a set threshold, for example, the set threshold is 0.8, and the preset product with a second correlation degree greater than 0.8 is selected as the recommended product; or the second correlation degree is sorted from high to low, and the first k preset products are selected as the recommended product, and k can be set according to actual conditions.

[0156] In addition to the second correlation degree reaching the set threshold and selecting the top k products according to the second correlation degree, the preset recommendation rule can also consider the price factor of the product. The price interval can be set, and the product with a price within the acceptable range of the user is preferentially recommended. The sales factor of the product can also be considered, and the product with high sales is preferentially recommended, because the product with high sales usually has good reputation and quality. In addition, the evaluation information of the product can also be combined, and the product with good evaluation, such as the product with a good comment rate reaching a certain percentage, is preferentially recommended. At the same time, the product related to or complementary to the product previously purchased by the user can also be recommended according to the historical purchase record of the user.

[0157] That is, the second correlation degree between the description target and the preset product is determined according to the similarity between the semantics of the description target and the semantics of the preset product, and if the second correlation degree reaches the set threshold, it indicates that the correlation degree between the description target and the preset product is high, and the preset product can be stored in the associated product set corresponding to the description target. If the user object shows positive emotion to the description target, it is very likely that the user object shows positive emotion to the preset product with high correlation degree to the description target, so the user object can be randomly selected from the associated product set corresponding to the description target. The product that can be recommended, or the product that meets the recommendation rule can be further selected from the associated product set, for example, the recommendation rule can be that the second correlation degree is greater than 90%.

[0158] In addition, the method proposed in the present application is not limited to recommending products, but can also directly recommend other information information with consistent content subject to the user object according to the information information interested by the user object.

[0159] In the above, the second correlation degree between the description target and a preset product can also be obtained based on a model, that is, the model is trained with [a certain description target, a certain product] as a sample and the real correlation degree between the description target and the product as a sample label. The trained model can output a product with high second correlation degree to the description target according to the description target input into the model.

[0160] For example, the method of selecting a recommended product in the associated product set according to the description target interested by the user object can also be completed by a neural network, and the network structure can include multiple fully connected layers, activation functions and an output layer, which is not limited by the present application.

[0161] The structure of the neural network can be set to include 3 fully connected layers, the number of neurons in the first layer is 128, the second layer is 64, and the third layer is 32. The activation function can be selected as the ReLU function, and the Softmax function is used in the output layer for multi-classification. The training method adopts the stochastic gradient descent method, and the learning rate is set to 0.001. The loss function selects the cross-entropy loss function to measure the difference between the model prediction result and the true label. In the training process, the determined description target with positive sentiment of the user object and the description target with negative sentiment are used as samples to train the model, the number of training rounds is set to 100, and the batch size is set to 32. By continuously adjusting the parameters of the model, the value of the loss function gradually decreases, thereby improving the prediction accuracy of the model.

[0162] As shown in FIG. 12, a logic diagram for product recommendation based on a neural network is provided in an embodiment of the present application. The determined description target with positive sentiment of the user object and the description target with negative sentiment can be used as samples to train the model. For example, in FIG. 12, the description target with positive sentiment of the user object is summarized as description target table 1, and the description target with negative sentiment of the user object is summarized as description target table 2. The neural network is trained according to the description target table 1 and the description target table 2. The trained neural network can be deployed in an actual application scenario. The description target extracted from the information is input into the trained neural network, or the existing description target is directly input into the trained neural network. The neural network outputs whether the user object is interested in the description target. For the interested description target, products are recommended from the corresponding associated product set.

[0163] In addition, the neural network can also predict the interest degree of the user object in unwatched information content and unviewed products.

[0164] When the user object publishes a new comment, the information data such as the information browsed by the user object and the comment information published by the user object can be collected and updated in real time, so as to update the description target list of interest and uninterest of the user object. Subsequently, when the user object enters the target page, products can be selected from the associated product set corresponding to the description target of interest of the user object, and a recommendation list, a recommendation card, etc. are generated according to the selected products to display in real time on the terminal.

[0165] It can be understood that in the specific embodiments of the present application, if the comment information of the user object or the information data is obtained, the above-mentioned data needs to be known by the user object or the maker of the information, and further needs to be actively authorized by the user object or the maker, and the collection, use and processing of the related data need to comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0166] Further, the above data after use will be deleted in time.

[0167] To sum up, the present application realizes personalized product recommendation for user objects by performing sentiment analysis on comment information, extracting keywords from information information and comment information, and recommending products to user objects based on user object preferences. The above three steps can be completed by three subsystems respectively, and the three subsystems can be deployed on the background server corresponding to the information information. The subsystems cooperate with each other to complete product recommendation, improve the experience and click-through rate of user objects.

[0168] As shown in FIG. 13, it is a product recommendation system schematic diagram provided by an embodiment of the present application. Taking information information as a video as an example, the product recommendation system includes a sentiment analysis subsystem, a keyword extraction subsystem and a product selection subsystem.

[0169] The sentiment analysis subsystem is responsible for automatically obtaining the above comment information after a certain number of comment information is published by the user object, and performing sentiment analysis on the comment information to obtain the sentiment tendency result corresponding to the comment information. The keyword extraction subsystem is responsible for content analysis of the video and the comment information, that is, keyword extraction is performed to obtain the description target. The description target of the video can be extracted from the title, subtitle, introduction, partition and other meta information corresponding to the video. Then it is determined which description target the sentiment tendency of the user object is directed to. Whether the user object shows positive emotion or negative emotion to each description target is recorded in the positive emotion-description target table and the negative emotion-description target table. In addition, it can also be recorded in other forms other than tables, which is not limited in the present application.

[0170] Then, the preferences of the user object are analyzed based on the positive emotion-description target table and the negative emotion-description target table. This process can be completed with the help of a model. For a new description target that appears subsequently, the model can be used to predict the sentiment tendency of the user object to the new description target. If it is positive emotion, the associated product set corresponding to the description target can be further obtained to select a recommended product from the associated product set. The recommended product is displayed to the user object when the user object enters the target page. In addition, information information related to the recommended product can also be pushed to the user object.

[0171] The specific process of the model predicting the emotional tendency of the user object to the new description target is as follows: first, the text information of the new description target is preprocessed, including word segmentation, adding special symbols, format conversion and other operations, so as to meet the input requirements of the model. Then, the preprocessed new description target is input into the trained model, and the model calculates and analyzes the input through the internal neural network structure. The prediction of the model is based on the relationship between the description target and the emotional tendency learned in the training process, and according to the characteristics and context information of the new description target, a vector representing the probabilities of positive, neutral and negative emotions is output. Finally, according to the probability distribution, the class with the highest probability is selected as the prediction result to determine the emotional tendency of the user object to the new description target.

[0172] In one specific example, the model uses a neural network with a network structure set to include 3 fully connected layers, with 128 neurons in the first layer, 64 in the second layer, and 32 in the third layer. The activation function is ReLU, and the Softmax function is used in the output layer for multi-classification. The training method uses the stochastic gradient descent method with a learning rate of 0.001. The loss function is the cross-entropy loss function, which is used to measure the difference between the model's prediction results and the true labels. In the training process, the determined description targets with positive emotions and those with negative emotions of the user object are used as samples to train the model, with 100 training rounds and a batch size of 32. The specific process of the model predicting the emotional tendency of the user object to the new description target is as follows: first, the text information of the new description target is preprocessed, including word segmentation, adding special symbols, format conversion and other operations, so as to meet the input requirements of the model. Then, the preprocessed new description target is input into the trained model, and the model calculates and analyzes the input through the internal neural network structure. The prediction of the model is based on the relationship between the description target and the emotional tendency learned in the training process, and according to the characteristics and context information of the new description target, a vector representing the probabilities of positive, neutral and negative emotions is output. Finally, according to the probability distribution, the class with the highest probability is selected as the prediction result to determine the emotional tendency of the user object to the new description target.

[0173] In summary, the present application combines sentiment analysis of comment information and content extraction (keyword extraction) of information information to recommend products to user objects. By analyzing the emotional sentiment of the user object, the user object's preferences for the description target and the product associated with the description target can be more accurately understood; at the same time, due to the diversity of information information and the extensive content, various description targets can be extracted from a large amount of information. Combining sentiment analysis of comment information and content extraction of information information and comment information can improve the accuracy of product recommendation and the satisfaction of user objects.

[0174] Based on the same inventive concept, the application further provides a product recommendation device. As shown in FIG. 14, which is a structural schematic diagram of the product recommendation device 140, the product recommendation device 140 can include:

[0175] An acquisition unit 1401 is configured to acquire comment information published by a user object for at least one information information, and acquire attribute information sets corresponding to the at least one information information, respectively.

[0176] An extraction unit 1402 is configured to, for each information information, extract a corresponding description target set based on at least one of the attribute information set corresponding to the information information and the comment information, each description target representing a content subject capable of triggering the user object to generate an emotional tendency.

[0177] A determination unit 1403 is configured to, for each description target set, perform semantic analysis on the comment information associated with the description target set to obtain an emotional tendency result corresponding to the comment information, and determine a preference degree of the user object for at least one description target included in the description target set based on the emotional tendency result.

[0178] A selection unit 1404 is configured to, based on the obtained preference degree for each description target, select a recommended product for the user object from the associated product set corresponding to each description target.

[0179] In a possible implementation, the extraction unit 1402 is specifically configured to obtain the description targets in each description target set by the following method:

[0180] Performing word segmentation on the attribute information set and the comment information corresponding to one information information, respectively, to obtain words, wherein the word type of the word corresponding to the attribute information set is an attribute word, and the word type of the word corresponding to the comment information is a comment word.

[0181] For each word, determining a position distance between the word and each other word, and determining a target evaluation value of the word based on the obtained position distances.

[0182] Among the words, at least one keyword satisfying a keyword selection condition is determined, each keyword being an attribute keyword or a comment keyword.

[0183] Based on the word meaning of the at least one keyword, the at least one keyword is classified, and each keyword satisfying a preset classification is taken as a description target.

[0184] In a possible implementation, the extraction unit 1402 is specifically configured to:

[0185] An additive evaluation value of the one word is obtained based on a benchmark evaluation value of the one word and benchmark evaluation values of at least one other word corresponding to the one word in a connection relationship, the connection relationship indicating that a position distance between two words of the same word type satisfies a preset distance condition.

[0186] The additive evaluation value is taken as a new benchmark evaluation value of the one word, and the operation of obtaining the additive evaluation value is repeatedly performed until a preset iteration stop condition is reached; wherein, in the first iteration, the benchmark evaluation value of the one word and the benchmark evaluation values of each other word are preset.

[0187] The additive evaluation value after iteration is stopped is taken as a target evaluation value corresponding to the one word.

[0188] In a possible implementation, the determination unit 1403 is specifically configured to:

[0189] Obtain the comment words in each word;

[0190] In each comment word, each comment word satisfying a preset composition condition of a word structure is subjected to secondary word segmentation processing to obtain corresponding comment sub-words;

[0191] The comment sub-words obtained after the secondary word segmentation processing and the comment words not subjected to the secondary word segmentation processing are taken as reprocessing words;

[0192] Based on the first association degree of each reprocessing word and each other reprocessing word, an emotional tendency result corresponding to the comment information is obtained.

[0193] In a possible implementation, the determination unit 1403 is specifically configured to:

[0194] If the description target included in the description target set is extracted based on one of the attribute information set corresponding to the information information and the comment information, the preference degree of the user object for the description target is determined based on the emotional tendency result;

[0195] If the description target included in the description target set is extracted based on both the attribute information set and the comment information, the preference degree of the user object for the first description target extracted based on the comment information is determined based on the emotional tendency result.

[0196] In a possible implementation, the determination unit 1403 is further configured to:

[0197] Based on the relevance degree of the first description target and the second description target extracted based on the attribute information set, the preference degree of the user object for the second description target is determined in combination with the emotional tendency result.

[0198] In a possible implementation, the apparatus further includes a storage unit 1405, configured to:

[0199] Before selecting the recommended product for the user object from the associated product set based on the obtained preference degree for each description target, for each description target, a second association degree between the description target and each preset product is determined, and a corresponding relationship between a preset product and a description target whose second association degree reaches a set threshold is stored in the associated product set corresponding to the description target.

[0200] The selecting unit 1404 is specifically configured to:

[0201] Select at least one recommended product that meets a preset recommendation rule for the user object from the associated product set based on the obtained preference degree for each description target.

[0202] In a possible implementation, the selecting unit 1404 is specifically configured to:

[0203] For each preset product, a semantic mapping relationship between a description target and the preset product is determined.

[0204] Based on the semantic mapping relationship, a second association degree between the description target and the preset product is determined.

[0205] The apparatus can be used to execute the method shown in the embodiments of the present application, and therefore, the functions of the functional modules of the apparatus and the like can be referred to the descriptions of the foregoing embodiments.

[0206] For the sake of convenience, each part is divided into modules (or units) according to functions and described respectively. Of course, in the implementation of the present application, the functions of the modules (or units) can be implemented in the same or multiple software or hardware.

[0207] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works with other related parts to achieve a predetermined target, and can be implemented entirely or partially by using software, hardware (such as a processing circuit or a memory) or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of an overall module or unit that includes the functions of the module or unit.

[0208] After introducing the product recommendation method and apparatus of the exemplary embodiments of the present application, the computer device of another exemplary embodiment of the present application is introduced.

[0209] Those skilled in the art can understand that each aspect of the present application can be implemented as a system, a method or a program product. Therefore, each aspect of the present application can be embodied in a form of entirely hardware, entirely software (including firmware, microcode, etc.) or a combination of hardware and software, which can be generically referred to as "circuitry", "module" or "system".

[0210] Please refer to FIG. 15, based on the same technical concept, the present application also provides a computer device. In one embodiment, the computer device can be the server shown in FIG. 1, and the computer device is shown in FIG. 15, which includes a memory 1501, a communication module 1503 and one or more processors 1502.

[0211] The memory 1501 is used to store computer programs executed by the processor 1502. The memory 1501 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system and programs required for running instant messaging functions; the data storage area can store various instant messaging information and operation instruction sets, etc.

[0212] The memory 1501 can be a volatile memory such as a random-access memory (RAM); the memory 1501 can also be a non-volatile memory such as a read-only memory, a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD); or the memory 1501 can be any other medium capable of carrying or storing desired program codes in the form of instructions or data structures and capable of being accessed by a computer, but is not limited to this. The memory 1501 can be a combination of the above memories.

[0213] The processor 1502 can include one or more central processing units (CPUs) or digital processing units, etc. The processor 1502 is used to implement the above product recommendation method when calling the computer programs stored in the memory 1501.

[0214] The communication module 1503 is used to communicate with terminal devices and other servers.

[0215] The specific connection medium between the memory 1501, the communication module 1503 and the processor 1502 is not limited in the embodiments of the present application. In FIG. 15, the memory 1501 and the processor 1502 are connected through the bus 1504, which is described by a thick line in FIG. 15. The connection mode between other components is only illustrative and is not limited. The bus 1504 can be divided into an address bus, a data bus, a control bus and the like. For the convenience of description, only one thick line is used to describe the bus 1504 in FIG. 15, but it is not described that there is only one bus or only one type of bus.

[0216] The memory 1501 stores a computer storage medium, and the computer storage medium stores computer executable instructions. The computer executable instructions are used to implement the product recommendation method of the embodiments of the present application. The processor 1502 is used to execute the product recommendation method of each embodiment.

[0217] In another embodiment, the computer device can also be a terminal, such as the terminal device shown in FIG. 1. In this embodiment, the structure of the computer device can be as shown in FIG. 16, which includes a communication component 1610, a memory 1620, a display unit 1630, a camera 1640, a sensor 1650, an audio circuit 1660, a Bluetooth module 1670, a processor 1680 and the like.

[0218] The communication component 1610 is used to communicate with a server. In some embodiments, a wireless fidelity (WiFi) module can be included. The WiFi module belongs to a short-range wireless transmission technology. The computer device can help the object to transmit and receive information through the WiFi module.

[0219] The memory 1620 can be used to store software programs and data. The processor 1680 executes various functions and data processing of the terminal device by running the software programs or data stored in the memory 1620. The memory 1620 can include a high-speed random access memory and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device or other volatile solid-state memory device. The memory 1620 stores an operating system that enables the terminal device to run. In the present application, the memory 1620 can store an operating system and various application programs, and can also store a computer program for implementing the product recommendation method of the embodiments of the present application.

[0220] The display unit 1630 can also be used to display information input by the object or information provided to the object, as well as the graphical object interface (GUI) of various menus of the terminal device. Specifically, the display unit 1630 can include a display screen 1632 arranged on the front of the terminal device. Among them, the display screen 1632 can be configured in the form of a liquid crystal display, a light-emitting diode, etc. The display unit 1630 can be used to display information display interfaces and the like in the embodiments of the present application.

[0221] The display unit 1630 can also be used to receive input digital or character information, generate signal input related to object settings and function control of the terminal device. Specifically, the display unit 1630 can include a touch screen 1631 arranged on the front of the terminal device, which can collect touch operations of the object thereon or nearby, such as clicking buttons, dragging scroll boxes, etc.

[0222] Among them, the touch screen 1631 can be overlaid on the display screen 1632, or the touch screen 1631 and the display screen 1632 can be integrated to realize the input and output functions of the terminal device. After integration, it can be simply referred to as a touch display screen. The display unit 1630 in the present application can display application programs and corresponding operation steps.

[0223] The camera 1640 can be used to capture still images, and the object can publish the images taken by the camera 1640 through the application. The camera 1640 can be one or more. The object generates an optical image through a lens and projects it to a photosensitive element. The photosensitive element can be a charge coupled device (CCD) or a complementary metal-oxide-semiconductor (CMOS) phototransistor. The photosensitive element converts the optical signal into an electrical signal, and then transmits the electrical signal to the processor 1680 to convert it into a digital image signal.

[0224] The terminal device can also include at least one sensor 1650, such as an acceleration sensor 1651, a distance sensor 1652, a fingerprint sensor 1653, a temperature sensor 1654. The terminal device can also be configured with a gyroscope, a barometer, a hygrometer, a thermometer, an infrared sensor, a light sensor, a motion sensor, and other sensors.

[0225] The audio circuit 1660, the speaker 1661 and the microphone 1662 can provide an audio interface between a user and the terminal device. The audio circuit 1660 can convert received audio data into an electrical signal, transmit the electrical signal to the speaker 1661, and convert the electrical signal into an audible signal output by the speaker 1661. The terminal device can also be configured with a volume button for adjusting the volume of the audible signal. On the other hand, the microphone 1662 converts a sound signal collected into an electrical signal, which is received by the audio circuit 1660 and converted into audio data, which is output to the communication component 1610 for transmission to another terminal device, for example, or to the memory 1620 for further processing.

[0226] The Bluetooth module 1670 is used to interact with other Bluetooth devices having a Bluetooth module through a Bluetooth protocol. For example, the terminal device can establish a Bluetooth connection with a wearable computer device (e.g., a smart watch) having a Bluetooth module through the Bluetooth module 1670, and thus interact with the wearable computer device.

[0227] The processor 1680 is the control center of the terminal device, which connects various parts of the terminal device through various interfaces and lines, and performs various functions of the terminal device and processes data by running or executing software programs stored in the memory 1620 and calling data stored in the memory 1620. In some embodiments, the processor 1680 can include one or more processing units; the processor 1680 can also integrate an application processor and a baseband processor, wherein the application processor mainly processes the operating system, the object interface, and the application program, and the baseband processor mainly processes wireless communication. It can be understood that the baseband processor can also not be integrated into the processor 1680. In the present application, the processor 1680 can run an operating system, an application program, an object interface display and a touch response, and a product recommendation method of the embodiments of the present application. In addition, the processor 1680 is coupled with the display unit 1630.

[0228] In some possible implementations, various aspects of the product recommendation method provided by the present application can also be implemented in the form of a program product, which includes a computer program for causing a computer device to execute the steps of the product recommendation method according to various exemplary embodiments of the present application described above in the specification when the program product is run on the computer device, for example, the computer device can execute the steps of the various embodiments.

[0229] The program product can employ any combination of one or more computer-readable media. The computer-readable media can be a computer-readable storage medium or a computer-readable signal medium. The computer-readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium include the following: an electrical connection having one or more wires, a portable computer diskette, 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 disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0230] The program product of the embodiments of the present application can employ a compact disc read-only memory (CD-ROM) and include a computer program, and can be run on a computer device. However, the program product of the present application is not limited thereto, and in the present document, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with a command execution system, apparatus, or device.

[0231] The computer-readable signal medium can include a computer-readable storage medium that is configured to store and deliver a computer program, and can propagate in or on a carrier wave modulated onto a baseband signal or a carrier signal. The computer-readable signal medium can be any computer-readable medium that is not a computer-readable storage medium and that can communicate, propagate, or transport programming for use by or in connection with a command execution system, apparatus, or device.

[0232] The computer program contained in the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, and the like, or any suitable combination thereof.

[0233] The computer program for performing the operations of the present application can be written in any combination of one or more programming languages, including an object-oriented programming language such as Java, C++, and the like, and a conventional procedural programming language such as the "C" programming language or a similar programming language. The computer program can be executed entirely on the computer device, partially on the computer device, as a standalone software package, partially on the computer device and partially on a remote computer device, or entirely on a remote computer device or server. In the case of a remote computer device, the remote computer device can be connected to the computer device through any kind of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer device (for example, connected to the Internet using an Internet service provider).

[0234] To sum up, the application provides a product recommendation method, device, computer equipment, computer readable storage medium and computer program product. The server respectively acquires comment information of a user object for at least one information information published, and respectively acquires attribute information set corresponding to each of the at least one information information. The information information has rich forms, including video, audio, picture, text and the like, and involves sports, painting, music and the like. Different types of information information can reflect the interest points of the user from multiple dimensions, for example, video information can show the preference of vision, and audio information can reflect the preference of sound content. The acquisition of the public information information and comment information enables the server to collect a large amount of data related to the interest of the user, so as to more comprehensively and accurately determine the preference of the user object. Compared with the traditional method of determining the preference by only relying on the use evaluation of the user on the product, the method greatly improves the comprehensiveness and accuracy of determining the preference of the user, and provides a more solid foundation for subsequent personalized recommendation.

[0235] Further, each description target in the description target set is obtained by respectively performing word segmentation processing on the attribute information set and the comment information corresponding to one information information to obtain each word segmentation, determining the position distance between each word segmentation and each other word segmentation for each word segmentation, determining the target evaluation value of the word segmentation based on the obtained position distance, determining at least one keyword in the word segmentation, which satisfies the keyword selection condition, classifying the at least one keyword based on the word meaning of the at least one keyword, and taking each keyword meeting the preset classification as one description target. The word segmentation processing splits the information information and the comment information into smaller units, which is convenient for subsequent analysis. The determination of the position distance between the word segmentations can reflect the degree of association of them in the text, and the word segmentations with closer position distance may be more relevant in semantics. By calculating the target evaluation value and screening the keywords, the key information can be extracted from a large number of word segmentations, and the keywords represent the core content of the information information and the comment information. Taking the keywords meeting the preset classification as the description target makes the extraction of the description target more accurate and targeted, provides an accurate basis for subsequent analysis of the preference of the user for specific content, and improves the accuracy of the extraction of the description target.

[0236] Further, the position distance between each word and each other word is determined respectively, and based on the obtained position distances, a target evaluation value of a word is determined based on a reference evaluation value of the word and reference evaluation values of at least one other word having a connection relationship with the word, an additive evaluation value of the word is obtained based on the reference evaluation value of the word and the reference evaluation values of the at least one other word, the connection relationship represents that the position distance between two words of the same word type satisfies a preset distance condition, the additive evaluation value is taken as a new reference evaluation value of the word, the operation of obtaining the additive evaluation value is repeatedly performed until a preset iteration stop condition is reached, and the additive evaluation value after iteration is stopped is taken as the target evaluation value corresponding to the word. In the text, the more connection relationships a word has with other words, the more content it is associated with, and the more likely it is to be key information. By iteratively calculating the additive evaluation value, the degree of association can be continuously strengthened, and the evaluation value of the key word gradually increases. The target evaluation value obtained finally can more accurately reflect the importance of the word in the text, so as to more accurately determine the target evaluation value and improve the accuracy of the target evaluation value determination, thereby providing a more reliable basis for subsequent keyword screening and description target determination.

[0237] Further, the comment information associated with the description target set is subjected to semantic analysis to obtain a sentiment tendency result corresponding to the comment information, the comment words in the words are obtained, each comment word in which satisfies a preset structure condition is subjected to secondary word segmentation processing to obtain corresponding comment sub-words, and the comment sub-words obtained after the secondary word segmentation processing and the comment words not subjected to the secondary word segmentation processing are taken as reprocessing words, and the sentiment tendency result corresponding to the comment information is obtained based on a first association degree of each reprocessing word with each other reprocessing word. For some compound words or words composed of roots and affixes, secondary word segmentation processing can split them into smaller meaningful units. These comment sub-words can capture more fine-grained language features and structures and have stronger generalization ability in different text scenarios. When encountering a word that has not appeared before, the sub-words can solve the problem by combining existing sub-words, thereby improving the processing capability and effect. The sentiment tendency result is obtained based on the first association degree of the reprocessing words, which can more comprehensively consider the association of each part in the comment information, thereby more accurately analyzing the sentiment tendency of the comment information and providing a more accurate basis for determining the user's preference for the description target.

[0238] Further, if the description target included in the description target set is extracted based on one of the attribute information set corresponding to the information information and the comment information, the preference degree of the user object for the description target is determined based on the sentiment tendency result. This case simplifies the analysis process, and when the description target is extracted from only one information source, the user preference is directly determined according to the sentiment tendency result, avoiding complex correlation analysis. In this case, the sentiment tendency result can directly reflect the attitude of the user to the description target, improve the efficiency of determining the user preference, and enable the server to provide personalized recommendations for the user more quickly.

[0239] Further, if the description target included in the description target set is extracted based on both the attribute information set and the comment information, the preference degree of the user object for the first description target extracted based on the comment information is determined based on the sentiment tendency result. The comment information is the content directly expressing the user's opinion, and the first description target extracted therefrom can better reflect the user's immediate interest and emotion. Determining the user's preference for the first description target based on the sentiment tendency result can more accurately capture the user's true preference and improve the pertinence of determining the user's preference, providing a more accurate direction for subsequent product recommendations.

[0240] Further, based on the relevance of the first description target and the second description target extracted based on the attribute information set, and in combination with the sentiment tendency result, the preference degree of the user object for the second description target is determined. In actual situations, the content described by the comment information may not be completely consistent with the content described by the information information, but there may be some correlation between the two. By calculating the relevance of the first description target and the second description target, the closeness of the association between them can be determined. If the relevance is high, it can be reasonably predicted that the sentiment tendency of the user to the second description target is similar to that of the first description target. This method can infer the user's preference through correlation analysis in the case where the user does not directly express the sentiment to the second description target, expanding the scope of user preference analysis and improving the rationality of determining the user's preference, providing the user with more extensive and potentially desirable product recommendations.

[0241] Further, before selecting the recommended products for the user object from the associated product set of each description target based on the obtained preference degree of each description target, a second correlation degree between each description target and each preset product is determined respectively for each description target, and a corresponding relationship between a preset product whose second correlation degree reaches a set threshold and a description target is stored in the associated product set corresponding to the description target. At least one recommended product that meets a preset recommendation rule is selected from the associated product set for the user object based on the obtained preference degree of each description target. By calculating the second correlation degree between the description target and the preset product, the product with a high correlation degree with the description target can be selected from a large number of preset products. These products with a high correlation degree are more likely to meet the user's interest and demand. Storing these products in the associated product set provides a precise product pool for subsequent recommendation. Selecting the recommended product from the associated product set based on the user's preference degree for the description target can provide the user with a recommendation that meets the user's personalized demand, improve the accuracy of product recommendation, and increase the user's satisfaction and click-through rate for the recommended product.

[0242] Further, the second correlation degree between each description target and each preset product is determined by respectively obtaining a semantic mapping relationship between a description target and a preset product for each preset product, and determining the second correlation degree between a description target and a preset product based on the semantic mapping relationship. The semantic mapping relationship can measure the similarity and correlation between the description target and the preset product from the semantic level. By calculating the cosine similarity of the semantic vectors of the two to determine the semantic mapping relationship, the matching degree of them in the semantic level can be more accurately reflected. The second correlation degree determined based on the semantic mapping relationship can more accurately select the preset product related to the description target, improve the accuracy of the determination of the second correlation degree, and provide a more reliable basis for constructing the associated product set and performing product recommendation.

[0243] Further, for video type information, a multi-modal method is adopted, i.e., using deep learning image recognition, speech recognition and semantic understanding technology to obtain the description target. In terms of image recognition, a target detection algorithm based on convolutional neural network (CNN) is selected, such as Faster R-CNN or YOLO series algorithm, to extract multiple image frames from the video, input them into the trained CNN model, and identify the objects, scenes and other information in the image. In terms of speech recognition, an end-to-end speech recognition model is used, such as DeepSpeech or Wav2Vec 2.0, to extract and preprocess the audio content of the video, and then input it into the speech recognition model to convert the audio signal into text information. In terms of semantic understanding, a pre-trained language model based on Transformer architecture is used, such as BERT or GPT, to input the text information obtained by speech recognition into the language model for semantic analysis and extraction of the description target corresponding to the video. Through the multi-modal method, the results of image recognition, speech recognition and semantic understanding are fused, which can comprehensively analyze the video content from multiple angles. Different modalities can provide different information, image recognition can obtain visual information in the video, speech recognition can obtain text information in the audio, and semantic understanding can deeply analyze these information. The combination of multiple modalities makes the extracted description target more comprehensive and accurate, and can cover more key content in the video, providing more rich information for subsequent analysis of user preferences and product recommendation.

[0244] Further, after obtaining the comment information, the comment information without expressing emotional tendency and the comment information with neutral emotional tendency are filtered. This can reduce the amount of data for subsequent processing, and focus on the comment information with clear emotional tendency. In the process of emotional classification, multiple classification (positive, neutral, negative) may be required, and after filtering, it becomes emotional binary classification (positive, negative), simplifying the classification process and improving the efficiency of emotional classification. At the same time, the interference of irrelevant information is reduced, so that the analysis result is more focused on the real preference of the user, and the accuracy of the analysis is improved.

[0245] Further, the BERT model adopts a pre-training + fine-tuning training process before use. In the pre-training stage, a large number of sample labels without samples are used to train the model, and these samples can use open source natural language processing data sets. Pre-training can improve the feature learning ability and generalization ability of the model, so that the model has basic language processing ability. In the fine-tuning training stage, real reviews are used as samples, and the real sentiment tendency results corresponding to the real reviews are used as sample labels, and the two are combined into a fine-tuning data set to fine-tune the model. Fine-tuning training can make the model adapt to a specific application scenario and improve the accuracy of the model in actual review information sentiment analysis. Through this training method, the model can better understand the semantics and sentiment of the review information, providing a guarantee for accurately obtaining the sentiment tendency result of the review information.

[0246] Further, the neural network is trained according to the description target with positive and negative emotions of the user object, and the network structure includes multiple fully connected layers, activation functions and an output layer. By continuously adjusting the parameters of the model, the value of the loss function is gradually reduced, thereby improving the prediction accuracy of the model. The trained neural network can be deployed in actual application scenarios, and the description target extracted from the information information is input into the trained neural network, and the neural network can output whether the user object is interested in the description target. For the description target of interest, products are recommended from the corresponding associated product set. In addition, the neural network can also predict the interest degree of the user object in unwatched information content and unvisited products. The neural network has strong learning and prediction ability, and can discover the potential relationship between the description target and the user interest through learning a large number of samples. By continuously optimizing the model parameters, the prediction accuracy of the model can be improved, and more products that meet the potential needs of the user are provided, thereby expanding the range and accuracy of the recommendation.

[0247] It should be noted that although several units or sub-units of the apparatus are mentioned in the foregoing detailed description, such division is merely exemplary and not mandatory. Indeed, features and functions of two or more units described above can be embodied in one unit, according to the embodiments of the present application. Conversely, features and functions of one unit described above can be further divided into units embodied by multiple units.

[0248] In addition, although the operations of the method of the present application are described in a specific order in the drawings, this is not required or implied that the operations must be performed in that specific order, or that all of the shown operations must be performed to achieve the desired result. Additionally or alternatively, certain steps can be omitted, combined into a single step, and / or divided into multiple steps.

[0249] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable code.

[0250] The present application is described in reference to the flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program commands. These computer program commands can be provided to a processor of a general purpose computer, a special purpose computer, an embedded processor or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0251] These computer program commands can also be stored in a computer readable storage medium that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable storage medium produce an article of manufacture including instructions which implement the function specified in the flowchart illustrations and / or block diagrams block or blocks.

[0252] These computer program commands can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0253] Each technical feature in the above embodiments can be combined in any manner, and for the sake of brevity, not all possible combinations are described, however, it will be understood by those skilled in the art that the present application embraces all such possible combinations.

[0254] The above-described embodiments are merely illustrative of several embodiments of the present application, which are described in more detail and in a specific manner, but should not be construed as limiting the scope of the patent. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A product recommendation method, executed by a computer device, comprising: obtaining, respectively, comment information published by a user object for at least one information information, and obtaining, respectively, a set of attribute information corresponding to each of the at least one information information; for each of the information information, extracting a corresponding set of description targets based on at least one of the set of attribute information and the comment information corresponding to the information information, each of the description targets representing a content subject triggering a sentiment tendency of the user object; for each of the set of description targets, performing semantic analysis on the comment information associated with the set of description targets to obtain a sentiment tendency result corresponding to the comment information, and determining, based on the sentiment tendency result, a preference degree of the user object for at least one description target included in the set of description targets; and based on the obtained preference degree for each of the description targets, selecting a recommended product for the user object from a set of associated products corresponding to each of the description targets. 2.The method of claim 1, wherein each of the description targets in the set of description targets is obtained by: respectively, and each of the attribute information sets and the comment information corresponding to the one information information is subjected to word segmentation processing to obtain respective words, a token type of a token corresponding to the set of attribute information is an attribute token, and a token type of a token corresponding to the comment information is a comment token; for each token, determining a position distance between the token and each of other tokens, and determining a target evaluation value of the token based on the obtained position distances; from the tokens, determining at least one keyword whose target evaluation value satisfies a keyword selection condition, each of the keywords being an attribute keyword or a comment keyword; based on a word meaning of the at least one keyword, classifying the at least one keyword, and taking each of the keywords satisfying a preset classification as a description target. 3.The method of claim 2, wherein the determining a position distance between the token and each of other tokens, and determining a target evaluation value of the token based on the obtained position distances comprises: obtaining an overlapping evaluation value of the token based on a reference evaluation value of the token and a reference evaluation value of each of at least one other token having a connection relationship with the token, the connection relationship representing that a position distance between two tokens of the same token type satisfies a preset distance condition; taking the overlapping evaluation value as a new reference evaluation value of the token, repeating the operation of obtaining the overlapping evaluation value until a preset iteration stop condition is reached; wherein, at the first iteration, the reference evaluation value of the token and the reference evaluation value of each of the other tokens are preset; taking the overlapping evaluation value after iteration as the target evaluation value corresponding to the token. 4.The method of claim 2 or 3, wherein the performing semantic analysis on the comment information associated with a set of description targets to obtain a sentiment tendency result corresponding to the comment information comprises: obtaining comment tokens from the tokens; for each of the comment tokens, performing secondary tokenization processing on the comment token to obtain corresponding comment sub-tokens, when a token structure of the comment token satisfies a preset composition condition. the each comment sub-word obtained after the secondary word segmentation processing, and the each comment word not subjected to the secondary word segmentation processing, as reprocessing words; obtaining a sentiment tendency result corresponding to the comment information based on the first association degree of each reprocessing word with each other reprocessing word.

5. The method of any one of claims 1-4, wherein the preference degree of the user object for at least one description target included in the one set of description targets is determined based on the sentiment tendency result, including: if the description target included in the set of description targets is extracted based on one of the set of attribute information and the comment information corresponding to the one information information, then the preference degree of the user object for the description target is determined based on the sentiment tendency result.

6. The method of any one of claims 1-5, wherein the preference degree of the user object for at least one description target included in the one set of description targets is determined based on the sentiment tendency result, including: if the description target included in the set of description targets is extracted based on both the set of attribute information and the comment information, then the preference degree of the user object for a first description target extracted based on the comment information is determined based on the sentiment tendency result.

7. The method of claim 5 or 6, further comprising: determining the preference degree of the user object for a second description target extracted based on the set of attribute information based on the correlation between the first description target and the second description target, and the sentiment tendency result.

8. The method of any one of claims 1-7, wherein before the recommended product is selected for the user object from the set of associated products corresponding to each description target based on the obtained preference degree for each description target, the method further comprises: determining a second association degree between each description target and each preset product, and storing a corresponding relationship between a preset product whose second association degree reaches a set threshold and the one description target in the set of associated products corresponding to the one description target; selecting at least one recommended product for the user object from the set of associated products based on the obtained preference degree for each description target, including: selecting at least one recommended product for the user object from the set of associated products that satisfies a preset recommendation rule.

9. The method of claim 8, wherein the second association degree between each description target and each preset product is determined, including: obtaining a semantic mapping relationship between the one description target and each preset product; 10. A product recommendation device, comprising: an obtaining unit configured to obtain comment information published by a user object for at least one information information, and obtain a set of attribute information corresponding to each of the at least one information information, respectively; ​ ​ ​ ​ ​ ​ ​ ​ extracting units, configured to extract, for each of the information information, a corresponding description target set based on at least one of the attribute information set and the comment information corresponding to the information information, each of the description targets representing a content subject capable of triggering the user object to generate an emotional tendency; determining units, configured to respectively perform semantic analysis on the comment information associated with each of the description target sets, to obtain an emotional tendency result corresponding to the comment information, and to respectively determine a preference degree of the user object for at least one description target included in the one description target set based on the emotional tendency result; and selecting units, configured to select, based on the obtained preference degrees of the description targets, a recommended product for the user object from the associated product sets corresponding to the description targets respectively. 11.A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method in any one of claims 1 to 9 when executing the computer program. 12.A computer storage medium, having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method in any one of claims 1 to 9. 13.A computer program product, comprising a computer program, wherein the computer program, when executed by a processor, implements the steps of the method in any one of claims 1 to 9.

Images