A personalized opinion label generation method and system for e-commerce platforms

Abstract

The present invention belongs to the field of computer and Internet technology, and provides a method and system for generating personalized opinion labels for e-commerce platforms. The method includes obtaining the comment semantic information and behavior records of users in the e-commerce platform; based on the comment semantic information, using an explicit preference tracker to obtain the user's explicit preferences; based on the behavior records, using an implicit preference tracker to obtain the user's implicit preferences; according to the user's explicit preferences and the user's implicit preferences, using a label generator to generate opinion labels corresponding to the user's preference labels; wherein the explicit preference tracker uses a hierarchical heterogeneous graph from the bottom to the top information propagation path to obtain the user's explicit preferences; the implicit preference tracker uses a multi-class behavior graph to obtain the user's implicit preferences. The present invention mines explicit and implicit personalized preferences from user comments and user behaviors respectively, and the two preferences enhance and complement each other, making the mined personalized information more comprehensive and accurate.

Description

Technical Field

[0001] The present invention belongs to the field of computer and Internet technology, and in particular, relates to a method and system for generating personalized opinion tags for an e-commerce platform. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Text summarization refers to a text compression technology that extracts key information from specific text and automatically generates short summaries. The text summarization task plays an important role in natural language processing. By accurately and quickly capturing the key points of information in massive amounts of information, it greatly speeds up the information filtering process and effectively alleviates the increasingly serious problem of information overload. In e-commerce platforms, text summarization technology mainly relies on a large amount of comments left by users on the platform. By analyzing the key content in the comments in multiple dimensions, it can fully understand product information and help users quickly grasp the pros and cons of products. This is of great significance to improving user satisfaction, enhancing user experience, and promoting product and platform development.

[0004] At present, the text summary generation task based on the e-commerce platform has been further developed on the basis of the review summary generation task, and has produced more fine-grained subtasks, such as opinion summary generation, product description generation, product title or keyword generation, recommendation generation, etc. In terms of the completeness of information summary, the product description generation task mainly aims to summarize the comprehensive and complete information of the product, and usually generates a long descriptive text. Other tasks are mainly aimed at being brief and concise, advocating the generation of a single or less main key information of a product or review, while ignoring less important information. From the above analysis, the research work of existing tasks either generates all key information indiscriminately, or only generates the most critical information, and cannot take into account the completeness of the extracted information and the order of the key points of the information. From the product perspective, the completeness of the extracted information helps users to fully understand the product information. From the user perspective, the order of the key points can ensure that the information that users first obtain is the content that they are most interested in.

[0005] In order to generate a comprehensive and orderly summary of key information for reviews, the opinion label generation task came into being. According to the product aspects described in the reviews, such as price, service, quality, etc., the opinion label generation task first clusters the reviews. A category of reviews describing the same product aspect will be used to generate an opinion label for the product in a certain aspect, and finally the generated opinion labels are sorted according to the size of the class. The proposal of the opinion label generation task provides a new idea for the review summary generation task, and its comprehensive and orderly characteristics make up for the shortcomings of the previous tasks. However, the order in the traditional opinion label generation task is determined by the size of the review cluster, which reflects the user's general preference for the product, but ignores the user's personalized preference. Summary of the invention

[0006] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides a personalized opinion label generation method and system for e-commerce platforms, which defines a personalized opinion label generation task, extracts comprehensive information of products from product reviews, and mines personalized preferences based on user reviews and user behaviors to determine the order of product key information, so that the key information at the top of the order conforms to the personalized preferences of a specific user rather than the universal preferences of the public.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A first aspect of the present invention provides a method for generating personalized opinion tags for an e-commerce platform.

[0009] A personalized opinion label generation method for an e-commerce platform, comprising:

[0010] Obtain the comment semantic information and behavior records of users on the e-commerce platform;

[0011] Based on the comment semantic information, an explicit preference tracker is used to obtain the user's explicit preference;

[0012] Based on the behavior records, an implicit preference tracker is used to obtain the user's implicit preference;

[0013] According to the user's explicit preference and the user's implicit preference, a label generator is used to generate opinion labels corresponding to the user's preference labels;

[0014] Among them, the explicit preference tracker adopts the information propagation path from the bottom layer to the top layer of the hierarchical heterogeneous graph to obtain the user's explicit preference; the implicit preference tracker adopts multiple types of behavior graphs to obtain the user's implicit preference.

[0015] Furthermore, the aspect tag represents information about a certain aspect of the product.

[0016] Furthermore, the hierarchical heterogeneous graph includes a word layer, an aspect layer, a product layer and a user layer from the bottom layer to the top layer, and the path from the bottom layer to the top layer is defined as a WAPU meta-path, and the path from the word layer, the aspect layer to the user layer is defined as a WAU meta-path.

[0017] Furthermore, the first user preference is obtained through the information propagation of the WAPU meta-path; and the error accumulation from the product layer to the user layer in the WAPU is reduced through the WAU meta-path to obtain the second user preference.

[0018] Furthermore, the method of obtaining the user's implicit preferences based on the behavior records by using an implicit preference tracker specifically includes: constructing the multi-category behavior graph based on users, products, clicks, favorites, and orders; based on a single behavior, aggregating the user representation and product representation of the multi-category behavior graph by using a multi-layer aggregation method to obtain a single user representation and a single product representation; based on multi-type behaviors, obtaining the user representation and product representation of multi-type behaviors;

[0019] Aggregate user representations of multiple types of behaviors of several other users similar to the user to obtain the user's personalized preferences based on behavior;

[0020] Among them, behaviors include: clicking, collecting and placing orders.

[0021] Furthermore, after obtaining the user representation and product representation, it includes: constructing a behavior UP matrix according to the user representation and product representation learned by LightGCN, aligning the behavior UP matrix with the comment adjacency matrix in the explicit preference tracker, and obtaining the user's implicit preference.

[0022] Furthermore, the step of using a label generator to generate opinion labels corresponding to labels of user preferences according to user explicit preferences and user implicit preferences specifically includes:

[0023] Phase 1: Obtain the aspect tags that the user is interested in, use the tag generator based on the first user preference, obtain the probability of each aspect tag, sort the probabilities of the aspect tags, and select the aspect tags with the highest probability.

[0024] The second stage: using the set number of aspect labels with the highest probability, and using the label generator to obtain the opinion label corresponding to each aspect label.

[0025] A second aspect of the present invention provides a personalized opinion label generation system for an e-commerce platform.

[0026] A personalized opinion label generation system for an e-commerce platform, comprising:

[0027] A data acquisition module is configured to: acquire comment semantic information and behavior records of users on the e-commerce platform;

[0028] An explicit preference tracking module is configured to: obtain the user's explicit preference based on the comment semantic information using an explicit preference tracker;

[0029] An implicit preference tracking module is configured to: obtain the user's implicit preference based on the behavior record using an implicit preference tracker;

[0030] A label generation module is configured to: generate opinion labels corresponding to labels of user preference aspects by using a label generator according to user explicit preferences and user implicit preferences;

[0031] Among them, the explicit preference tracker adopts the information propagation path from the bottom layer to the top layer of the hierarchical heterogeneous graph to obtain the user's explicit preference; the implicit preference tracker adopts multiple types of behavior graphs to obtain the user's implicit preference.

[0032] A third aspect of the present invention provides a computer-readable storage medium.

[0033] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps in the method for generating personalized opinion labels for an e-commerce platform as described in the first aspect above.

[0034] A fourth aspect of the present invention provides a computer device.

[0035] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for generating personalized opinion labels for an e-commerce platform as described in the first aspect above are implemented.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The present invention fills the gap in the personalized direction of opinion tag generation tasks and creatively proposes a unified framework for jointly extracting users' explicit and implicit personalized preferences. Different from the traditional review summary generation task, the present invention has the following advantages:

[0038] 1. In order to better mine the personalized preferences of users, the present invention designs two personalized information mining modules based on the comments in the e-commerce platform, respectively mining explicit and implicit personalized preferences from user comments and user behaviors, and denoising according to the different characteristics of different data to generate comprehensive and orderly personalized opinion sequence labels. The two preferences enhance and complement each other, making the mined personalized information more comprehensive and accurate.

[0039] 2. Take into account the information completeness and order of key points in generating review summaries, and generate comprehensive and orderly summaries personalized for specific users. This can not only comprehensively reflect the main features of all aspects of the product, but also arrange the order of key points according to user interests, so that the aspects that users are interested in are placed in the most prominent position.

[0040] 3. A new unified framework for mining users' explicit and implicit preferences is proposed, which solves the sparsity problem of comments. Based on the different characteristics of user comments and user behavior data, a targeted denoising method is proposed to make the extracted user preferences more comprehensive and accurate.

[0041] 4. Based on real e-commerce platform data, the present invention has achieved good improvements in various generation indicators and ranking indicators. It is proved that in real scenarios, the personalized opinion tags generated by the present invention can accurately capture user interests, fully reflect product information, greatly shorten the time for users to obtain key information, improve the user experience, and promote the development of e-commerce platforms and increase platform revenue. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0043] Figure 1 is a flow chart of a method for generating personalized opinion tags for an e-commerce platform, shown in Embodiment 1 of the present invention;

[0044] Figure 2 It is a structural diagram of a hierarchical heterogeneous graph shown in the first embodiment of the present invention. DETAILED DESCRIPTION

[0045] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0046] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0047] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0048] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the methods and systems according to various embodiments of the present disclosure. It should be noted that each box in the flowchart or block diagram can represent a module, a program segment, or a part of a code, and the module, program segment, or a part of a code may include one or more executable instructions for implementing the logical functions specified in each embodiment. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, or they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the flowchart and / or block diagram, and the combination of boxes in the flowchart and / or block diagram can be implemented using a dedicated hardware-based system that performs a specified function or operation, or can be implemented using a combination of dedicated hardware and computer instructions.

[0049] Embodiment 1

[0050] like Figure 1 As shown, this embodiment provides a method for generating personalized opinion labels for an e-commerce platform. This embodiment uses the method applied to a server as an example. It can be understood that the method can also be applied to a terminal, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network servers, cloud communications, middleware services, domain name services, security services CDN, and big data and artificial intelligence platforms. The terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited to this. The terminal and the server can be directly or indirectly connected via wired or wireless communication, which is not limited in this application. In this embodiment, the method includes the following steps:

[0051] Obtain the comment semantic information and behavior records of users on the e-commerce platform;

[0052] Based on the comment semantic information, an explicit preference tracker is used to obtain the user's explicit preference;

[0053] Based on the behavior records, an implicit preference tracker is used to obtain the user's implicit preference;

[0054] According to the user's explicit preference and the user's implicit preference, a label generator is used to generate opinion labels corresponding to the user's preference labels;

[0055] Among them, the explicit preference tracker adopts the information propagation path from the bottom layer to the top layer of the hierarchical heterogeneous graph to obtain the user's explicit preference; the implicit preference tracker adopts multiple types of behavior graphs to obtain the user's implicit preference.

[0056] This embodiment is based on the mainstream generation framework: Transformer technology, and encodes product reviews through the transformer encoding layer to obtain the main information of the product, so as to achieve the purpose of extracting complete information. On this basis, this embodiment proposes two modules for tracking user preferences, namely the review-based explicit preference tracker (REPT) and the behavior-based implicit preference tracker (BIPT), so as to extract explicit and implicit user preferences from user reviews and user behaviors. Finally, by fusing product information and user preference information in the transformer decoding layer and using a personalized ranking-aware label generator (PRT), this embodiment can effectively generate personalized, comprehensive and effective opinion labels. The workflow diagram of this embodiment is shown in the figure. Figure 1 As shown in the figure, it is mainly divided into three parts: (1) Review-based Explicit Preference Tracker (REPT) (2) Behavior-based Implicit Preference Tracker (BIPT) and (3) Personalized Ranking-aware Tag Generator (PRT). The following is a detailed introduction:

[0057] 1. Basic framework transformer (Transformer)

[0058] First, the basic framework transformer of the embodiment is introduced. It is the current mainstream generation model, which contains multiple identical encoding layers and multiple identical decoding layers. The encoding layer encodes the product review into a word vector and adds position information to each word. Then, multiple feedforward neural network layers (FFN) and multi-head self-attention layers are connected, and residual connections and normalization operations are performed on each layer to finally obtain the global product review representation h. enc Next, the decoding layer combines h enc The standard output predicts the word to be generated at the current moment. The structure of the decoding layer is similar to that of the encoding layer, except that the multi-head cross-attention layer is used instead of the multi-head self-attention layer, so that the attention to each word of the input comment is constantly changed at each moment of generation.

[0059] For the personalized label generation task, we improved the basic transformer framework. We incorporated personalized representation information into the decoding layer and considered the sorting information during decoding. Before introducing the details of the improvement, we first introduce two personalized information mining modules.

[0060] 2. Review-based Explicit Preference Tracker (REPT)

[0061] Compared with the traditional review summary generation task that only uses the user's historical review text or predefined user attributes as input to mine the user's personalized information, we designed a review-based explicit preference tracker (REPT) to jointly model users, products and reviews through a hierarchical heterogeneous graph (HHG). In the hierarchical heterogeneous graph, deep information interaction can be carried out between products and products, users and users, and users and products, which is conducive to mining the potential relationship between them and alleviates the sparsity problem of reviews to a certain extent. Since the relationship between user preferences and opinion labels is difficult to capture, we establish aspect labels between the two to explore the relationship between them. Each opinion label describes the information of a certain aspect of the product, that is, it corresponds to a specific aspect label. Aspect labels are predefined, so users' preferences on opinion labels can be mined by tracking their preferences on aspect labels. Based on this, we also introduced aspects in the hierarchical heterogeneous graph. The complete hierarchical heterogeneous graph structure is shown in the figure below. Figure 2 The hierarchical heterogeneous graph contains four layers, from top to bottom: user layer (user), product layer (product), aspect layer (aspect), and word layer (word).

[0062] In the hierarchical heterogeneous graph, the information of all nodes comes from the word layer. Each word node in the word layer is a word in the comment. They are encoded through an open source pre-trained vocabulary to obtain an initial representation. Subsequently, the hierarchical heterogeneous graph propagates information between layers from bottom to top. For the user layer, product layer, and aspect layer, the neighbors of each node in the layer are heterogeneous nodes in the lower layer. For ease of description, we use uppercase letters to represent set representations and lowercase letters to represent element representations. Specifically, for any node i in these three layers, the process of propagating information from the bottom-level neighbor j to i is:

[0063] h′ i =GAT(h i ,{h j ,j∈N i})(1)

[0064] It can be expressed as:

[0065]

[0066]

[0067] e ij =FFN([h i ,Whj ])(4)

[0068] Among them, h i represents the representation of node i before updating. FFN(*) is a feedforward neural network. ij A graph attention score is calculated, which assigns different weights to the information of neighboring nodes. The final updated representation h′ of node i i is the weighted sum of all heterogeneous neighbor nodes.

[0069] We define the path of information propagation from the bottom layer to the top layer of the hierarchical heterogeneous graph as the WAPU meta-path (word-aspect-product-user meta-path), that is, the semantic information of the review is uploaded from the word layer to the aspect layer, then to the product layer, and finally to the user layer. Through the information propagation of the WAPU meta-path, we can obtain the representation of the user node as a representation of user preferences based on the review.

[0070] In order to alleviate the sparsity of reviews, further deepen the interaction between users and products, and enhance the accuracy of learning user preference representations, we introduce the WAU meta-path (word-aspect-user meta-path). The information propagation and aggregation methods of the WAU meta-path are the same as those of the WAPU, but the information at the user level is directly propagated by the aspect layer, and the connection between users and aspects is defined as the aspects contained in the reviews posted by users. The shorter information propagation path avoids the error accumulation from the product layer to the user layer in WAPU. In order to enhance the information interaction between products, capture more fine-grained information links, and alleviate the information errors caused by the noise of reviews, we apply contrastive learning to enhance the representation between the product layer of WAPU and the user layer of WAU.

[0071] We have reason to believe that if users are satisfied with a product, then the product information and user preference information will be consistent to a certain extent. For any product p, the set of users who are satisfied with it is represented by u + , and the remaining users are denoted as u - We use the contrastive loss to represent the product p Representation with positive user Approximately close to the negative user representation Try to stay away from it. This process can be expressed as:

[0072]

[0073] in, is the user representation learned through the WAU meta-path. By applying contrastive learning, we can obtain more accurate product representations in the WAPU meta-path, so the user representation aggregated from the product representation is also more accurate. Therefore, we mine user personalized preferences based on reviews through hierarchical heterogeneous graphs, enhance the interaction between users and products, and effectively alleviate the sparsity and noise problems of reviews.

[0074] 3. Behavior-Based Implicit Preference Tracker (BIPT)

[0075] Since there are many users who only consume but do not comment on e-commerce platforms, it is impossible to use historical comment information to obtain user preferences for such users. Fortunately, they have left a lot of behavior records on the platform. For this reason, we designed a behavior-based implicit preference tracker (BIPT) to mine users' personalized preferences by tracking the clicks, favorites, and orders left by users when using the platform. Based on the user's various behaviors, we use a multiple behavior graph (MBG) to model the user's implicit preferences. The graph mainly contains two types of nodes: users and products, and three types of edges: clicks, favorites, and orders.

[0076] We use the IDs of users and products as the initial embeddings of nodes in the multi-class behavior graph and model different types of edges in the multi-class behavior graph separately. Taking click behavior as an example, for a given user and product interaction click behavior adjacency matrix S c , the user's ID encoding matrix Product ID matrix Where |U|, |P| are the number of users and products in the dataset, d is the ID embedding dimension, and we use multi-layer LightGCN to aggregate the representations of users and products in the behavior graph. User representation The aggregation process can be expressed as:

[0077]

[0078]

[0079] Among them, L is the number of layers of LightGCN, D -1 YesS c The product represents the diagonal matrix of The polymerization process is to use Alternative After obtaining the user and product representations of a single behavior, we fuse the three different types of behavior representations to obtain the user and product representations of multiple types of behaviors. For example, the fusion of user representations can be expressed as:

[0080]

[0081] Since the input of multi-type behavior graph does not contain semantic information, we only explore the association relationship between users based on node representation, mine similar users, and then use the similarity relationship to propagate information among the nodes in the graph nodes based on semantic representation, namely, the Review-based Explicit Preference Tracker (REPT).

[0082] For any user in the behavior graph The similarity scores between the user and other users in the set can be calculated by inner product. We select the k users with the highest similarity scores. s By aggregating the user representations of similar users in the review-based explicit preference tracker (REPT), the user’s behavior-based personalized preference representation can be obtained. The process can be expressed as:

[0083]

[0084]

[0085]

[0086] in, is the user node representation under the WAU meta-path in the Review-based Explicit Preference Tracker (REPT), w u,ui For similar users u i The contribution weight of user u.

[0087] The large amount of noise in user behavior will seriously interfere with the learning of user association relationships. In order to make the user similarity relationship explored by the behavior-based implicit preference tracker (BIPT) reliable, we reconstruct the comment relationship between users and products in the behavior graph so that the user's behavior on the product is aligned with the comment behavior, thereby supervising the learning of the behavior graph representation. Specifically, we first construct the behavior UP matrix I based on the user and product representations learned by LightGCN b , and then the behavior UP matrix is ​​combined with the review adjacency matrix I in the review-based explicit preference tracker (REPT) r Alignment, the specific process can be expressed as:

[0088]

[0089]

[0090] Where |I r | represents the number of elements in the adjacency matrix.

[0091] Therefore, we mined users' implicit preferences based on multi-category behavior graphs, and designed a denoising module based on the characteristics of user behavior data to alleviate the noise of user behavior and ensure the accuracy of user personalized preferences.

[0092] 4. Personalized Rank-Aware Tag Generator (PRT)

[0093] Through the review-based explicit preference tracker (REPT) and behavior-based implicit preference tracker (BIPT), we can capture users’ explicit preferences. and implicit preferences To incorporate personalized information, we improve the decoding layer of the Transformer to consider generating ordered opinion labels based on user preferences.

[0094] Since the user's preference in opinion labels can be transformed into the user's preference for predefined aspects, we generate ordered aspect labels to represent the order of opinion labels. Therefore, we divide the generation into two stages. First, we generate ordered aspect labels with user preferences, and then for each aspect label generated in the first stage, we generate the opinion label corresponding to the aspect. The unified paradigm of the generation process can be expressed as:

[0095]

[0096] p(y t )=TransDec([e1,e2,…,e t ])(15)

[0097] Among them, x t is the input at time t, Fusion(*) is a linear layer used to fuse personalized preference representations into the input, y t is the output at time t, and TransDec(*) is the decoding layer structure of the traditional transformer.

[0098] In the first generation phase, we first generate the aspect labels that users are interested in. We take the product representations from the Review-based Explicit Preference Tracker (REPT) as the initial input, i.e. Generate the probability of each aspect in the predefined aspect set, that is, y1 = a i Then select the k with the highest probability g The second stage of generation is carried out.

[0099] In the first stage, in order to make the size of the aspect label generation probability consistent with the user's preferred aspect order, we use a ranking-based loss function to supervise the generation of aspect labels. We use g i represents the generated aspect label, li Represents the aspect labels of real user preferences, then the ranking-based loss function can be expressed as:

[0100]

[0101] Among them, ΔNDCG(i,j) represents the g i and g j The absolute change in the ranking indicator NDCG after the order of the two aspect labels is swapped.

[0102] In the second stage of generation, we target the k generated in the first stage. g Each aspect label generates a corresponding opinion label.

[0103]

[0104] y i,t =w i,t-1 ,t>1(18)

[0105] where w i,t-1 Indicates aspect a i The t-1th word in the corresponding opinion label. In order to maximize the generation probability, we use negative log-likelihood loss to supervise the generation of opinion labels:

[0106]

[0107] in Represents the probability of generating a true opinion label.

[0108] Therefore, the improved Transformer decoding layer is more suitable for the task of generating personalized opinion labels. It can generate opinion labels that users are interested in based on their preferences and ensure that the higher the ranking, the more interested the user is.

[0109] 5. Multi-task joint training

[0110] We use a multi-task learning framework to uniformly minimize contrast loss, selection loss, ranking loss, and generation loss. The objective function is expressed as:

[0111] L=λ1L c +λ2L s +λ3L r +λ4L g (20)

[0112] Among them, λ1,λ2,λ3,λ4 are hyperparameters used to balance the weights of various loss functions. Therefore, each component of the joint model can be trained in an end-to-end manner.

[0113] Embodiment 2

[0114] This embodiment provides a personalized opinion label generation system for an e-commerce platform.

[0115] A personalized opinion label generation system for an e-commerce platform, comprising:

[0116] A data acquisition module is configured to: acquire comment semantic information and behavior records of users on the e-commerce platform;

[0117] An explicit preference tracking module is configured to: obtain the user's explicit preference based on the comment semantic information using an explicit preference tracker;

[0118] An implicit preference tracking module is configured to: obtain the user's implicit preference based on the behavior record using an implicit preference tracker;

[0119] A label generation module is configured to: generate opinion labels corresponding to labels of user preference aspects by using a label generator according to user explicit preferences and user implicit preferences;

[0120] Among them, the explicit preference tracker adopts the information propagation path from the bottom layer to the top layer of the hierarchical heterogeneous graph to obtain the user's explicit preference; the implicit preference tracker adopts multiple types of behavior graphs to obtain the user's implicit preference.

[0121] It should be noted that the data acquisition module, explicit preference tracking module, implicit preference tracking module and label generation module are the same as the examples and application scenarios implemented by the steps in Embodiment 1, but are not limited to the contents disclosed in Embodiment 1. It should be noted that the above modules as part of the system can be executed in a computer system such as a set of computer executable instructions.

[0122] Embodiment 3

[0123] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the method for generating personalized opinion labels for an e-commerce platform as described in the first embodiment above are implemented.

[0124] Embodiment 4

[0125] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for generating personalized opinion labels for an e-commerce platform as described in the first embodiment above are implemented.

[0126] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0127] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0128] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0129] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0130] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0131] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A personalized opinion label generation method for an e-commerce platform, characterized in that: include: Obtain the comment semantic information and behavior records of users on the e-commerce platform; Based on the comment semantic information, an explicit preference tracker is used to obtain the user's explicit preference; Based on the behavior records, an implicit preference tracker is used to obtain the user's implicit preference; According to the user's explicit preference and the user's implicit preference, a label generator is used to generate opinion labels corresponding to the user's preference labels; The explicit preference tracker uses a hierarchical heterogeneous graph to propagate information from the bottom layer to the top layer to obtain the user's explicit preference; the implicit preference tracker uses a multi-class behavior graph to obtain the user's implicit preference; In the hierarchical heterogeneous graph, the information of all nodes comes from the word layer. Each word node in the word layer is a word in the comment, which is encoded through an open source pre-trained vocabulary to obtain an initial representation. Subsequently, the hierarchical heterogeneous graph propagates information between layers from bottom to top. For the user layer, product layer, and aspect layer, the neighbors of each node in the layer are the heterogeneous nodes of the lower layer. Capital letters are used to represent set representations, and lowercase letters are used to represent element representations. For any node i in these three layers, the process of propagating information from the bottom-level neighbor j to i is as follows: h′ i =GAT(h i ,{h j ,j∈N i }) It can be expressed as: e ij =FFN([h i , Wh j ]) Among them, h i represents the representation of node i before updating. FFN(*) is a feedforward neural network. ij A graph attention score is calculated, which assigns different weights to the information of neighboring nodes. The final updated representation of node i is h' i is the weighted sum of all heterogeneous neighbor nodes; Use the IDs of users and products as the initial embeddings of nodes in the multi-class behavior graph, and model different types of edges in the multi-class behavior graph separately; take click behavior as an example, for a given user and product interaction click behavior, the adjacency matrix S c , the user's ID encoding matrix Product ID matrix Where |U|, |P| are the number of users and products in the dataset, d is the ID embedding dimension, and multi-layer LightGCN is used to aggregate the representations of users and products in the behavior graph; user representation The aggregation process is expressed as: Among them, L is the number of layers of LightGCN, D -1 YesS c The product of the diagonal matrix The polymerization process is to use Alternative After obtaining the user and product representations of a single behavior, different types of behavior representations are fused to obtain user and product representations of multiple types of behaviors.

2. The method for generating personalized opinion tags for an e-commerce platform according to claim 1, characterized in that: The aspect tag represents information about a certain aspect of the product.

3. The method for generating personalized opinion tags for an e-commerce platform according to claim 1, characterized in that: The hierarchical heterogeneous graph includes the word layer, aspect layer, product layer and user layer from the bottom layer to the top layer, and defines the path from the bottom layer to the top layer as the WAPU meta-path, and defines the path from the word layer, aspect layer to the user layer as the WAU meta-path.

4. The method for generating personalized opinion tags for an e-commerce platform according to claim 3, characterized in that: The first user preference is obtained through information propagation of the WAPU meta-path; and the error accumulation from the product layer to the user layer in the WAPU is reduced through the WAU meta-path to obtain the second user preference.

5. The method for generating personalized opinion tags for an e-commerce platform according to claim 1, characterized in that: The method of obtaining the user's implicit preferences based on the behavior records by using an implicit preference tracker specifically includes: constructing the multi-category behavior graph based on users, products, clicks, favorites, and orders; based on a single behavior, aggregating the user representation and product representation of the multi-category behavior graph by using a multi-layer aggregation method to obtain a single user representation and a single product representation; based on multi-type behaviors, obtaining the user representation and product representation of multi-type behaviors; Aggregate user representations of multiple types of behaviors of several other users similar to the user to obtain the user's personalized preferences based on behavior; Among them, behaviors include: clicking, collecting and placing orders.

6. The method for generating personalized opinion tags for an e-commerce platform according to claim 5, characterized in that: After obtaining the user representation and product representation, it includes: constructing a behavior UP matrix based on the user representation and product representation learned by LightGCN, aligning the behavior UP matrix with the comment adjacency matrix in the explicit preference tracker, and obtaining the user's implicit preference.

7. The method for generating personalized opinion tags for an e-commerce platform according to claim 1, characterized in that: The step of using a label generator to generate opinion labels corresponding to user preference labels based on user explicit preferences and user implicit preferences specifically includes: Phase 1: Obtain the aspect tags that the user is interested in, use the tag generator based on the first user preference, obtain the probability of each aspect tag, sort the probabilities of the aspect tags, and select the aspect tags with the highest probability. The second stage: using the set number of aspect labels with the highest probability, and using the label generator to obtain the opinion label corresponding to each aspect label.

8. A personalized opinion label generation system for e-commerce platforms, characterized in that: include: A data acquisition module is configured to: acquire comment semantic information and behavior records of users on the e-commerce platform; An explicit preference tracking module is configured to: obtain user explicit preferences based on the comment semantic information using an explicit preference tracker; An implicit preference tracking module is configured to: obtain the user's implicit preference based on the behavior record using an implicit preference tracker; A label generation module is configured to: generate opinion labels corresponding to labels of user preference aspects by using a label generator according to user explicit preferences and user implicit preferences; The explicit preference tracker uses a hierarchical heterogeneous graph to propagate information from the bottom layer to the top layer to obtain the user's explicit preference; the implicit preference tracker uses a multi-class behavior graph to obtain the user's implicit preference; In the hierarchical heterogeneous graph, the information of all nodes comes from the word layer. Each word node in the word layer is a word in the comment, which is encoded through an open source pre-trained vocabulary to obtain an initial representation. Subsequently, the hierarchical heterogeneous graph propagates information between layers from bottom to top. For the user layer, product layer, and aspect layer, the neighbors of each node in the layer are the heterogeneous nodes of the lower layer. Capital letters are used to represent set representations, and lowercase letters are used to represent element representations. For any node i in these three layers, the process of propagating information from the bottom-level neighbor j to i is as follows: h′ i =GAT(h i ,{h j ,j∈N i }) It can be expressed as: e ij =FFN([h i , Wh j ]) Among them, h i represents the representation of node i before updating. FFN(*) is a feedforward neural network. ij A graph attention score is calculated, which assigns different weights to the information of neighboring nodes. The final updated representation h′ of node i i is the weighted sum of all heterogeneous neighbor nodes; Use the IDs of users and products as the initial embeddings of nodes in the multi-class behavior graph, and model different types of edges in the multi-class behavior graph separately; take click behavior as an example, for a given user and product interaction click behavior, the adjacency matrix S c , the user's ID encoding matrix Product ID matrix Where |U|, |P| are the number of users and products in the dataset, d is the ID embedding dimension, and multi-layer LightGCN is used to aggregate the representations of users and products in the behavior graph; user representation The aggregation process is expressed as: Among them, L is the number of layers of LightGCN, D -1 YesS c The product of the diagonal matrix The polymerization process is to use Alternative After obtaining the user and product representations of a single behavior, different types of behavior representations are fused to obtain user and product representations of multiple types of behaviors.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the method for generating personalized opinion labels for an e-commerce platform as described in any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the method for generating personalized opinion labels for an e-commerce platform as described in any one of claims 1-7 are implemented.

Images