User classification method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention provides a user classification method and device, electronic equipment and a storage medium, and relates to the technical field of comment information processing, the electronic equipment can obtain multiple comments sent by a to-be-classified user in a live broadcast room and time sequence information of the multiple comments; performing semantic analysis on each comment to obtain a user emotion reflected by each comment; determining an emotion fluctuation condition based on the user emotion reflected by the plurality of comments and the time sequence information of the plurality of comments; and inputting the live broadcast room behavior characteristics of the to-be-classified users into a user classification model, and outputting user categories by the user classification model based on a pre-learned corresponding relationship between the live broadcast room behavior characteristics and the user categories. In addition, even if the user is a high-value user with a high shopping intention, negative comments may be made, so that classification can be performed based on emotion fluctuation conditions reflected by multiple comments of the same user in a period of time. In this way, the accuracy of classifying the users in the live broadcasting room can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of comment information processing, and in particular to a user classification method, device, electronic device and storage medium. Background Art

[0002] With the development of the internet, livestreaming has become increasingly popular. During livestreaming, the host explains the product and sells it to users in the room. Livestreaming often involves a variety of users, including high-value users with a strong purchasing intent and low-value users who maliciously disrupt the livestream.

[0003] Users can interact with the livestreamer by leaving comments in the livestream room. To ensure that the livestreamer focuses on comments from high-value users, negative comments within the livestream room need to be filtered. Current techniques analyze each comment within the livestream room. If a comment is negative, the user who posted it is marked as a low-quality user, and comments from low-quality users are filtered.

[0004] However, high-value users may also make negative comments. For example, if a high-value user asks a question multiple times but receives no response from the host, they may make a negative comment like, "Why is the host ignoring me?" Current technologies would label these high-quality users as low-quality. Therefore, current technologies cannot accurately classify users within a live broadcast room. Summary of the Invention

[0005] The purpose of the embodiments of the present application is to provide a user classification method, device, electronic device, and storage medium to improve the accuracy of classifying users in a live broadcast room. The specific technical solution is as follows:

[0006] In a first aspect, an embodiment of the present application provides a user classification method, the method comprising:

[0007] Obtain multiple comments sent by the user to be classified in the live broadcast room within a preset time period and time sequence information of the multiple comments;

[0008] Perform semantic analysis on each comment to obtain the user sentiment reflected in each comment;

[0009] Determining the emotional fluctuations of the user to be classified within the preset time period based on the user emotions reflected by the multiple comments and the time sequence information of the multiple comments;

[0010] The live broadcast room behavior characteristics of the user to be classified are input into a pre-trained user classification model, so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior characteristics and the user category, wherein the live broadcast room behavior characteristics at least include the emotional fluctuations, and the user category represents the willingness of the user to be classified to shop in the live broadcast room.

[0011] Optionally, the step of performing semantic analysis on each comment to obtain the user sentiment reflected by each comment includes:

[0012] Each comment is input into a pre-trained semantic analysis model so that the semantic analysis model encodes each comment, obtains a feature vector corresponding to each comment, and determines a sentiment score corresponding to each comment based on the user sentiment reflected by the feature vector.

[0013] Optionally, the step of determining the emotion score corresponding to each comment based on the user emotion reflected by the feature vector includes:

[0014] For each comment, determine the sentiment score corresponding to the comment based on the correspondence between the pre-learned feature vector and the sentiment score; or,

[0015] According to the timestamp order of the multiple comments, for the first comment, the emotion score corresponding to the first comment is determined based on the pre-learned correspondence between the feature vector and the emotion score; for non-first comments, the similarity between the feature vector corresponding to the comment and the feature vector corresponding to the previous comment is calculated, and the emotion score corresponding to the comment is determined based on the similarity and the emotion score corresponding to the previous comment.

[0016] Optionally, the step of determining the emotional fluctuation of the user to be classified within the preset time period based on the user emotions reflected by the multiple comments and the time sequence information of the multiple comments includes:

[0017] For each comment, based on the similarity between the feature vector corresponding to the comment and the feature vector corresponding to the previous comment and the emotion score corresponding to the comment, the emotional fluctuations of the user to be classified within the preset time period are constructed according to a preset feature construction method, wherein the preset feature construction method is determined based on the format of the features that can be processed by the user classification model.

[0018] Optionally, the live broadcast room behavior characteristics further include comment interval characteristics and / or historical live broadcast room purchase characteristics;

[0019] Before the step of inputting the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior characteristics and the user category, the method further includes:

[0020] Calculating, based on the timestamp information of the multiple comments, the variance of the timestamps corresponding to the multiple comments as the comment interval feature; and / or determining, based on the timestamp information of the multiple comments, the time interval between each comment and adjacent comments as the comment interval feature;

[0021] and / or,

[0022] Obtain the historical purchase records of the user to be classified in the live broadcast room;

[0023] Based on the historical purchase records, the historical live broadcast room purchase characteristics of the person to be classified are determined, wherein the historical live broadcast room purchase characteristics include at least one of the total consumption amount, purchase frequency, return rate, relevance of purchased categories, price sensitivity and complaint pattern.

[0024] Optionally, the step of inputting the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior characteristics and the user category includes:

[0025] Inputting the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model;

[0026] The user classification model outputs that the user to be classified is a user with high purchase intention when the emotion fluctuation situation indicates that the user to be classified is always in a positive emotion;

[0027] The user classification model outputs that the user to be classified is a user with low purchase intention when the emotional fluctuation situation indicates that the user to be classified is always in a negative mood;

[0028] The user classification model outputs the user to be classified as a user with high purchase intention when the emotion fluctuation situation indicates that the user to be classified changes from positive emotion to negative emotion and meets the preset screening conditions;

[0029] Among them, the preset screening condition is that the comment interval characteristics of the user to be classified reflect that the time interval between the multiple comments is not a fixed interval and the publishing frequency of the multiple comments is not greater than the preset frequency, or the historical live broadcast room purchase characteristics reflect that the purchasing habits of the user to be classified are consistent with the purchasing habits of users shopping in the live broadcast room.

[0030] Optionally, the training method of the user classification model includes:

[0031] Obtain multiple live broadcast room behavior feature samples and the user category truth value corresponding to each live broadcast room behavior feature sample;

[0032] Inputting the live broadcast room behavior feature sample into an initial user classification model, so that the initial user classification model predicts the user category corresponding to the live broadcast room behavior feature sample based on current model parameters and outputs a user category prediction value;

[0033] Based on the difference between the true value of the user category corresponding to the live broadcast room behavior feature sample and the predicted value of the user category, the parameters of the initial user classification model are adjusted by back propagation until the initial user classification model meets the convergence conditions, thereby obtaining a trained user classification model.

[0034] Optionally, after the step of inputting the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior characteristics and the user category, the method further includes:

[0035] In a case where the user to be classified is a user with high purchase intention, determining, from the plurality of comments, comments to be processed whose corresponding user emotions are positive emotions, wherein the user emotions are determined based on the semantics corresponding to each comment by a pre-trained semantic analysis model;

[0036] The comments to be processed are input into a pre-trained auxiliary answer model, so that the auxiliary answer model extracts questions about the product from the comments to be processed, integrates the extracted questions, and generates auxiliary answers corresponding to the integrated questions based on pre-stored product information, wherein the auxiliary answers are used to assist the anchor in answering questions raised by users with high purchasing intention.

[0037] In a second aspect, an embodiment of the present application provides a user classification device, the device comprising:

[0038] A comment acquisition module is used to acquire multiple comments sent by the user to be classified in the live broadcast room within a preset time period and the time sequence information of the multiple comments;

[0039] Semantic analysis module, used to perform semantic analysis on each comment to obtain the user sentiment reflected in each comment;

[0040] A fluctuation determination module, configured to determine the emotional fluctuation of the user to be classified within the preset time period based on the user emotions reflected by the multiple comments and the time sequence information of the multiple comments;

[0041] The category determination module is used to input the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model, so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior characteristics and the user category, wherein the live broadcast room behavior characteristics at least include the emotional fluctuations, and the user category represents the willingness of the user to be classified to shop in the live broadcast room.

[0042] Optionally, the semantic analysis module includes:

[0043] The semantic analysis submodule is used to input each comment into a pre-trained semantic analysis model so that the semantic analysis model encodes each comment, obtains a feature vector corresponding to each comment, and determines the emotion score corresponding to each comment based on the user emotion reflected by the feature vector.

[0044] Optionally, the semantic analysis submodule includes:

[0045] A first semantic analysis unit is used to determine, for each comment, a corresponding sentiment score of the comment based on a pre-learned correspondence between a feature vector and a sentiment score;

[0046] The second semantic analysis unit is used to determine the emotion score corresponding to the first comment according to the timestamp order of the multiple comments, based on the pre-learned correspondence between the feature vector and the emotion score; for non-first comments, calculate the similarity between the feature vector corresponding to the comment and the feature vector corresponding to the previous comment, and determine the emotion score corresponding to the comment based on the similarity and the emotion score corresponding to the previous comment.

[0047] Optionally, the fluctuation situation determination module includes:

[0048] The fluctuation determination submodule is used to construct the emotional fluctuation of the user to be classified within the preset time period for each comment based on the similarity between the feature vector corresponding to the comment and the feature vector corresponding to the previous comment, as well as the emotional score corresponding to the comment, in accordance with a preset feature construction method, wherein the preset feature construction method is determined based on the format of the features that can be processed by the user classification model.

[0049] Optionally, the live broadcast room behavior characteristics further include comment interval characteristics and / or historical live broadcast room purchase characteristics;

[0050] The device further comprises:

[0051] A first interval feature determination module is configured to calculate, based on the timestamp information of the multiple comments, the variance of the timestamps corresponding to the multiple comments as the comment interval feature;

[0052] a second interval feature determination module, configured to determine, based on the timestamp information of the plurality of comments, a time interval between each comment and an adjacent comment as the comment interval feature;

[0053] and / or,

[0054] A purchase record acquisition module is used to obtain the historical purchase records of the user to be classified in the live broadcast room;

[0055] A purchase feature determination module is used to determine the historical live broadcast room purchase features of the person to be classified based on the historical purchase records, wherein the historical live broadcast room purchase features include at least one of the total consumption amount, purchase frequency, return rate, relevance of purchased categories, price sensitivity and complaint pattern.

[0056] Optionally, the category determination module includes:

[0057] A model input submodule, configured to input the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model;

[0058] A first category determination submodule is configured for the user classification model to output that the user to be classified is a user with high purchase intention when the emotion fluctuation condition indicates that the user to be classified is always in a positive mood;

[0059] A second category determination submodule is configured for the user classification model to output that the user to be classified is a user with low purchase intention when the emotion fluctuation condition indicates that the user to be classified is always in a negative mood;

[0060] A third category determination submodule is configured for the user classification model to output that the user to be classified is a user with high purchase intention when the emotion fluctuation condition indicates that the user to be classified has changed from positive emotion to negative emotion and meets the preset screening conditions;

[0061] Among them, the preset screening condition is that the comment interval characteristics of the user to be classified reflect that the time interval between the multiple comments is not a fixed interval and the publishing frequency of the multiple comments is not greater than the preset frequency, or the historical live broadcast room purchase characteristics reflect that the purchasing habits of the user to be classified are consistent with the purchasing habits of users shopping in the live broadcast room.

[0062] Optionally, the device further includes:

[0063] A sample acquisition module is used to obtain multiple live broadcast room behavior feature samples and the true value of the user category corresponding to each live broadcast room behavior feature sample;

[0064] A sample input module is used to input the live broadcast room behavior feature sample into the initial user classification model, so that the initial user classification model predicts the user category corresponding to the live broadcast room behavior feature sample based on the current model parameters and outputs the user category prediction value;

[0065] The parameter adjustment module is used to adjust the parameters of the initial user classification model by back propagation based on the difference between the true value of the user category corresponding to the live broadcast room behavior feature sample and the predicted value of the user category until the initial user classification model meets the convergence condition, thereby obtaining a trained user classification model.

[0066] Optionally, the device further includes:

[0067] a comment screening module, configured to, when the user to be classified is a user with high purchase intention, determine, from the plurality of comments, comments to be processed whose corresponding user emotions are positive, wherein the user emotions are determined based on the semantics corresponding to each comment using a pre-trained semantic analysis model;

[0068] The auxiliary answer module is used to input the pending comments into a pre-trained auxiliary answer model, so that the auxiliary answer model extracts questions about the product from the pending comments, integrates the extracted questions, and generates auxiliary answers corresponding to the integrated questions based on pre-stored product information, wherein the auxiliary answers are used to assist the anchor in answering questions raised by users with high purchasing intention.

[0069] In a third aspect, an embodiment of the present application provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0070] Memory for storing computer programs;

[0071] The processor is configured to implement any of the method steps described in the first aspect when executing a program stored in the memory.

[0072] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method steps described in any one of the first aspects are implemented.

[0073] Beneficial effects of the embodiments of the present application:

[0074] In the solution provided by the embodiments of the present application, an electronic device can obtain multiple comments and their time sequence information posted by a user to be classified in a live broadcast room within a preset time period; perform semantic analysis on each comment to obtain the user's emotions reflected in each comment; determine the emotional fluctuations of the user to be classified within the preset time period based on the user emotions reflected in the multiple comments and their time sequence information; and input the user's live broadcast room behavioral characteristics into a pre-trained user classification model, so that the user classification model outputs a user category for the user to be classified based on the pre-learned correspondence between the live broadcast room behavioral characteristics and user categories. The live broadcast room behavioral characteristics include at least emotional fluctuations, and the user category represents the user's willingness to shop in the live broadcast room. Even high-value users with a high willingness to shop may post negative comments. For example, if a user posts multiple comments but never receives a response, they may post a comment to complain. Since such users often experience emotional transitions during the process of posting multiple comments, for example, from inquiring to waiting to anxious, classification can be based on the emotional fluctuations reflected in multiple comments from the same user over a period of time. In this way, it is possible to avoid misjudging high-value users as low-value users based solely on a negative comment, and the accuracy of classifying users in the live broadcast room can be improved.

[0075] Of course, it is not necessary to achieve all the advantages described above at the same time when implementing any product or method of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other embodiments can also be obtained based on these drawings.

[0077] Figure 1 A flowchart of a user classification method provided in an embodiment of the present application;

[0078] Figure 2 Based on Figure 1 A flow chart of a method for determining purchase characteristics of a historical live broadcast room according to the illustrated embodiment;

[0079] Figure 3 for Figure 1 A specific flow chart of step S104 in the embodiment shown;

[0080] Figure 4 Based on Figure 1 A flow chart of a method for training a user classification model according to the illustrated embodiment;

[0081] Figure 5 Based on Figure 1 A flow chart of the auxiliary answer generation method of the illustrated embodiment;

[0082] Figure 6 Based on Figure 1 A schematic diagram of a user classification method according to the illustrated embodiment;

[0083] Figure 7 A schematic diagram of the structure of a user classification device provided in an embodiment of the present application;

[0084] Figure 8 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0085] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field based on this application are within the scope of protection of this application.

[0086] In order to improve the accuracy of classifying users in a live broadcast room, the embodiments of the present application provide a user classification method, apparatus, electronic device, computer-readable storage medium, and computer program product. The following first introduces a user classification method provided by the embodiments of the present application.

[0087] The user classification method provided in the embodiment of the present application can be applied to any electronic device that needs to classify users in a live broadcast room, for example, a server, a computer device, etc., without specific limitation here. For the sake of clarity, it is subsequently referred to as an electronic device in this article.

[0088] like Figure 1 As shown, a user classification method includes:

[0089] S101, obtaining multiple comments sent by the user to be classified in the live broadcast room within a preset time period and time sequence information of the multiple comments;

[0090] S102, performing semantic analysis on each comment to obtain the user sentiment reflected by each comment;

[0091] S103, determining the emotional fluctuation of the user to be classified within the preset time period based on the user emotions reflected by the multiple comments and the time sequence information of the multiple comments;

[0092] S104: Inputting the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model, so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior characteristics and the user category.

[0093] Among them, the behavioral characteristics of the live broadcast room at least include the emotional fluctuations, and the user category represents the willingness of the user to be classified to shop in the live broadcast room.

[0094] As can be seen, in embodiments of the present application, an electronic device can obtain multiple comments and their time sequence information posted by a user to be classified in a live broadcast room within a preset time period; perform semantic analysis on each comment to obtain the user's emotions reflected in each comment; determine the emotional fluctuations of the user to be classified within the preset time period based on the user's emotions reflected in the multiple comments and their time sequence information; and input the user's live broadcast room behavioral characteristics into a pre-trained user classification model, so that the user classification model outputs a user category for the user to be classified based on the pre-learned correspondence between the live broadcast room behavioral characteristics and user categories. The live broadcast room behavioral characteristics include at least emotional fluctuations, and the user category represents the user's willingness to shop in the live broadcast room. Even high-value users with a high willingness to shop may post negative comments. For example, if a user posts multiple comments but never receives a response, they may post a comment to complain. Since such users often experience emotional transitions during the process of posting multiple comments, for example, from inquiring to waiting to anxious, classification can be performed based on the emotional fluctuations reflected in multiple comments from the same user over a period of time. In this way, it is possible to avoid misjudging high-value users as low-value users based solely on a negative comment, and the accuracy of classifying users in the live broadcast room can be improved.

[0095] In step S101, the electronic device may obtain multiple comments and time sequence information of the multiple comments sent by the user to be classified in the live broadcast room within a preset time period. The preset time period can be set according to actual needs, for example, 5 minutes, 10 minutes, etc., and is not specifically limited here. The user to be classified can be a user who sent a comment in the live broadcast room.

[0096] The above-mentioned time sequence information can be represented by timestamp information or relative time. For example, the time when the first comment of the user to be classified is posted and the time interval between the posting of non-first comments and the posting of the previous comment can be recorded. Alternatively, the broadcast start time and the time after the broadcast start that each comment is posted can be recorded, etc., and no specific limitation is given here.

[0097] Since multiple comments sent by the same user within a preset time period are usually coherent rather than just a single independent text, the comment content can reflect the user emotions of the user to be classified when the comment is sent. Therefore, in order to determine the emotional fluctuations of the user to be classified within the preset time period, the electronic device can perform semantic analysis on each comment to obtain the user emotions reflected by each comment, that is, execute step S102.

[0098] For example, multiple comments can be input into a semantic analysis model based on BERT (Bidirectional Encoder Representations from Transformers) or other open source models for semantic analysis, and the user sentiment reflected in each comment output by the model can be obtained.

[0099] After obtaining the user emotions reflected by each comment, the electronic device can determine the emotional fluctuations of the user to be classified within a preset time period based on the user emotions reflected by the multiple comments and the time sequence information of the multiple comments, that is, execute step S103.

[0100] For example, assuming a preset time period of 10 minutes, we can obtain multiple comments posted by a user in a livestream room within 10 minutes, along with the time sequence information for each comment. Suppose the user posted comments 1-3 in sequence within 10 minutes. Comment 1 is "Is the clothing size too large or too small?" Comment 2 is "Why is the anchor ignoring me?" Comment 3 is "Is this store a scam?" Then, we can input comments 1-3 into the semantic analysis model and find that the user emotions reflected in comments 1-3 are: inquiry, waiting, and anxiety, respectively. Furthermore, we can determine the user's emotional fluctuations over the 10 minutes: from inquiry to waiting, and then to anxiety.

[0101] Since the live broadcast room behavior characteristics of the user to be classified can reflect the user's willingness to shop in the live broadcast room, the electronic device can input the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model, so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior characteristics and the user category, that is, execute step S104.

[0102] Among them, live broadcast room behavioral characteristics may include at least emotional fluctuations. User categories can represent the willingness of the users to be classified to make purchases in the live broadcast room. The specific division method of user categories can be set according to actual needs. For example, user categories may include users with high purchase intention and users with low purchase intention.

[0103] The above-mentioned user classification model can be a model that can learn the correspondence between live broadcast room behavior characteristics and user categories and perform classification tasks. For example, it can be an RF (Random Forest) model, RNN (Recurrent Neural Network), or Transformer model, which is not specifically limited here.

[0104] It can be seen that in the embodiment of the present application, even if the user is a high-value user with a high willingness to shop, they may still make negative comments. For example, if a user has made comments many times and has never received a response, the user may make a comment to complain. Since such users usually have emotional changes in the process of making multiple comments, for example, from asking to waiting to being anxious, it is possible to classify them based on the emotional fluctuations reflected in multiple comments made by the same user over a period of time. In this way, it is possible to avoid misjudging high-value users as low-value users based solely on a negative comment, and to improve the accuracy of classifying users in the live broadcast room.

[0105] As an implementation of an embodiment of the present application, the above step of performing semantic analysis on each comment to obtain the user emotion reflected by each comment may include:

[0106] Each comment is input into a pre-trained semantic analysis model so that the semantic analysis model encodes each comment, obtains a feature vector corresponding to each comment, and determines a sentiment score corresponding to each comment based on the user sentiment reflected by the feature vector.

[0107] To assess the user sentiment reflected in a comment, the electronic device can input each comment into a pre-trained semantic analysis model. The semantic analysis model then encodes each comment, generating a corresponding feature vector. Based on the user sentiment reflected in the feature vector, the model then determines a sentiment score for each comment. The sentiment score can reflect the degree of user sentiment. For example, the sentiment score can range from 0 to 10, with a higher score indicating more positive user sentiment.

[0108] The semantic analysis model can be a model capable of performing semantic analysis on text, specifically a BERT-based semantic analysis model or other open-source model for semantic analysis, which is not specifically limited here. The semantic analysis model can be trained by first pre-training, i.e., learning a general language representation from large-scale unlabeled text to obtain a pre-trained semantic analysis model. The pre-trained semantic analysis model is then adapted to the task of user emotion recognition. During this process, the parameters of the semantic analysis model are fine-tuned until the semantic analysis model meets the convergence conditions, thereby obtaining a trained semantic analysis model.

[0109] As can be seen, in the embodiments of the present application, the electronic device can input each comment into a pre-trained semantic analysis model, so that the semantic analysis model encodes each comment, obtains a feature vector corresponding to each comment, and determines the emotion score corresponding to each comment based on the user emotion reflected by the feature vector. In this way, the comments can be processed by the semantic analysis model to obtain the emotion score corresponding to the comment. By quantifying user emotions in the form of emotion scores, user emotions can be accurately represented.

[0110] As an implementation of an embodiment of the present application, the step of determining the emotion score corresponding to each comment based on the user emotion reflected by the feature vector may include one of the following two implementations:

[0111] For each comment, determine the sentiment score corresponding to the comment based on the correspondence between the pre-learned feature vector and the sentiment score; or,

[0112] According to the timestamp order of the multiple comments, for the first comment, the emotion score corresponding to the first comment is determined based on the pre-learned correspondence between the feature vector and the emotion score; for non-first comments, the similarity between the feature vector corresponding to the comment and the feature vector corresponding to the previous comment is calculated, and the emotion score corresponding to the comment is determined based on the similarity and the emotion score corresponding to the previous comment.

[0113] In a first embodiment, the semantic analysis model can determine the sentiment score corresponding to each comment based on the pre-learned correspondence between feature vectors and sentiment scores. For example, the electronic device can input comments 1 through 3 into the semantic analysis model in sequence. The semantic analysis model can then encode comments 1 through 3 to obtain feature vectors 1 through 3. Based on the pre-learned correspondence between feature vectors and sentiment scores, the model can determine sentiment scores of 9, 5, and 3 for comments 1 through 3.

[0114] In a second embodiment, the electronic device may further input the timestamp corresponding to each comment into the semantic analysis model. The semantic analysis model may then determine the first comment and non-first comments based on the order of the timestamps of each comment, and perform different processing on the first and non-first comments:

[0115] For the first comment, the semantic analysis model can determine the sentiment score corresponding to the first comment based on the correspondence between the pre-learned feature vector and the sentiment score.

[0116] For non-first comments, the semantic analysis model can calculate the similarity between the feature vector corresponding to the comment and the feature vector corresponding to the previous comment. Next, the semantic analysis model can determine the sentiment score corresponding to the comment based on the similarity corresponding to the non-first comment and the sentiment score corresponding to the previous comment.

[0117] Specifically, the similarity can be cosine similarity, Euclidean distance, Manhattan distance, etc. Since the feature vector can represent the user sentiment reflected in the comment and the textual meaning of the comment, the similarity can simultaneously reflect the similarity between the user sentiment and the textual meaning of the comment and the previous comment.

[0118] For example, the electronic device can input comments 1-comment 3 and the timestamp corresponding to each comment into the semantic analysis model. The semantic analysis model can encode comments 1-comment 3 to obtain feature vectors 1-feature vector 3, and based on the timestamp corresponding to each comment, determine the order of the comments as comment 1, comment 2, and comment 3.

[0119] For comment 1, the semantic analysis model can determine the sentiment score of 8 corresponding to comment 1 based on the pre-learned correspondence between feature vectors and sentiment scores. For comments 2 and 3, the electronic device can first calculate the cosine similarity between feature vector 2 and feature vector 1. If the calculated cosine similarity is 0.5, the semantic analysis model can combine the sentiment score of 8 corresponding to comment 1 to determine the sentiment score of 4 corresponding to comment 2. Next, the electronic device can calculate the cosine similarity between feature vector 3 and feature vector 2. If the calculated cosine similarity is 0.5, the semantic analysis model can combine the sentiment score of 4 corresponding to comment 1 to determine the sentiment score of 2 corresponding to comment 2.

[0120] It can be seen that in the embodiments of the present application, the semantic analysis model can determine the sentiment score corresponding to each comment through one of the above two implementation methods. In the first implementation method, the semantic analysis model can directly determine the sentiment score corresponding to each comment, which is more efficient. In the second implementation method, the semantic analysis model can combine the similarity between the feature vectors corresponding to adjacent comments to determine the sentiment score corresponding to non-first comments, which is more accurate.

[0121] As an implementation of an embodiment of the present application, the step of determining the emotional fluctuation of the user to be classified within the preset time period based on the user emotions reflected by the multiple comments and the time sequence information of the multiple comments may include:

[0122] For each comment, based on the similarity between the feature vector corresponding to the comment and the feature vector corresponding to the previous comment and the emotion score corresponding to the comment, the emotion fluctuation of the user to be classified within the preset time period is constructed according to the preset feature construction method.

[0123] Since sentiment scores can only reflect user sentiment from a single dimension, such as positive, neutral, or negative, in some cases, even if two comments have the same sentiment score, the actual user sentiment may be different. The actual user sentiment can be reflected by the feature vector extracted by the semantic analysis model.

[0124] For example, review 1 reads, "It's expensive, but I still bought it; it's quite good." Review 2 reads, "It's OK, not that great either." Both reviews have a sentiment score of 6, indicating that the user sentiment reflected in both reviews is neutral to positive. Although reviews 1 and 2 have the same sentiment score, the eigenvector 1 and eigenvector 2 for review 1 and 2 reveal that the user sentiment reflected in these reviews is actually different. Specifically, review 1 tends to be positive, while review 2 tends to be perfunctory.

[0125] If the similarity between the feature vectors corresponding to two comments decreases, it indicates that the content or user emotion expressed by the user to be classified has suddenly changed. Therefore, the change in similarity can reflect the outbreak of potential negative emotions. When determining the emotional fluctuations of the user to be classified, the electronic device can construct the emotional fluctuations of the user to be classified within a preset time period based on the similarity between the feature vector corresponding to each comment and the feature vector corresponding to the previous comment, as well as the emotion score corresponding to the comment, according to a preset feature construction method.

[0126] The preset feature construction method can be determined based on the format of features that can be processed by the user classification model. For example, if the user classification model is an RF model, text features can be constructed, such as: "User A's emotional fluctuations within 10 minutes: asking, waiting, anxious", "Feature vector similarity between the first and second comments: 0.7", "Feature vector similarity between the second and third comments: 0.3", and so on.

[0127] As can be seen, in the embodiments of the present application, the electronic device can construct, for each comment, the emotional fluctuations of the user to be classified within a preset time period based on the similarity between the feature vector corresponding to the comment and the feature vector corresponding to the previous comment, as well as the emotion score corresponding to the comment, in accordance with a preset feature construction method. The preset feature construction method is determined based on the format of features that can be processed by the user classification model. Because similarity can simultaneously reflect the similarity of user emotions corresponding to adjacent comments and the similarity of textual meaning, user emotions can be accurately determined based on similarity, thereby obtaining an accurate picture of emotional fluctuations.

[0128] As an implementation method of an embodiment of the present application, the live broadcast room behavior features may further include comment interval features. In this case, before the step of inputting the live broadcast room behavior features of the user to be classified into a pre-trained user classification model so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior features and the user category, the method may further include:

[0129] Calculating the variance of the timestamps corresponding to the multiple comments as the comment interval feature based on the timestamp information of the multiple comments; and / or

[0130] The time interval between each comment and adjacent comments is determined according to the timestamp information of the multiple comments as the comment interval feature.

[0131] When the live broadcast room behavior characteristics include a comment interval characteristic, the electronic device may determine the comment interval characteristic by any of the following three implementation methods:

[0132] In the first embodiment, since the time distribution of comments posted by normal users in a live broadcast room is relatively natural, while malicious bots (automated programs) often post a large number of comments in a short period of time, to determine whether the user posting the comments is a malicious bot, the electronic device can calculate the variance of the timestamps corresponding to the multiple comments based on their timestamp information, using this as a comment interval feature. For example, assuming the variance of the timestamps corresponding to three comments is 10 seconds, the comment interval feature can be: 'Timestamp variance of three comments', Value 10s.

[0133] In the second embodiment, since legitimate users typically post comments as live content progresses, comment intervals are not fixed, while malicious bots (automated programs) typically post comments at fixed intervals. Therefore, to determine whether a commenting user is a malicious bot, the electronic device can determine the time interval between each comment and adjacent comments based on the timestamp information of multiple comments, using this as a comment interval feature. For example, assuming the time interval between the first and second comments is 5 seconds, the comment interval feature can be 'Time interval between the first and second comments', with a value of 5s.

[0134] In a third embodiment, the electronic device can calculate the variance of the timestamps corresponding to the multiple comments based on the timestamp information of the multiple comments, and determine the time interval between each comment and its adjacent comments. Furthermore, the variance of the timestamps corresponding to the multiple comments and the time interval between each comment and its adjacent comments are used as comment interval features.

[0135] Since the third embodiment is a combination of the first embodiment and the second embodiment, the third embodiment will not be described in detail here.

[0136] As can be seen, in the embodiments of the present application, the electronic device can determine the comment interval feature using one of the three aforementioned implementations. Since the comment interval feature takes into account the comment patterns of malicious bots, namely, posting a large number of comments in a short period of time and with fixed comment intervals, malicious bots can be accurately identified using the comment interval feature by calculating the variance of the timestamps of multiple comments and the time intervals between adjacent comments.

[0137] As an implementation method of the embodiment of the present application, the above live broadcast room behavior characteristics may also include historical live broadcast room purchase characteristics. In this case, Figure 2 As shown, before the step of inputting the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior characteristics and the user category, the method may further include:

[0138] S201, obtaining the historical purchase records of the user to be classified in the live broadcast room;

[0139] In some cases, relying solely on multiple comments from a user to be classified may not fully determine the user's willingness to shop in the live broadcast room. Since the user's historical purchase records can reflect the user's shopping habits, the electronic device can obtain the user's historical purchase records in the live broadcast room.

[0140] S202: Based on the historical purchase records, determine the historical live broadcast room purchase characteristics of the person to be classified.

[0141] After obtaining the historical purchase records of the user to be classified, the electronic device can determine the historical live broadcast purchase characteristics of the person to be classified based on the historical purchase records. The historical live broadcast purchase characteristics may include at least one of the following: total consumption amount, purchase frequency, return rate, relevance of purchased categories, price sensitivity, and complaint patterns.

[0142] The following will explain the purchase characteristics of these historical live broadcast rooms respectively:

[0143] The total consumption amount can reflect the total consumption amount of the users to be classified in the live broadcast room over a period of time. If the total consumption amount of the users to be classified is high, then the willingness of the users to be classified to shop in the live broadcast room is usually high. The total consumption amount can be calculated in a variety of ways. For example, the consumption amount of the users to be classified in the live broadcast room over a period of time can be added up to get the total consumption amount. Among them, a period of time can be 2 weeks, 30 days, 60 days, etc. For another example, the total consumption amount Score can be calculated using the following formula activity :

[0144]

[0145] Among them, N is the number of purchases made by the user to be classified in the current live broadcast room within a period of time, p i is the consumption amount of the user to be classified during the i-th shopping in the current live broadcast room, Δt i is the span between the date of the i-th shopping and the current date, and λ is a preset time decay coefficient, which is used to achieve a trend in which shopping records closer to the current date have higher weights.

[0146] The purchase frequency can reflect the frequency with which the users to be classified shop in the live broadcast room over a period of time. If the purchase frequency is high, it means that the users to be classified are more likely to repurchase. In one embodiment, a personalized purchase frequency calculation model can be constructed for the users to be classified. For example, if user A shops once a week, then at a similar time point next week, the probability of user A repurchasing will also be high. For another example, if user B shops once every two weeks and user B has not shopped for two consecutive weeks, then the probability of user B repurchasing at the current time point is also relatively high.

[0147] Purchase frequency can be calculated in a variety of ways. For example, the purchase frequency can be calculated by calculating the ratio between the number of purchases made by the user to be classified in the live broadcast room and the number of days. For another example, the purchase frequency MatchScore of the user to be classified can be calculated using the following formula:

[0148]

[0149] Among them, T now is the current timestamp, t n is the time of the last purchase of the user to be classified, μ Δt is the average of historical purchase time intervals, σ Δt is the standard deviation of historical purchase intervals, and ∈ is a preset minimum value to prevent the denominator from being 0. The closer the MatchScore is to 1, the greater the possibility of repurchase for the user to be classified.

[0150] The return rate reflects the proportion of returned items among the items purchased by the users being categorized over a period of time. For example, the return rate can be calculated by calculating the ratio of the number of returned items to the number of items purchased by the users being categorized over a period of time.

[0151] In some cases, the return habits of the user to be classified follow an N-1 return pattern, meaning they accustomed to purchasing N items, returning N-1 items, and retaining one item. For example, when purchasing clothing, to avoid multiple returns and exchanges, the user to be classified might purchase multiple sizes simultaneously, ultimately retaining only the one that fits best and returning the rest. Although these users to be classified often return items, they still contribute to sales. Therefore, if a user to be classified has an N-1 return pattern, their return behavior can be combined with their purchase behavior when calculating their corresponding return rate. This means the user to be classified is considered to have purchased only one item, and the N-1 returned items are not counted.

[0152] The purchased category correlation can reflect the product categories that the user to be classified tends to purchase. When calculating the purchased category correlation corresponding to the user to be classified, the proportion of the purchase quantity of each product category, such as clothing, beauty products, electronics, etc., can be calculated separately to obtain the purchased category correlation corresponding to each product category. Even if the user to be classified has a high return rate in a certain category, but has a high repurchase rate for the category in the current live broadcast room, then the user to be classified is still a user worthy of special attention.

[0153] Since each product category may contain low-priced products and high-priced products, we can further calculate the purchased category relevance of the low-priced products and the purchased category relevance of the high-priced products in a certain product category for the users to be classified.

[0154] Price sensitivity can reflect the user-to-be-classified's sensitivity to promotions. If the user-to-be-classified makes a purchase when the product price is in the low price range and the low price range occurs less frequently, it can be considered that the user-to-be-classified is price-sensitive. For example, if a live broadcast studio holds a promotional price reduction once a year, the user-to-be-classified will only purchase during the promotional price reduction, indicating that the user-to-be-classified is price-conscious and easily stimulated to buy by price reductions. However, if the low price range occurs more frequently, that is, low prices are more common, then the user-to-be-classified's purchase behavior at that price point does not necessarily mean that they are price-sensitive.

[0155] For example, the following formula can be used to calculate the single-point price sensitivity S(p i ):

[0156]

[0157] Among them, p i is the i-th discrete price point, for example, 199 yuan, 249 yuan, and so on. f(p i ) is the price point p i The frequency of occurrence of can be estimated by the frequency of occurrence of this price point in history. i ) is the user to be classified at price point p i The purchase behavior intensity can be specifically at the price point p i ∈ is a preset minimum value used to avoid the denominator being 0.

[0158] Furthermore, the following formula can be used to calculate the overall price sensitivity S of the users to be classified: norm :

[0159]

[0160] n is the total number of discrete price points. The overall price sensitivity can be understood as: when the frequency of a price point is low, that is, f(p i ) is small, but the user to be classified happens to purchase at this price point, that is, B(p i ) is larger, then the sensitivity of the users to be classified to the price point will be increased. When the frequency of a price point is high, that is, f(p i ) is large, and the number of purchases of the users to be classified is also large, that is, B(p i ) is larger, indicating that the users to be classified have less reaction to price changes and are less price sensitive.

[0161] Complaint patterns can indicate whether the user being classified is a professional bad reviewer. Negative reviews from professional bad reviewers typically use a fixed script and often complain about similar issues across multiple livestreams. Therefore, the user's complaint texts regarding products similar to those sold in the current livestream can be fed into a pre-trained semantic analysis model. The semantic analysis model analyzes the complaint texts and outputs a probability that the user being classified is a professional bad reviewer. If the content of multiple complaint texts from the user being classified is similar, the probability that the user is a professional bad reviewer is high. If the user only complains about a specific product and the complaint text is high quality, this indicates that the user is a rational, high-end user.

[0162] As can be seen, in the embodiments of the present application, the electronic device can obtain the historical purchase records of the user to be classified in the live broadcast room; based on these historical purchase records, the historical live broadcast room purchase characteristics of the user to be classified can be determined, where these historical live broadcast room purchase characteristics include at least one of total spending, purchase frequency, return rate, relevance of purchased categories, price sensitivity, and complaint patterns. In some cases, the user to be classified may be critical in the comment section, for example, commenting, "These pants don't fit me." However, these users are simply particular about product details. If these users have a high repurchase rate for products in a certain category, they are still target customers worthy of attention. If users are classified based solely on a single comment, these users may be missed. Therefore, when classifying users, historical live broadcast room purchase characteristics can also be used as a reference. Since historical live broadcast room purchase characteristics can reflect the shopping habits of the user to be classified, user classification based on historical live broadcast room purchase characteristics can further improve the accuracy of user classification.

[0163] As an implementation method of the present application, Figure 3 As shown, the step of inputting the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior characteristics and the user category may include:

[0164] S301, inputting the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model; if the emotional fluctuation situation indicates that the user to be classified is always in a positive mood, or if the emotional fluctuation situation indicates that the user to be classified changes from a positive emotion to a negative emotion, and if the preset screening conditions are met, executing step S302; if the emotional fluctuation situation indicates that the user to be classified is always in a negative mood, executing step S303;

[0165] The electronic device can input the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model. When the live broadcast room behavior characteristics are manifested in different situations, the user classification model can output the corresponding category of the user to be classified.

[0166] S302, the user classification model outputs that the user to be classified is a user with high purchase intention;

[0167] When the emotional fluctuation indicates that the user to be classified always has positive emotions, it means that multiple comments sent by the user to be classified within the preset time period are all positive comments. Then the user classification model can output that the user to be classified is a user with high purchase intention.

[0168] When the emotional fluctuation indicates that the user to be classified changes from positive emotion to negative emotion and meets the preset screening conditions, the user classification model can also output that the user to be classified is a user with high purchase intention.

[0169] Among them, the preset screening conditions are that the comment interval characteristics of the user to be classified reflect that the time interval between multiple comments is not a fixed interval and the publishing frequency of multiple comments is not greater than the preset frequency, or the historical live broadcast room purchase characteristics reflect that the purchasing habits of the user to be classified are consistent with the purchasing habits of users shopping in the live broadcast room.

[0170] If the user to be classified meets the preset screening criteria, it means that the user to be classified is not using a malicious bot, and the user's historical live broadcast room purchase characteristics indicate that the user to be classified is likely to make purchases in the current live broadcast room. For example, the user's historical live broadcast room purchase characteristics meet at least one of the following: the total consumption amount of the user to be classified reaches the preset total consumption amount threshold, the purchase frequency reaches the preset purchase frequency threshold, the return rate is lower than the preset return rate threshold, the relevance of purchased categories indicates that the user to be classified tends to make purchases in the current live broadcast room, the price sensitivity indicates that the user to be classified is relatively sensitive to the prices of goods in the current live broadcast room, and the complaint pattern indicates that the user to be classified is not a professional bad reviewer.

[0171] In this case, the user to be classified makes a negative comment, which may be because he did not get a timely response from the anchor, or there are indeed problems with the product after-sales service. Therefore, the user classification model can output the user to be classified as a user with high purchase intention.

[0172] S303, the user classification model outputs that the user to be classified is a user with low purchase intention;

[0173] When the emotional fluctuation indicates that the user to be classified always has negative emotions, it means that multiple comments sent by the user to be classified within a preset time period are all negative comments. In this case, the user classification model can output that the user to be classified is a user with low purchase intention.

[0174] As can be seen, in this embodiment of the present application, the user classification model can output the corresponding category of the user to be classified when the behavioral characteristics of the live broadcast room are manifested in different situations. In particular, if the user to be classified changes from positive emotion to negative emotion and meets the preset screening conditions, the user classification model will still output the user to be classified as a user with high purchase intention. This can improve the accuracy of classifying users in the live broadcast room.

[0175] As an implementation method of the present application, Figure 4 As shown, the training method of the above user classification model may include:

[0176] S401, obtaining multiple live broadcast room behavior feature samples and the user category truth value corresponding to each live broadcast room behavior feature sample;

[0177] When training the user classification model, the electronic device can obtain multiple live broadcast room behavior feature samples and the user category true value corresponding to each live broadcast room behavior feature sample.

[0178] S402: Input the live broadcast room behavior feature sample into an initial user classification model, so that the initial user classification model predicts the user category corresponding to the live broadcast room behavior feature sample based on current model parameters and outputs a user category prediction value;

[0179] Next, the electronic device can input the live broadcast room behavior feature samples into the initial user classification model. The initial user classification model will predict the user category corresponding to the live broadcast room behavior feature samples based on the current model parameters and output the user category prediction value.

[0180] S403, based on the difference between the true value of the user category corresponding to the live broadcast room behavior feature sample and the predicted value of the user category, the parameters of the initial user classification model are adjusted by back propagation until the initial user classification model meets the convergence conditions, thereby obtaining a trained user classification model.

[0181] Since the true value of the user category is the actual value of the user category, and the predicted value of the user category is predicted by the initial user classification model, in order to improve the accuracy of the user category output by the initial user classification model, the electronic device can adjust the parameters of the initial user classification model through back propagation based on the difference between the true value of the user category and the predicted value of the user category corresponding to the live broadcast room behavior feature sample.

[0182] After the parameter adjustment is complete, the electronic device can determine whether the initial user classification model meets the convergence criteria. If not, the process returns to step S401 and continues training the initial user classification model using the next live broadcast room behavior feature sample and the true value of the user category corresponding to the live broadcast room behavior feature sample. If the convergence criteria are met, the process stops, resulting in a fully trained user classification model.

[0183] It can be seen that in the embodiment of the present application, the electronic device can obtain multiple live broadcast room behavior feature samples and the true value of the user category corresponding to each live broadcast room behavior feature sample; input the live broadcast room behavior feature samples into the initial user classification model, so that the initial user classification model predicts the user category corresponding to the live broadcast room behavior feature sample based on the current model parameters and outputs the user category prediction value; based on the difference between the true value of the user category corresponding to the live broadcast room behavior feature sample and the user category prediction value, the parameters of the initial user classification model are adjusted through back propagation until the initial user classification model meets the convergence condition, thereby obtaining a trained user classification model. In this way, the user classification model can learn the correspondence between the live broadcast room behavior features and the user category, and then accurately output the user category of the user to be classified.

[0184] As an implementation method of the present application, Figure 5 As shown, after the step of inputting the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior characteristics and the user category, the method may further include:

[0185] S501, when the user to be classified is a user with high purchase intention, determining, from the plurality of comments, comments to be processed whose corresponding user emotions are positive emotions;

[0186] If the user being categorized as having a high purchase intent is considered a purchaser, the anchor can prioritize responding to their comments. This user's multiple comments may include both positive and negative ones. Since positive comments often include questions related to the product, while negative comments are often unrelated, the anchor only needs to respond to positive comments.

[0187] To filter out positive comments from the multiple comments posted by the user, the electronic device may determine comments with positive user emotions from the multiple comments posted by the user as comments to be processed. The user emotions may be determined based on the semantics of each comment using a pre-trained semantic analysis model.

[0188] S502: Input the pending comments into a pre-trained auxiliary answer model, so that the auxiliary answer model extracts questions about the product from the pending comments, integrates the extracted questions, and generates auxiliary answers corresponding to the integrated questions based on pre-stored product information.

[0189] To improve the host's efficiency in responding to comments, an auxiliary answer model can be pre-trained. This auxiliary answer model can be a model that can store product information and process text content. For example, it can be an LLM (Large Language Model).

[0190] In this case, after the electronic device identifies the pending reviews, it can input them into the auxiliary answering model. The auxiliary answering model can then extract questions about the product from the pending reviews, integrate the extracted questions, and generate a review summary, thus eliminating the need for the anchor to manually read the reviews one by one.

[0191] After obtaining a review summary, the auxiliary answer model can generate auxiliary answers corresponding to the review summary based on pre-stored product information. These auxiliary answers can help the host answer questions from users with high purchase intent, improving the host's response efficiency to reviews.

[0192] For example, suppose user A, who has a high purchase intent, posts comment 1, "Does this shoe run large or small?" and comment 2, "Is it suitable for running?" Comments 1 and 2 can be fed into the auxiliary answer model, which extracts the comment summaries for comments 1 and 2: "The user is interested in the shoe's size and intended use (running)." The auxiliary answer model then generates the auxiliary answer, "The size runs large, but is suitable for running."

[0193] For example, suppose user B, who has a high purchase intention, posts comment 3, "Does this headset support noise cancellation?" and comment 4, "Is it active noise cancellation or passive noise cancellation?" Comments 3 and 4 can then be fed into the auxiliary answer model, which extracts the comment summaries of comments 3 and 4: "Does this headset support noise cancellation?" The auxiliary answer model then generates the auxiliary answer, "Yes, this headset supports active noise cancellation, which effectively reduces ambient noise."

[0194] It can be seen that in the embodiment of the present application, by generating comment summaries and auxiliary answers for multiple comments, the pressure on the host can be reduced and the efficiency of the host's response to comments can be improved. In addition, by combining the semantic analysis model, the user classification model and the auxiliary answer model, the comments can be pre-screened by means of the semantic analysis model and machine learning (Machine Learning), and negative comments can be efficiently filtered out to avoid the host receiving too much worthless information. Furthermore, only the comments of users with high purchasing intentions can be input into the auxiliary answer model without the need to input all comments into the auxiliary answer model. In this way, the computational cost of the auxiliary answer model can be greatly reduced (by about 80%-90%), and worthless comments can be prevented from occupying the computing resources of the auxiliary answer model.

[0195] As an implementation method of the embodiment of the present application, a schematic diagram of a user classification method can be as follows: Figure 6 As shown, the electronic device can obtain comments 1 and 2 and timestamps of user A to be classified to form a session of user A to be classified. Comment 1 contains the content "Does this dress come in other colors?" with a timestamp of 00:03, and comment 2 contains the content "I just placed an order, when will it be shipped?" with a timestamp of 00:42.

[0196] The electronic device can input Comment 1, Comment 2, and the timestamp into the semantic analysis model. Through the semantic analysis model encoding, the feature vector of Comment 1 is obtained as [0.12, -0.34, 0.56, …, -0.87], and the feature vector of Comment 2 is [-0.32, -0.14, 0.16, …, 0.37]. Next, feature construction can be performed on Comment 1 and Comment 2, that is, the similarity between the feature vectors of adjacent comments is calculated. The similarity corresponding to Comment 1 is 1, and the similarity between the feature vector corresponding to Comment 2 and the feature vector corresponding to Comment 1 is 0.32. In addition, the above comments can be statistically analyzed as a whole to determine the total number of comments and the proportion of negative comments.

[0197] The electronic device can also obtain the purchase and return records of the user A to be classified, including: the purchased item is facial cleanser, the purchase time is 2025-4-22, and the return situation is none; the purchased item is sports shoes, the purchase time is 2025-4-23, and the return situation is none, etc.

[0198] By constructing features of purchase and return records, we can obtain the historical live broadcast purchase features of the user A to be classified. Specifically, the purchase behavior statistics of the user A to be classified are obtained, and the total number of purchases is 5, the number of purchases in the past 30 days is 2, the purchase frequency is 15, and the purchase time variance is 8. For return behavior, the return rate is 0.2, the return rate of sports shoes is 0.5, and the return rate of facial cleanser is 0. For product preferences, skin care products account for 60, sports shoes account for 40, the average purchase price is 180, and the highest purchase price is 600. For live broadcast product matching, the similarity is 0.85, indicating that the similarity between the product categories purchased historically by the user A to be classified and the product categories in the current live broadcast room is 0.85, the brand match is true, and the category match is also true.

[0199] Next, the electronic device can input the above characteristics of user A to be classified into a user classification model, which can then output a user category for user A to be classified. The user category can be a user with high purchase intent or a user with low purchase intent. If user A to be classified has high purchase intent, multiple comments from user A can be input into an auxiliary answer model. The auxiliary answer model generates comment summaries and auxiliary answers for the host to interact with user A to be classified.

[0200] In the technical solution of this application, the operations involved in obtaining, storing, using, processing, transmitting, providing and disclosing user personal information are all carried out with the user's authorization.

[0201] Corresponding to the above-mentioned user classification method, the embodiment of the present application further provides a user classification device. The user classification device provided by the embodiment of the present application is introduced below.

[0202] like Figure 7 As shown, a user classification device includes:

[0203] The comment acquisition module 701 is used to acquire multiple comments sent by the user to be classified in the live broadcast room within a preset time period and the time sequence information of the multiple comments;

[0204] Semantic analysis module 702, used to perform semantic analysis on each comment to obtain the user sentiment reflected by each comment;

[0205] Fluctuation determination module 703, configured to determine the emotion fluctuation of the user to be classified within the preset time period based on the user emotions reflected by the multiple comments and the time sequence information of the multiple comments;

[0206] The category determination module 704 is used to input the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model, so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior characteristics and the user category, wherein the live broadcast room behavior characteristics at least include the emotional fluctuations, and the user category represents the willingness of the user to be classified to shop in the live broadcast room.

[0207] As can be seen, in embodiments of the present application, an electronic device can obtain multiple comments and their time sequence information posted by a user to be classified in a live broadcast room within a preset time period; perform semantic analysis on each comment to obtain the user's emotions reflected in each comment; determine the emotional fluctuations of the user to be classified within the preset time period based on the user's emotions reflected in the multiple comments and their time sequence information; and input the user's live broadcast room behavioral characteristics into a pre-trained user classification model, so that the user classification model outputs a user category for the user to be classified based on the pre-learned correspondence between the live broadcast room behavioral characteristics and user categories. The live broadcast room behavioral characteristics include at least emotional fluctuations, and the user category represents the user's willingness to shop in the live broadcast room. Even high-value users with a high willingness to shop may post negative comments. For example, if a user posts multiple comments but never receives a response, they may post a comment to complain. Since such users often experience emotional transitions during the process of posting multiple comments, for example, from inquiring to waiting to anxious, classification can be performed based on the emotional fluctuations reflected in multiple comments from the same user over a period of time. In this way, it is possible to avoid misjudging high-value users as low-value users based solely on a negative comment, and the accuracy of classifying users in the live broadcast room can be improved.

[0208] As an implementation of the embodiment of the present application, the semantic analysis module 702 may include:

[0209] The semantic analysis submodule is used to input each comment into a pre-trained semantic analysis model so that the semantic analysis model encodes each comment, obtains a feature vector corresponding to each comment, and determines the emotion score corresponding to each comment based on the user emotion reflected by the feature vector.

[0210] As an implementation of an embodiment of the present application, the semantic analysis submodule may include:

[0211] A first semantic analysis unit is used to determine, for each comment, a corresponding sentiment score of the comment based on a pre-learned correspondence between a feature vector and a sentiment score;

[0212] The second semantic analysis unit is used to determine the emotion score corresponding to the first comment according to the timestamp order of the multiple comments, based on the pre-learned correspondence between the feature vector and the emotion score; for non-first comments, calculate the similarity between the feature vector corresponding to the comment and the feature vector corresponding to the previous comment, and determine the emotion score corresponding to the comment based on the similarity and the emotion score corresponding to the previous comment.

[0213] As an implementation of an embodiment of the present application, the fluctuation condition determination module 703 may include:

[0214] The fluctuation determination submodule is used to construct the emotional fluctuation of the user to be classified within the preset time period for each comment based on the similarity between the feature vector corresponding to the comment and the feature vector corresponding to the previous comment, as well as the emotional score corresponding to the comment, in accordance with a preset feature construction method, wherein the preset feature construction method is determined based on the format of the features that can be processed by the user classification model.

[0215] As an implementation of an embodiment of the present application, the live broadcast room behavior feature further includes a comment interval feature and / or a historical live broadcast room purchase feature. In this case, the apparatus may further include:

[0216] A first interval feature determination module is configured to calculate, based on the timestamp information of the multiple comments, the variance of the timestamps corresponding to the multiple comments as the comment interval feature;

[0217] a second interval feature determination module, configured to determine, based on the timestamp information of the plurality of comments, a time interval between each comment and an adjacent comment as the comment interval feature;

[0218] and / or,

[0219] A purchase record acquisition module is used to obtain the historical purchase records of the user to be classified in the live broadcast room;

[0220] A purchase feature determination module is used to determine the historical live broadcast room purchase features of the person to be classified based on the historical purchase records, wherein the historical live broadcast room purchase features include at least one of the total consumption amount, purchase frequency, return rate, relevance of purchased categories, price sensitivity and complaint pattern.

[0221] As an implementation of the embodiment of the present application, the category determination module 704 may include:

[0222] A model input submodule, configured to input the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model;

[0223] A first category determination submodule is configured for the user classification model to output that the user to be classified is a user with high purchase intention when the emotion fluctuation condition indicates that the user to be classified is always in a positive mood;

[0224] A second category determination submodule is configured for the user classification model to output that the user to be classified is a user with low purchase intention when the emotion fluctuation condition indicates that the user to be classified is always in a negative mood;

[0225] A third category determination submodule is configured for the user classification model to output that the user to be classified is a user with high purchase intention when the emotion fluctuation condition indicates that the user to be classified has changed from positive emotion to negative emotion and meets the preset screening conditions;

[0226] Among them, the preset screening condition is that the comment interval characteristics of the user to be classified reflect that the time interval between the multiple comments is not a fixed interval and the publishing frequency of the multiple comments is not greater than the preset frequency, or the historical live broadcast room purchase characteristics reflect that the purchasing habits of the user to be classified are consistent with the purchasing habits of users shopping in the live broadcast room.

[0227] As an implementation of the embodiment of the present application, the above-mentioned device may further include:

[0228] A sample acquisition module is used to obtain multiple live broadcast room behavior feature samples and the true value of the user category corresponding to each live broadcast room behavior feature sample;

[0229] A sample input module is used to input the live broadcast room behavior feature sample into the initial user classification model, so that the initial user classification model predicts the user category corresponding to the live broadcast room behavior feature sample based on the current model parameters and outputs the user category prediction value;

[0230] The parameter adjustment module is used to adjust the parameters of the initial user classification model by back propagation based on the difference between the true value of the user category corresponding to the live broadcast room behavior feature sample and the predicted value of the user category until the initial user classification model meets the convergence condition, thereby obtaining a trained user classification model.

[0231] As an implementation of the embodiment of the present application, the above-mentioned device may further include:

[0232] A comment screening module is configured to, when the user to be classified is a user with high purchase intention, determine, from the plurality of comments, comments to be processed whose corresponding user emotions are positive emotions, wherein the user emotions are determined based on the semantics corresponding to each comment using a pre-trained semantic analysis model;

[0233] The auxiliary answer module is used to input the pending comments into a pre-trained auxiliary answer model, so that the auxiliary answer model extracts questions about the product from the pending comments, integrates the extracted questions, and generates auxiliary answers corresponding to the integrated questions based on pre-stored product information, wherein the auxiliary answers are used to assist the anchor in answering questions raised by users with high purchasing intention.

[0234] The present application also provides an electronic device, such as Figure 8 Shown, including:

[0235] Memory 801, used for storing computer programs;

[0236] The processor 802 is configured to implement the user classification method described in any of the above embodiments when executing the program stored in the memory 801 .

[0237] Furthermore, the electronic device may further include a communication bus and / or a communication interface, and the processor 802, the communication interface, and the memory 801 communicate with each other via the communication bus.

[0238] As can be seen, in embodiments of the present application, an electronic device can obtain multiple comments and their time sequence information posted by a user to be classified in a live broadcast room within a preset time period; perform semantic analysis on each comment to obtain the user's emotions reflected in each comment; determine the emotional fluctuations of the user to be classified within the preset time period based on the user's emotions reflected in the multiple comments and their time sequence information; and input the user's live broadcast room behavioral characteristics into a pre-trained user classification model, so that the user classification model outputs a user category for the user to be classified based on the pre-learned correspondence between the live broadcast room behavioral characteristics and user categories. The live broadcast room behavioral characteristics include at least emotional fluctuations, and the user category represents the user's willingness to shop in the live broadcast room. Even high-value users with a high willingness to shop may post negative comments. For example, if a user posts multiple comments but never receives a response, they may post a comment to complain. Since such users often experience emotional transitions during the process of posting multiple comments, for example, from inquiring to waiting to anxious, classification can be performed based on the emotional fluctuations reflected in multiple comments from the same user over a period of time. In this way, it is possible to avoid misjudging high-value users as low-value users based solely on a negative comment, and the accuracy of classifying users in the live broadcast room can be improved.

[0239] The communication bus mentioned in the electronic device mentioned above may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. This communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, only one thick line is used in the figure, but this does not mean that there is only one bus or only one type of bus.

[0240] The communication interface is used for communication between the above electronic device and other devices.

[0241] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk memory. Alternatively, the memory may be at least one storage device located away from the processor.

[0242] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

[0243] In another embodiment provided by the present application, a computer-readable storage medium is further provided, wherein a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, the steps of any of the above-mentioned user classification methods are implemented.

[0244] In another embodiment provided by the present application, a computer program product including instructions is also provided, which, when executed on a computer, enables the computer to execute any user classification method in the above embodiments.

[0245] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When software is used for implementation, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website, computer, server or data center to another website, computer, server or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrations. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0246] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0247] Each embodiment in this specification is described in a related manner. Similar portions between embodiments can be referenced to each other. Each embodiment focuses on the differences between other embodiments. In particular, the device, electronic device, computer-readable storage medium, and computer program product embodiments are generally similar to the method embodiments, so their descriptions are relatively simple. For related portions, reference can be made to the descriptions of the method embodiments.

[0248] The above description is only a preferred embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application are included in the scope of protection of the present application.

Claims

1. A user classification method, characterized in that: The method comprises: Obtain multiple comments sent by the user to be classified in the live broadcast room within a preset time period and time sequence information of the multiple comments; Perform semantic analysis on each comment to obtain the user sentiment reflected in each comment; Determining the emotional fluctuations of the user to be classified within the preset time period based on the user emotions reflected by the multiple comments and the time sequence information of the multiple comments; The live broadcast room behavior characteristics of the user to be classified are input into a pre-trained user classification model, so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior characteristics and the user category, wherein the live broadcast room behavior characteristics at least include the emotional fluctuations, and the user category represents the willingness of the user to be classified to shop in the live broadcast room.

2. The method according to claim 1, characterized in that The step of performing semantic analysis on each comment to obtain the user sentiment reflected by each comment includes: Each comment is input into a pre-trained semantic analysis model so that the semantic analysis model encodes each comment, obtains a feature vector corresponding to each comment, and determines a sentiment score corresponding to each comment based on the user sentiment reflected by the feature vector.

3. The method according to claim 2, characterized in that The step of determining the emotion score corresponding to each comment based on the user emotion reflected by the feature vector includes: For each comment, determine the sentiment score corresponding to the comment based on the correspondence between the pre-learned feature vector and the sentiment score; or, According to the timestamp order of the multiple comments, for the first comment, the emotion score corresponding to the first comment is determined based on the pre-learned correspondence between the feature vector and the emotion score; for non-first comments, the similarity between the feature vector corresponding to the comment and the feature vector corresponding to the previous comment is calculated, and the emotion score corresponding to the comment is determined based on the similarity and the emotion score corresponding to the previous comment.

4. The method according to claim 3, characterized in that The step of determining the emotional fluctuation of the user to be classified within the preset time period based on the user emotions reflected by the multiple comments and the time sequence information of the multiple comments includes: For each comment, based on the similarity between the feature vector corresponding to the comment and the feature vector corresponding to the previous comment and the emotion score corresponding to the comment, the emotional fluctuations of the user to be classified within the preset time period are constructed according to a preset feature construction method, wherein the preset feature construction method is determined based on the format of the features that can be processed by the user classification model.

5. The method according to any one of claims 1 to 4, characterized in that The live broadcast room behavior characteristics also include comment interval characteristics and / or historical live broadcast room purchase characteristics; Before the step of inputting the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior characteristics and the user category, the method further includes: Calculating, based on the timestamp information of the multiple comments, the variance of the timestamps corresponding to the multiple comments as the comment interval feature; and / or determining, based on the timestamp information of the multiple comments, the time interval between each comment and adjacent comments as the comment interval feature; and / or, Obtain the historical purchase records of the user to be classified in the live broadcast room; Based on the historical purchase records, the historical live broadcast room purchase characteristics of the person to be classified are determined, wherein the historical live broadcast room purchase characteristics include at least one of the total consumption amount, purchase frequency, return rate, relevance of purchased categories, price sensitivity and complaint pattern.

6. The method according to claim 5, characterized in that The step of inputting the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior characteristics and the user category, includes: Inputting the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model; The user classification model outputs that the user to be classified is a user with high purchase intention when the emotion fluctuation situation indicates that the user to be classified is always in a positive emotion; The user classification model outputs that the user to be classified is a user with low purchase intention when the emotional fluctuation situation indicates that the user to be classified is always in a negative mood; The user classification model outputs the user to be classified as a user with high purchase intention when the emotion fluctuation situation indicates that the user to be classified changes from positive emotion to negative emotion and meets the preset screening conditions; Among them, the preset screening condition is that the comment interval characteristics of the user to be classified reflect that the time interval between the multiple comments is not a fixed interval and the publishing frequency of the multiple comments is not greater than the preset frequency, or the historical live broadcast room purchase characteristics reflect that the purchasing habits of the user to be classified are consistent with the purchasing habits of users shopping in the live broadcast room.

7. The method according to any one of claims 1 to 4, characterized in that The training method of the user classification model includes: Obtain multiple live broadcast room behavior feature samples and the user category truth value corresponding to each live broadcast room behavior feature sample; Inputting the live broadcast room behavior feature sample into an initial user classification model, so that the initial user classification model predicts the user category corresponding to the live broadcast room behavior feature sample based on current model parameters and outputs a user category prediction value; Based on the difference between the true value of the user category corresponding to the live broadcast room behavior feature sample and the predicted value of the user category, the parameters of the initial user classification model are adjusted by back propagation until the initial user classification model meets the convergence conditions, thereby obtaining a trained user classification model.

8. The method according to any one of claims 1 to 4, characterized in that After the step of inputting the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior characteristics and the user category, the method further includes: In a case where the user to be classified is a user with high purchase intention, determining, from the plurality of comments, comments to be processed whose corresponding user emotions are positive emotions, wherein the user emotions are determined based on the semantics corresponding to each comment by a pre-trained semantic analysis model; The comments to be processed are input into a pre-trained auxiliary answer model, so that the auxiliary answer model extracts questions about the product from the comments to be processed, integrates the extracted questions, and generates auxiliary answers corresponding to the integrated questions based on pre-stored product information, wherein the auxiliary answers are used to assist the anchor in answering questions raised by users with high purchasing intention.

9. A user classification device, characterized in that: The device comprises: A comment acquisition module is used to acquire multiple comments sent by the user to be classified in the live broadcast room within a preset time period and the time sequence information of the multiple comments; Semantic analysis module, used to perform semantic analysis on each comment to obtain the user sentiment reflected in each comment; A fluctuation determination module, configured to determine the emotional fluctuation of the user to be classified within the preset time period based on the user emotions reflected by the multiple comments and the time sequence information of the multiple comments; The category determination module is used to input the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model, so that the user classification model outputs the user category of the user to be classified based on the pre-learned correspondence between the live broadcast room behavior characteristics and the user category, wherein the live broadcast room behavior characteristics at least include the emotional fluctuations, and the user category represents the willingness of the user to be classified to shop in the live broadcast room.

10. The device according to claim 9, characterized in that The semantic analysis module includes: A semantic analysis submodule is configured to input each comment into a pre-trained semantic analysis model so that the semantic analysis model encodes each comment, obtains a feature vector corresponding to each comment, and determines a sentiment score corresponding to each comment based on the user sentiment reflected by the feature vector; and / or, The semantic analysis submodule includes: A first semantic analysis unit is used to determine, for each comment, a corresponding sentiment score of the comment based on a pre-learned correspondence between a feature vector and a sentiment score; A second semantic analysis unit is configured to determine, for the first comment, the sentiment score corresponding to the first comment according to the pre-learned correspondence between the feature vector and the sentiment score in the order of the timestamps of the plurality of comments; for non-first comments, calculate the similarity between the feature vector corresponding to the comment and the feature vector corresponding to the previous comment, and determine the sentiment score corresponding to the comment based on the similarity and the sentiment score corresponding to the previous comment; and / or, The fluctuation situation determination module includes: A fluctuation determination submodule is configured to construct, for each comment, the emotional fluctuation of the user to be classified within the preset time period based on the similarity between the feature vector corresponding to the comment and the feature vector corresponding to the previous comment, as well as the emotional score corresponding to the comment, in accordance with a preset feature construction method, wherein the preset feature construction method is determined based on the format of features that can be processed by the user classification model; and / or, The live broadcast room behavior characteristics also include comment interval characteristics and / or historical live broadcast room purchase characteristics; The device further comprises: A first interval feature determination module is configured to calculate, based on the timestamp information of the multiple comments, the variance of the timestamps corresponding to the multiple comments as the comment interval feature; a second interval feature determination module, configured to determine, based on the timestamp information of the plurality of comments, a time interval between each comment and an adjacent comment as the comment interval feature; and / or, A purchase record acquisition module is used to obtain the historical purchase records of the user to be classified in the live broadcast room; a purchase feature determination module, configured to determine, based on the historical purchase records, the historical live broadcast room purchase features of the person to be classified, wherein the historical live broadcast room purchase features include at least one of the following: total consumption amount, purchase frequency, return rate, relevance of purchased categories, price sensitivity, and complaint pattern; and / or, The category determination module includes: A model input submodule, configured to input the live broadcast room behavior characteristics of the user to be classified into a pre-trained user classification model; A first category determination submodule is configured for the user classification model to output that the user to be classified is a user with high purchase intention when the emotion fluctuation condition indicates that the user to be classified is always in a positive mood; A second category determination submodule is configured for the user classification model to output that the user to be classified is a user with low purchase intention when the emotion fluctuation condition indicates that the user to be classified is always in a negative mood; A third category determination submodule is configured for the user classification model to output that the user to be classified is a user with high purchase intention when the emotion fluctuation condition indicates that the user to be classified has changed from positive emotion to negative emotion and meets the preset screening conditions; Among them, the preset screening condition is that the comment interval characteristics of the user to be classified reflect that the time intervals of the multiple comments are not fixed and the publishing frequency of the multiple comments is not greater than the preset frequency, or the historical live broadcast room purchase characteristics reflect that the purchasing habits of the user to be classified are consistent with the purchasing habits of users who shop in the live broadcast room; and / or, The device further comprises: A sample acquisition module is used to obtain multiple live broadcast room behavior feature samples and the true value of the user category corresponding to each live broadcast room behavior feature sample; A sample input module is used to input the live broadcast room behavior feature sample into the initial user classification model, so that the initial user classification model predicts the user category corresponding to the live broadcast room behavior feature sample based on the current model parameters and outputs the user category prediction value; A parameter adjustment module is used to adjust the parameters of the initial user classification model by back propagation based on the difference between the true value of the user category corresponding to the live broadcast room behavior feature sample and the predicted value of the user category until the initial user classification model meets the convergence condition to obtain a trained user classification model; and / or, The device further comprises: a comment screening module, configured to, when the user to be classified is a user with high purchase intention, determine, from the plurality of comments, comments to be processed whose corresponding user emotions are positive, wherein the user emotions are determined based on the semantics corresponding to each comment using a pre-trained semantic analysis model; The auxiliary answer module is used to input the pending comments into a pre-trained auxiliary answer model, so that the auxiliary answer model extracts questions about the product from the pending comments, integrates the extracted questions, and generates auxiliary answers corresponding to the integrated questions based on pre-stored product information, wherein the auxiliary answers are used to assist the anchor in answering questions raised by users with high purchasing intention.

11. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; Memory for storing computer programs; A processor, configured to implement the method steps described in any one of claims 1 to 8 when executing a program stored in a memory.

12. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method steps of any one of claims 1 to 8 are implemented.

Images