Product classification model training and predicting method based on digital human

Abstract

The invention provides a product classification model training and predicting method based on a digital human, and belongs to the technical field of product classification and digital human. The training method comprises a training corpus collection step, a training corpus classification step, a digital human generation step, a model training step and a digital human model association step. The prediction method comprises the following steps: S710, receiving information of a to-be-classified product input by a user; s720, analyzing the information of the to-be-classified products to obtain a plurality of semantic segmented words; s730, taking the semantic segmented words as the input of the target product classification model; and S740, outputting a product classification result through a digital human model associated with the target product classification model. According to the technical scheme provided by the invention, the output result of the target product classification model can be output through the digital human, so that the output information of the product classification result is richer and more visual, the generated digital human model and the product classification model also have relatively strong relevance, and the user experience is good.

Description

Technical Field

[0001] The present invention belongs to the technical field of product classification and digital human, and in particular, relates to a product classification model training and prediction method based on digital human, a computer-readable storage medium for implementing the method, a computer program product and an electronic device. Background Art

[0002] Product type plays a moderating role in the impact on consumers' perceived value and perceived trust. Content-based recommendations and search product matching, collaborative filtering recommendations and experience product matching can make consumers perceive higher trust and perceived value. Therefore, product classification can achieve more accurate personalized recommendations, thereby improving user stickiness.

[0003] Product classification is a process of classifying the attributes of a product in multiple dimensions based on the existing rough product information, combined with the existing relevant network shared big data of the product or similar products, including product review information, product sales information, product logistics information, etc. When the user currently only obtains some rough or vague information about the product, such as only a name, a picture, a video, or an advertising screen, the need is to obtain more dimensional information about the product, and then obtain the possible classification of the product in multiple dimensions. The traditional method is often through a search engine, where users enter keywords to filter information and then classify it. This method is simple and direct, but the search results are not ideal and are not easy for users to accept and understand.

[0004] With the development of deep learning and artificial intelligence technology, it is now possible to automatically identify product feature information through classification models, provide reference information of multiple dimensions of the product based on shared big data, and then classify it from different angles, such as the product classification method proposed in Chinese invention patent publications CN118314403A and CN113553433A.

[0005] However, the current product classification method still only roughly gives a certain category attribute of the product from a single dimension, and the data category is not rich enough; at the same time, the classification results of the products to be classified with different attributes output by a single training model are also relatively boring in visualization, lacking affinity, and providing a poor user experience. Summary of the invention

[0006] In view of the above technical problems, the present invention proposes a product classification model training and prediction method based on digital human, a computer-readable storage medium, a computer program product and an electronic device for implementing the method.

[0007] In the first aspect of the present invention, a product classification model training method based on digital human is proposed, the training method comprises a training corpus collection step, a training corpus classification step, a digital human generation step, a model training step and a digital human model association step, each step is specifically implemented as follows:

[0008] Training corpus collection steps: Collect training corpus from product details and product reviews;

[0009] Training corpus classification step: dividing the training corpus into digital human generated corpus and model training corpus;

[0010] Digital human generation step: generating a digital human model based on the digital human generation corpus;

[0011] Model training step: training a benchmark product classification model based on the model training corpus to obtain a target product classification model;

[0012] Digital human model association step: Associating the digital human model with the target product classification model so that the output result of the target product classification model is output by the digital human model.

[0013] The output result of the target product classification model includes product information of multiple different dimensions; after obtaining the output result of the target product classification model, the method further includes:

[0014] Based on the target classification requirements, one or more dimensional information is selected from the multiple dimensional information, and based on the selected one or more dimensional information, the target classification of the product is determined.

[0015] The training corpus classification step specifically includes:

[0016] Determine review group information based on product review information obtained from the product review area;

[0017] The product review information and the review group information are used as the digital human generated corpus.

[0018] Furthermore, the training corpus classification step specifically includes:

[0019] Obtain first product details information from the first product details area;

[0020] Obtaining first product review information from a product review area corresponding to the first product;

[0021] After associating the first product detail information with the first product review information, the information is used as the model training corpus.

[0022] The training corpus collection step specifically includes:

[0023] For each product, determine the validity of each product review information in its product review area;

[0024] When a product review is valid, collect the product review as training corpus;

[0025] Among them, a product review information effectively includes:

[0026] There are a predetermined number of other users who give positive feedback to the product review information.

[0027] Furthermore, the training corpus collection step also includes:

[0028] Use web crawlers to crawl product review information and product detail information from the product detail area and product review area;

[0029] For each comment, obtain the user registration information corresponding to the comment by simulating clicks;

[0030] For each product detail information, multiple description keywords corresponding to the product are obtained through semantic analysis;

[0031] The product review information, user registration information and description keywords are used as training corpus. The model training steps specifically include:

[0032] Dividing the model training corpus into training samples and test samples;

[0033] After training the benchmark product classification model using the training samples, testing the trained benchmark product classification model using the test samples;

[0034] When the test result meets the predetermined requirement, the trained reference product classification model is output as the target product classification model.

[0035] Based on the target product classification model obtained by the model training method described in the first aspect, in the second aspect of the present invention, a product classification model prediction method based on digital human is proposed. The method uses the target product classification model to perform the following steps:

[0036] S710: receiving information of products to be classified input by a user;

[0037] S720: parsing the information of the product to be classified to obtain a plurality of semantic segmentations;

[0038] S730: Using the semantic segmentation as input of the target product classification model;

[0039] S740: Outputting a product classification result through a digital human model associated with the target product classification model.

[0040] The product classification results include product information in multiple dimensions, including one or a combination of the following:

[0041] The target population of the product, the product price range, the expected effects of the product, the product contraindications, the product logistics situation, and the product category label.

[0042] After obtaining the output result of the target product classification model, the method further includes:

[0043] Based on target classification requirements, select one or more dimensional information from the multiple dimensional information;

[0044] Based on the selected one or more dimensional information, a target classification of the product is determined.

[0045] The information of the product to be classified input by the user in step S710 includes one or a combination of the following:

[0046] Enter product pictures, product text descriptions, product video keyframes, and product voice descriptions.

[0047] The aforementioned product classification model training method and prediction method based on digital humans can be automatically implemented through various forms of electronic devices through computer program instructions; the computer program instructions can be stored in different forms of storage media and loaded into computer electronic devices for execution.

[0048] Therefore, in the third aspect of the present invention, a computer-readable storage medium is also provided for storing computer instructions. When the computer instructions are executed on an electronic device, the electronic device executes all or part of the steps of the aforementioned product classification model training method and prediction method based on digital humans.

[0049] In a fourth aspect of the present invention, a computer device is also proposed, comprising a processor and a memory, wherein the memory is used to store instructions, and the processor is used to call the instructions in the memory, so that the computer device executes the aforementioned product classification model training method and prediction method based on digital humans.

[0050] In a fifth aspect of the present invention, a computer program product is also proposed, the product comprising a computer program, and when the computer program is executed, the steps of the aforementioned product classification model training method and prediction method based on digital humans are implemented.

[0051] The digital human-based product classification model training and prediction method proposed in the present invention first obtains the target product classification model and the digital human model associated with it through a training corpus collection step, a training corpus classification step, a digital human generation step, a model training step and a digital human model association step, then parses the information of the product to be classified, obtains multiple semantic segmentations as input of the target product classification model, and outputs the product classification result through the digital human model associated with the target product classification model, so that the output of the product classification result is richer and more intuitive, and there is also a strong correlation between the generated digital human model and the product classification model, and the user experience is good.

[0052] Further advantages of the present invention will be further reflected in detail in the specific embodiments section in conjunction with the drawings of the specification. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0054] Figure 1 This is a schematic diagram of the main flow of a product classification model training method based on digital human according to an embodiment of the present invention.

[0055] Figure 2 yes Figure 1 Schematic diagram of the scene page for corpus collection in the method

[0056] Figure 3 is based on Figure 1 Schematic diagram of the method realizing the digital human output of product classification results

[0057] Figure 4 This is a schematic diagram of the main flow of a product classification model prediction method based on digital human in one embodiment of the present invention. DETAILED DESCRIPTION

[0058] First of all, it should be pointed out that the embodiments of the product classification model training method and prediction method based on digital humans mentioned in this section can be implemented through computer programs on electronic devices and systems configured with memory and processors. The electronic devices and systems can be in the form of physical machines, virtual machines, servers, clusters or any combination thereof.

[0059] Preferably, the specific form of the electronic device may also be a human-computer interaction terminal, and the human-computer interaction terminal may be a desktop terminal, a smart handheld terminal, a mobile terminal, etc. with a human-computer interaction interface.

[0060] In addition, the object of the method execution of the present invention is product classification. For the method flow executed by the computer, product classification and commodity classification are generally not distinguished. Therefore, in the embodiments, drawings and other introductions of the present invention, the technical features "commodity" and "product" are common terms, but in different contexts, commodity or product may be used respectively for better reading and understanding.

[0061] See first Figure 1 , Figure 1 It is a main flow diagram of a product classification model training method based on digital human according to an embodiment of the present invention.

[0062] Figure 1 The main process of the method shown includes a training corpus collection step, a training corpus classification step, a digital human generation step, a model training step, and a digital human model association step. Each step is specifically implemented as follows:

[0063] Training corpus collection steps: Collect training corpus from product details and product reviews;

[0064] Training corpus classification step: dividing the training corpus into digital human generated corpus and model training corpus;

[0065] Digital human generation step: generating a digital human model based on the digital human generation corpus;

[0066] Model training step: training a benchmark product classification model based on the model training corpus to obtain a target product classification model;

[0067] Digital human model association step: Associating the digital human model with the target product classification model so that the output result of the target product classification model is output by the digital human model.

[0068] Next, we will Figure 1 Based on the combination Figure 2 , and further introduces each step of the method.

[0069] The first step is to collect training corpus, which mainly collects training corpus from the product details area and product review area.

[0070] In an example, the corpus collection scenario is as follows Figure 2 As stated, Figure 2 yes Figure 1 Schematic diagram of the scenario page for corpus collection in the method.

[0071] Figure 2 The schematic diagram shows that the corpus collection scene page of a certain product includes a product details page area and a product review area.

[0072] Generally speaking, the product details page area usually introduces the basic attribute information of the product in detail ( Figure 2 The product reviews area displays the product reviews corresponding to the product's historical total sales records, including good reviews, medium reviews, and bad reviews.

[0073] Figure 2 The "Comment Area" shows the number of positive, neutral and negative reviews, and schematically shows a product review by "User ID 1" "The product is very useful...";

[0074] Corresponding to the comment, the comment area page also sets a browsing feedback button for other users to collect other users' feedback on the comment, which includes positive feedback (like / useful) and negative feedback (deprecation / useless); Figure 2 The positive feedback (likes) shown are represented by a right triangle, indicating that the number of positive feedback (likes / useful) is 199, and the number of negative feedback (disparagement / uselessness) is 25. Obviously, the number of positive feedback (likes / useful) is much greater than the number of negative feedback (disparagement / uselessness), so the comment can be considered a valid comment.

[0075] Comparison Figure 2 , the training corpus collection step specifically includes:

[0076] For each product, determine the validity of each product review information in its product review area;

[0077] When a product review is valid, collect the product review as training corpus;

[0078] Among them, a product review information effectively includes:

[0079] There are a predetermined number of other users who give positive feedback to the product review information.

[0080] Preferably, a product review information is valid when it meets one of the following conditions:

[0081] (1) More than a predetermined number of other users give positive feedback to the product review information;

[0082] (2) The word count of the product review exceeds the preset average;

[0083] (3) The product review information is a top review;

[0084] (4) The number of times the product review information is cited or replied to by others exceeds the preset value;

[0085] Understandably, in order to encourage users to post valuable comments, many online shopping websites usually automatically detect the semantics of user comments when they post them, including whether there are useless words, the number of words in the comment, whether the comment contains keywords, etc. At the same time, they also judge the value of the comment based on the number of times the user comment is cited and replied by others, and then display the valuable comment at the top, etc. Such valuable comments usually provide more real and valuable training corpus.

[0086] Therefore, for Figure 2 For the corpus collection scenario page, the collected training corpus includes at least two areas: product detail information obtained from the product detail area and product review information obtained from the product review area corresponding to the product.

[0087] For the subsequent training process, different training processes require different training corpus samples. Therefore, the corpus needs to be classified, that is, the training corpus classification step: the training corpus is divided into digital human generated corpus and model training corpus;

[0088] Specifically, the training corpus classification step includes:

[0089] Determine review group information based on product review information obtained from the product review area;

[0090] The product review information and the review group information are used as the digital human generated corpus.

[0091] by Figure 2 For example, based on Figure 2 The product review information shown identifies the reviewer as "User ID1". For this review information, the user registration information corresponding to the review is obtained by simulating a click. That is, after simulating a click on "User ID1", the user registration information corresponding to User ID1 can be obtained, which generally includes the user's gender, age, registration time, and may also include the user's historical shopping records, etc. This information has been anonymized and will not expose user privacy or be used for user identity identification, but it can provide anthropomorphic shopping scene corpus for subsequent digital human generation.

[0092] Therefore, the comment group information can include the above-mentioned user registration information, combined with the product review information, which can be used as the subsequent digital human generation corpus.

[0093] Next, the digital human generation step is entered to generate a digital human model based on the digital human generation corpus.

[0094] The "digital human" mentioned in this embodiment refers to a virtual human that imitates humans in appearance, behavior, thinking and emotion by relying on artificial intelligence technology. It has three main characteristics: (1) similar appearance, gender, personality and other characteristics to humans; (2) can convey human behavior through language, expression and action; (3) has the ability to perceive and recognize the external world and communicate and interact with people. At present, the classification of digital humans in the market mainly includes identity-based digital humans and service-based digital humans. Identity-based digital humans mainly include two types: virtual IP idols and human virtual avatars, which pay more attention to creativity and beauty and are widely used in the entertainment industry; service-based digital humans have certain functionality and interactivity, and can replace real people to provide users with services in various application scenarios, with a high degree of automation, such as virtual anchors, virtual customer service, virtual assistants, etc.

[0095] Based on artificial intelligence technology, the customer characteristics are portrayed by analyzing the collected product review information and the information of the reviewers as data, and then personalized virtual digital human appearance, sound effects and services are generated based on the customer characteristics. The appearance and interactive sound are both personalized modes, which can generate digital human images more accurately.

[0096] For related processing technologies of evaluation texts, please refer to the following prior arts:

[0097] Yang Senqi. Research on rural tourism evaluation text classification based on graph convolutional neural network and ERN I E-gram[D]. Sichuan Agricultural University, 2023.

[0098] Xia Huosong, Zhu Huiyi, Wei Fengrui. Research on sentiment sub-classification model of subjective product reviews, Intelligence Magazine, Vol. 32, No. 2, 2013.

[0099] In another aspect, it is also necessary to perform a model training step: training a benchmark product classification model based on the model training corpus to obtain a target product classification model.

[0100] Specifically, the model training step specifically includes:

[0101] Dividing the model training corpus into training samples and test samples;

[0102] After training the benchmark product classification model using the training samples, testing the trained benchmark product classification model using the test samples;

[0103] When the test result meets the predetermined requirement, the trained reference product classification model is output as the target product classification model.

[0104] It can be seen that when generating a digital human model, the focus is on customer portrait features, including the collected customer review information on the product and the population information of the reviewer itself; when training a classification model, it is necessary to combine product details information and product review information, because the purpose of the classification model is that after the user inputs a product information to be classified in the future, the classification model can output the corresponding classification results, and these classification results are often included in the valid review information of the product.

[0105] It can be seen that the training corpus of the classification model first has a certain preprocessing process, which specifically includes: obtaining the first product detail information obtained from the first product detail area;

[0106] Obtaining first product review information from a product review area corresponding to the first product;

[0107] After associating the first product detail information with the first product review information, the information is used as the model training corpus.

[0108] It can be understood that the benchmark product classification model is a type of machine learning, and its task is to predict the category of a sample by learning the characteristics of the sample. Training based on the benchmark product classification model means learning how to determine which category a sample belongs to from the labeled training data, and then making predictions based on the learned knowledge when faced with new data.

[0109] The benchmark product classification model can be selected from the prior art. As an example, a common benchmark product classification model feature extractor, feature conversion module and feature classifier, etc., are used, and the present invention will not elaborate on this.

[0110] The principle of model training based on the benchmark product classification model is similar to the process of other model training in the prior art, including the number of training times / termination conditions, distinction between training samples / test samples, etc.

[0111] Specifically, after the benchmark product classification model is trained using the training samples, the trained benchmark product classification model is tested using test samples; when the test results meet predetermined requirements, the trained benchmark product classification model is output as the target product classification model.

[0112] The test results here meet the predetermined requirements, which can be that multiple result outputs of the trained benchmark product classification model all meet user expectations; it can also be that the difference in the output results of the trained benchmark product classification model for the same or similar inputs is less than a predetermined value, etc.

[0113] However, unlike the prior art, the present invention is further improved in that it also includes a digital human model association step: associating the digital human model with the target product classification model so that the output result of the target product classification model is output by the digital human model.

[0114] Since the digital human model is generated based on product review information and reviewer information, and the target product classification model uses a corpus sample formed by associating product detail information with the product review information during training, this means that when each target product classification model outputs product classification results for different products to be classified, there is a digital human model associated with the product that can be used to output the corresponding classification results.

[0115] Preferably, a digital human model associated with the product outputs the corresponding classification result, which can be a digital human voice broadcast of the result. At the same time, the digital human has the appearance, sound effects and services corresponding to the product, and can provide richer visual, auditory, text and graphic information. Based on this, users can receive relevant information more intuitively, thereby improving user experience.

[0116] In order to better reflect the process, Figure 1-Figure 2 Based on Figure 3 ,based on Figure 1 The method realizes a schematic diagram of a digital human outputting a product classification result.

[0117] Figure 3 In the example, the target product classification model has been trained and can be loaded to the user terminal for use. At this time, the user enters the information of the product to be classified by inputting a product picture, a product text description, a product video key frame, or a product voice description.

[0118] It can be understood that the information of the product to be classified input by the user here is vague and rough. For example, the user only knows the product name, or only sees a product picture, or sees an advertisement clip containing the product.

[0119] The target product classification model extracts the corresponding features to be classified through the feature extractor, and converts them into feature vectors supported by the model through the feature conversion module. After sending them to the feature classifier, the classification results can be analyzed.

[0120] As an advantage, different from the single classification result of the prior art, the classification result here is actually based on the vague and rough information of part of the product to be classified input by the aforementioned user. Through the target product classification model, based on the network shared data used to train the target product classification model (the corpus sample formed by the association of product detail information and the product review information), more detailed information corresponding to the product can be output. Each of these detailed information represents a possible subsequent classification dimension. Which classification dimension to adopt will be selected by the user according to the subsequent specific purpose (i.e., the target classification requirement).

[0121] exist Figure 3 In the example, the product classification result includes one or a combination of the following:

[0122] The target population of the product, the product price range, the expected effects of the product, the product contraindications, the product logistics situation, and the product category label.

[0123] It can be seen that the product classification results include information of multiple dimensions.

[0124] For example, if the purpose of product classification by the user (i.e., target classification demand) is to determine the advertising target of the product, i.e., the target population, then the dimension information that the user is concerned about may be "product applicable population". Based on this dimension information, the user can obtain the specific population classification (target classification) of the product, such as middle school students, the elderly, and other classified populations;

[0125] As another example, if the user's purpose of product classification (i.e., target classification requirement) is to determine the product's price category (high-end / low-end), then the dimensional information that the user is concerned about may be "product price range" + "product category label", where "product price range" provides the current price range of the product or similar products (competing products); and "product category label" provides more auxiliary information, for example, the product category label may be "suitable for white-collar workers", "suitable for home", "suitable for office", etc.; based on the combination of the two dimensional information, the user can determine the product's price category (target classification) as high-end (for example, high-end pricing only suitable for office white-collar workers) or low-end (ordinary household daily necessities).

[0126] Similarly, users can choose to use information in other dimensions, such as product contraindications and product logistics conditions, based on actual target classification needs, to facilitate the final target classification needs of the product.

[0127] Therefore, it is preferred that although Figure 1 Although not shown, after obtaining the multiple dimension information of the product classification result, the method further includes:

[0128] Product target classification determination step: based on the target classification requirements, one or more dimensional information is selected from the multiple dimensional information, and based on the selected one or more dimensional information, the target classification of the product is determined.

[0129] As a further advantage, unlike the prior art, Figure 3 A digital human model is also shown. The digital human model is associated with the product, and the result is announced through the digital human voice. At the same time, the digital human has the appearance, sound effects and services corresponding to the product, and can provide richer visual, auditory, text and graphic information. Based on this, users can receive relevant information more intuitively, thereby improving user experience.

[0130] As mentioned above, since the digital human model is generated based on product review information and review crowd information, and the target product classification model uses a corpus sample formed by associating product detail information with the product review information during training, this means that when each target product classification model outputs a product classification result for different commodities to be classified, there is a digital human model associated with the commodity that can be used to output the corresponding classification result. Therefore, the above embodiment does not simply introduce a voice broadcast intermediate to mechanically broadcast the output result, but fully considers the specific attributes of the commodity classification information to be output to select a targeted digital human model generated based on product review information and review crowd information, so that the demonstration of the entire broadcast process is more friendly and shows a natural process, which will not make the user feel abrupt.

[0131] Figure 3 The example actually uses Figure 1 The method is a model prediction method after model training. Therefore, Figure 4 A schematic diagram of the main flow of a product classification model prediction method based on digital human according to an embodiment of the present invention is further provided.

[0132] exist Figure 4 In the method, the method includes steps S710-S740, and each step is specifically implemented as follows:

[0133] S710: receiving information of products to be classified input by a user;

[0134] S720: parsing the information of the product to be classified to obtain a plurality of semantic segmentations;

[0135] S730: Using the semantic segmentation as input of the target product classification model;

[0136] S740: Outputting a product classification result through a digital human model associated with the target product classification model.

[0137] The product classification results include product information in multiple dimensions, including one or a combination of the following:

[0138] The target population of the product, the product price range, the expected effects of the product, the product contraindications, the product logistics situation, and the product category label.

[0139] Preferably, although Figure 4 Not shown, but after obtaining the output result of the target product classification model, the method further includes:

[0140] S750: Select one or more dimensional information from the multiple dimensional information based on target classification requirements;

[0141] Based on the selected one or more dimensional information, a target classification of the product is determined.

[0142] The information of the product to be classified input by the user in step S710 includes the following product fuzzy information or a combination thereof:

[0143] Enter product pictures, product text descriptions, product video keyframes, and product voice descriptions.

[0144] When the method is implemented using a computer program, the training corpus collection step further includes:

[0145] Use web crawlers to crawl product review information and product detail information from the product detail area and product review area;

[0146] For each comment, obtain the user registration information corresponding to the comment by simulating clicks;

[0147] For each product detail information, multiple description keywords corresponding to the product are obtained through semantic analysis;

[0148] The product review information, user registration information and description keywords are used as training corpus.

[0149] Similar corpus crawling techniques can be found in:

[0150] Wang Jinghua. Research on flour raw material classification and price index prediction based on e-commerce platform crawler data[D]. Henan University of Technology, 2024.

[0151] Other technologies, principles, algorithms or models not elaborated in detail in this application can be found in the prior art, including the prior art documents cited in the embodiments of this application, which constitute an integral part of the technical solution of the present invention.

[0152] However, it should be pointed out that the present invention provides some specific prior art documents for the specific implementation of some steps, but for those skilled in the art, it does not mean that the implementation of the relevant parts can only adopt the relevant technical means recorded in these listed prior art documents. Those skilled in the art can know that the implementation means of the relevant parts are not unique, and the documents provided in the embodiment part are only one of the ways or reference means that can be implemented. Of course, those skilled in the art can also use other prior art known in the art (for example, for web page data crawling, there are other types of prior art documents in this field) to implement the relevant steps or methods. The listed documents are only for reference and are not used as exhaustive examples or limitations on the embodiments of the present invention. The scope of protection of this application shall be subject to the scope recorded in the claims.

[0153] At the same time, in the specific implementation of this application, if user-related data is involved, when the embodiment of this application is applied to a specific product or technology, the user's permission or consent must be obtained, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0154] The foregoing has shown and described the method embodiments and system of the present invention, but it is understandable to those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A product classification model training method based on digital human, characterized in that: The training method comprises: Training corpus collection steps: Collect training corpus from product details and product reviews; Training corpus classification step: dividing the training corpus into digital human generated corpus and model training corpus; Digital human generation step: generating a digital human model based on the digital human generation corpus; Model training step: training a benchmark product classification model based on the model training corpus to obtain a target product classification model; Digital human model association step: associating the digital human model with the target product classification model so that the output result of the target product classification model is output by the digital human model; The output result of the target product classification model includes product information of multiple different dimensions.

2. A product classification model training method based on digital human as claimed in claim 1, characterized in that: The training corpus classification step specifically includes: Determine review group information based on product review information obtained from the product review area; The product review information and the review group information are used as the digital human generated corpus.

3. A product classification model training method based on digital human as claimed in claim 1, characterized in that: The training corpus collection step specifically includes: For each product, determine the validity of each product review information in its product review area; When a product review is valid, collect the product review as training corpus; Among them, a product review information effectively includes: There are a predetermined number of other users who give positive feedback to the product review information.

4. The product classification model training method based on digital human as claimed in claim 1, characterized in that: The model training steps specifically include: Dividing the model training corpus into training samples and test samples; After training the benchmark product classification model using the training samples, testing the trained benchmark product classification model using the test samples; When the test result meets the predetermined requirement, the trained reference product classification model is output as the target product classification model.

5. The product classification model training method based on digital human as claimed in claim 1, characterized in that: The training corpus classification step specifically includes: Obtain first product details information from the first product details area; Obtaining first product review information from a product review area corresponding to the first product; After associating the first product detail information with the first product review information, the information is used as the model training corpus.

6. A product classification model training method based on digital human as claimed in claim 1, characterized in that: The training corpus collection step also includes: Use web crawlers to crawl product review information and product detail information from the product detail area and product review area; For each comment, obtain the user registration information corresponding to the comment by simulating clicks; For each product detail information, multiple description keywords corresponding to the product are obtained through semantic analysis; The product review information, user registration information and description keywords are used as training corpus.

7. A product classification model prediction method based on digital human, wherein the product classification model is a target product classification model trained by a product classification model training method based on digital human according to any one of claims 1 to 6, characterized in that: The prediction method comprises the following steps: S710: receiving information of products to be classified input by a user; S720: parsing the information of the product to be classified to obtain a plurality of semantic segmentations; S730: Using the semantic segmentation as input of the target product classification model; S740: Outputting a product classification result through a digital human model associated with the target product classification model.

8. The product classification model prediction method based on digital human as claimed in claim 7, characterized in that: The product classification results include one or a combination of the following: The target population of the product, the product price range, the expected effects of the product, the product contraindications, the product logistics situation, and the product category label.

9. The product classification model prediction method based on digital human as claimed in claim 7, characterized in that: The information of the product to be classified input by the user in step S710 includes one or a combination of the following: Enter product pictures, product text descriptions, product video keyframes, and product voice descriptions.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, implement the steps of a product classification model training method based on a digital human as described in any one of claims 1 to 6; or, when executed by a processor, implement the steps of a product classification model prediction method based on a digital human as described in any one of claims 7 to 9.

Images