Commodity category determination method and system based on AI model

Abstract

The invention discloses a commodity category determination method and system based on an AI model, and the method comprises the steps: executing a supplement operation through a first AI model when the judgment result shows that there is no commodity category in the commodity detail information of a commodity, generating an original commodity category, constructing a category database according to the commodity category of a target platform, and determining the commodity category of the target platform according to the category database. And performing similarity retrieval through the second AI model to obtain a to-be-selected commodity category set, performing structured processing on the original commodity category and the to-be-selected commodity category set, inputting a plurality of third AI models to obtain a plurality of third output results, performing step-by-step judgment operation according to the third output results, and determining a target commodity category. According to the method and the corresponding system, a commodity category library is prevented from being locally created in advance in the e-commerce ERP, and the commodity category matching and commodity publishing efficiency is greatly improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and particularly to a method and system for determining product categories based on an AI model. Background Art

[0002] With the development of the cross-border e-commerce industry, the sales methods of many sellers are gradually changing from a single-platform selling model to a multi-platform selling model. When publishing products on multiple platforms through an e-commerce ERP, product information is usually collected from the procurement platform, i.e., the original platform. After storing the product information in the e-commerce ERP, it is then published to the corresponding e-commerce platform, i.e., the target platform.

[0003] Since there are significant differences in the formats and standards of product information among various procurement platforms and e-commerce platforms, in order to achieve normal matching of product categories between the procurement platform and the e-commerce platform, it is usually necessary to pre-create a local product category library on the e-commerce ERP that can cover the product categories of each procurement platform, and identify the corresponding relationships between each local product category and the product categories of each e-commerce platform. Additionally, as the product categories of the procurement platform and the e-commerce platform are updated, the data in the local product category library needs to be continuously maintained and updated. Thus, a large amount of manpower is required to create and maintain the product category library. Otherwise, incorrect product categories may lead to errors in product publishing.

[0004] Other technical problems related to this application will be further elaborated later. The above content is only used to assist in understanding the technical solution of this application and does not mean that all of the above content is prior art. Summary of the Invention

[0005] The main objective of this application is to provide a method and system for determining product categories based on an AI model, which can avoid pre-creating a product category library locally on the e-commerce ERP. When publishing products, detailed and accurate product categories can be configured for different e-commerce platforms respectively, significantly improving the efficiency of product category matching and product publishing.

[0006] To achieve the above objective, this application proposes a method for determining product categories based on an AI model for an e-commerce ERP system. The method includes the following steps: Step S1: Determine whether there is product category information in the product detail information of the product. If there is, use the corresponding product category as the original product category; If not, perform a supplement operation through the first AI model, where the supplement operation generates the original product category based on the product detail information; Step S2: Construct a category database based on the product categories of the target platform, and perform a similarity search on the original product category through the second AI model to extract a set of candidate product categories from the category database; Step S3: Structurally process the original product category and the set of candidate product categories, and input the results of the structural processing into multiple third AI models to obtain multiple third output results. The structural processing is to process the original product category and the set of candidate product categories into query instructions that can be directly input into the third AI models; Step S4: Perform a step-by-step judgment operation on the third output results to obtain the target product category of the product on the target platform. The step-by-step judgment operation includes: Step S41: Execute the first judgment logic. The first judgment logic includes determining whether all the third output results are the same. If so, use the same candidate product category as the target product category. If not, execute Step S42; Step S42: Execute the second judgment logic. The second judgment logic is to determine whether there are any identical third output results. If so, use the candidate product category with the majority of identical results as the target product category. If not, execute Step S43; Step S43: Execute the third judgment logic. The third judgment logic is to determine whether all the third output results are different. If so, use the candidate product category output by a trusted third AI model as the target product category.

[0007] Other features and technical effects of this application are described and explained in the later part of the specification. The technical problem-solving ideas and related product design solutions of this application are as follows: When a product is cross-platform published, the e-commerce ERP creates a local product category library to identify the corresponding relationships of product categories on different platforms, but it requires a lot of effort to maintain the local product category library. The applicant found that if the product category required for product publication can be directly generated through an AI model (i.e., an AI system), then the pre-creation of a local product category library in the e-commerce ERP can be avoided, and the time for continuously maintaining the product category library can be reduced.

[0008] After actual verification, it is found that although many current AI models (generative large models) are powerful and have great advantages in semantic understanding and text processing, due to the large amount of product detail information of products, using part or all of the product detail information as the corpus basis, the product categories generated after calling the AI model vary greatly, and the generated product categories cannot accurately match the product categories on the e-commerce platform.

[0009] On this basis, the applicant also found that although the product categories generated by the AI model (the first AI model) cannot match the product categories of the e-commerce platform to be released, there is a certain similarity between the generated product categories and the product categories that should actually be matched on the e-commerce platform to be released (the target product categories). Further call the AI model (the second AI model) for similarity retrieval, and retrieve the N candidate product categories closest to the generated product categories from each product category of the e-commerce platform, and form a candidate product category set with the N candidate product categories. The applicant found that when N≥5, in most scenarios, the target product category will necessarily be included in the candidate product category set.

[0010] At this time, the AI model (the third AI model) can be called again to screen out the target product category from the candidate product category set. After verification, it is found that if only relying on the third AI model for semantic understanding, the target product category cannot be completely and accurately screened out, and there will be a relatively large deviation. However, if multiple third AI models (more than 3) are called for semantic screening and multiple temporary product categories (the third output results) are output respectively, then in most scenarios, the same temporary product category among them is the target product category. In addition, after verification, when N≤10, the probability of the same temporary product category appearing can be better improved, and thus the target product category can be obtained faster.

[0011] This method for determining product categories can avoid pre-creating a complex product category library locally in the e-commerce ERP, and also avoid the cumbersome continuous maintenance and upgrade of the product category library. Moreover, when releasing products, the product information collected from any procurement platform can be randomly released to multiple e-commerce platforms. Even if the original product category is missing in the collected product information, the accurate product category of the e-commerce platform can be matched, thus greatly improving the efficiency of product category matching and product release.

[0012] Other implementation schemes and technical effects are described later.

[0013] Furthermore, this application also includes systems corresponding to various methods. The system includes the functional modules involved in this application, executes the operation instructions of the corresponding functional modules or corresponding methods, and outputs relevant data information to the front-end interface of the system. The system is stored in a server and / or computer device including a processor, and the processor is used to execute the operation instructions of the system.

[0014] Declaration: The functional modules of this application can be integrated with each other, can also exist independently, or one functional module can be a sub-module of another functional module; the step numbers such as S1 and S2 do not limit the sequence of the corresponding operation steps. Brief Description of the Drawings

[0015] The accompanying drawings are used to provide a further understanding of the present application and do not constitute a limitation on the present application; the content shown in the accompanying drawings may be the actual data of the embodiments and falls within the protection scope of the present application.

[0016] Figure 1 It is a schematic diagram of the function modules of an e-commerce ERP system in an embodiment of the present application; Figure 2 It is a schematic flowchart of a method for determining product categories based on an AI model in an embodiment of the present application; Figure 3 It is a schematic diagram of model interaction of a method for determining product categories based on an AI model in an embodiment of the present application; Figure 4 It is a schematic diagram of the principle of a method for determining product categories based on an AI model in an embodiment of the present application. Detailed implementation manners

[0017] In order to make the objectives, technical solutions and advantages of the present application clearer and more understandable, the embodiments of the present application will be further described in detail below in conjunction with the accompanying drawings through specific implementation manners. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0018] Refer to Figure 2 The present application proposes a method for determining product categories based on an AI model, which is used in an e-commerce ERP system or an e-commerce platform system. In an embodiment of the present application, the method includes steps S1 to S4 as follows.

[0019] Step S1: Determine whether there is product category information in the product detail information of the product. If so, use the corresponding product category as the original product category; if not, perform a supplementary operation through the first AI model, and the supplementary operation generates the original product category based on the product detail information.

[0020] Refer to Figure 3, this solution uses multiple AI models with different functions. The first AI model is used to perform product category matching on the input product detail information or the field information disassembled from the product detail information when the product category does not exist in the product detail information of the product, generate the product category of the product on the original platform, or the local product category of the product on the e-commerce ERP, or the product category that best matches the product detail information, and output the above product category as the original product category; the second AI model is used to calculate the vector similarity between the vectorized original product category and the category database, and select the N candidate product categories most similar to the original product category from the product categories of the target platform according to the similarity size, and output the product category set, where 5 ≤ N ≤ 10; the third AI model is used to calculate the similarity between the input structured original product category and the set of candidate product categories, and output one candidate product category with the highest similarity to the original product category in the set of candidate product categories. The number of the third AI models is at least 3. The first AI model, the second AI model, and the third AI model may not be the same AI model.

[0021] Obtain the product detail information on the original platform through the function interface of the original platform and split it according to the data type of the product detail information. Specifically, when the data type is structured data or semi-structured data, split it by extraction; when the data type is unstructured data, split it by semantic analysis method or image extraction and recognition method. After splitting, the field information of the product is obtained, and the field information includes at least one of the product title, product description, product picture, product category, and other information. Determine whether the product category field in the field information is empty. If not, extract the field value of the product category field to obtain the original product category of the product. If so, confirm that there is no product category information in the product detail information of the product.

[0022] When there is no product category information, perform a supplementary operation through the first AI model. Specifically, determine the field characteristics of the product detail information, where the field characteristics include text characteristics and visual feature statistical characteristics, and the visual features include the product detail information in the form of pictures. Analyze each field characteristic of the product detail information based on text analysis and / or image recognition, and match the product category to obtain the original product category. In addition, the product category matching of the first AI model can be a rule-based method, such as keyword matching, decision tree, etc., or a machine learning method, such as a text classification model, or a hybrid method combining rules and machine learning.

[0023] Step S2: Construct a category database according to the product categories of the target platform, and perform a similarity search on the original product category through the second AI model to extract a set of candidate product categories from the category database.

[0024] Reference Figure 4, the embedding model is used to compress each high-dimensional sparse product category into a low-dimensional dense vector. Specifically, after vectorizing the product categories in the target platform, a category database is obtained. After vectorizing the original product categories, the original product vector data is obtained, and the mapping relationship between each product category and the vector is saved. The pre-trained language model can also be used to encode the text segment into a vector to achieve vectorization processing.

[0025] The vector similarity is calculated by the second AI model, and the N candidate vector data most similar to the vector data is extracted from the category database. Based on the mapping relationship, the product categories corresponding to the candidate vector data are obtained, and a set of candidate product categories is obtained. This solution further proposes to calculate the vector similarity by means of cosine similarity, Euclidean distance, Manhattan distance, dot product, etc. For example, when the second AI model calculates the vector similarity by cosine similarity, the calculation formula is,

[0026] The range is [-1, 1], and the larger the value, the more similar; when the second AI model calculates the vector similarity by Euclidean distance, the calculation formula is,

[0027] The range is [0, +∞), and the smaller the value, the more similar.

[0028] In summary, after the vectorization operation and then calculating the vector similarity, the original word order structure of the product category can be broken, so that the semantic expression is not limited by the word order structure. Compared with directly calculating the semantic similarity of the product category, it can accurately distinguish and identify the text similarity and the content with inconsistent expressed meanings, and obtain an accurate set of candidate product categories.

[0029] Furthermore, this solution proposes to perform a threshold judgment on the similarity corresponding to the candidate product categories calculated by the second AI model to reversely verify the accuracy of the first AI model in matching the original product categories. Specifically, monitor the similarity corresponding to each candidate product category. When the similarity retrieved by the similarity is lower than the preset value, propose to calibrate the original product category based on the dynamic threshold and the confidence probability value. After verification, when the value range of the preset value is [0.71, 0.9], it can better screen the candidate product categories and verify the matching accuracy of the second AI model.

[0030] The confidence probability value represents the probability that the product output by the first AI model belongs to the original product category. For example, if the model predicts that the probability of a product belonging to "mobile phone accessories" is 0.92, then the confidence probability value is 92%. In practical applications, the first AI model has the situation of less training data for unpopular product categories and overfitting for popular product categories, which may lead to a low or high confidence probability value. In response to this, the applicant proposes to calibrate the original product category according to a dynamic threshold. The dynamic threshold is determined by a basic threshold and a popularity index. The popularity index is the logarithm of the number of products published in the original product category within a preset time (the past 30 days), that is, popularity index = log(number of products in the product category in the past 30 days + 1). At this time, the calculation formula of the dynamic threshold is,

[0031] Determine whether the confidence probability value is lower than the dynamic threshold. If it is lower, trigger the category adjustment logic. According to the product category constraint rules, determine whether there is a conflict between the original product category and the product description of the product. Filter the product category after matching for compliance on the e-commerce platform, and associate sub-product categories according to the product description or / and the original product category. Confirm the sub-product category according to the hierarchical relationship between the historical product category and the sub-product category, and use the sub-product category as the original product category, where the historical product category is the product category to which the product belongs in the historical product release data. The product category constraint rules are the requirements for classifying products into each product category, that is, the conditions that each product must meet when being classified into each product category, including one or more of the physical attribute requirements, functional use requirements, usage scenario requirements, target population requirements, brand and model requirements, regulations and industry standards, and platform special rules of the product.

[0032] In an embodiment, when the product details information of a product records "degradable mobile phone case", the product description is "material = corn starch-based plastic", and the output of the first AI model is "mobile phone accessories > protective case", at this time the confidence probability value is 68%, which is lower than the dynamic threshold of 75% for the mobile phone accessories category, triggering the category adjustment logic. At this time, the original product category "mobile phone accessories > protective case" conflicts with the product description "corn starch-based plastic". According to "corn starch-based plastic", it is associated with the "environmental protection materials" sub-category. Query the seller's product release history and determine that the accuracy rate of the seller in the "environmental protection products" product category is 92%. Thus, the original product category is adjusted to obtain the original product category as "mobile phone accessories > environmental protection materials zone > degradable protective case".

[0033] In summary, by combining the confidence probability value with the dynamic threshold mechanism, the problems of excessive intervention in high-frequency product categories and insufficient coverage of long-tail product categories when the first AI model matches product categories are solved, the stability and credibility of the model are enhanced, computing resources are saved, and the accuracy of product category matching is improved.

[0034] Step S3: Structurally process the original product category and the set of candidate product categories, and input the structural processing result into multiple third AI models to obtain multiple third output results.

[0035] The structural processing is to process the original product category and the set of candidate product categories into query instructions that can be directly input into the third AI model. Specifically, using the original product category and the set of candidate product categories as input variables, a structural model is constructed based on each third AI model, and the structural processing result is obtained based on the output result of the structural model. The construction of the structural model includes determining the corresponding query instruction template based on the retrieval mode of each third AI model, and filling the original product category and the set of candidate product categories into the query instruction template. The structural processing result is the query instruction of each third AI model, and the role of the structural model is to combine the query instruction templates of each third AI model with the structured corpus.

[0036] When the retrieval mode of the third AI model is keyword- or semantic similarity-based retrieval, that is, the text retrieval mode, the query instruction template is a natural language template, and the input variables, that is, the original product category and the set of candidate product categories, are converted into text combination instructions; when the retrieval mode is combined text, image and other data retrieval, that is, the multi-modal retrieval mode, the query instruction template is a multi-modal template, and the input variables are converted into a combination of image feature vectors and text descriptions; when the retrieval mode is database field- or knowledge graph-based retrieval, that is, the structured retrieval mode, the query instruction template is a structured query template, and the input variables are converted into field formats.

[0037] The third AI model calculates the similarity between the original product category and each candidate product category in the set of candidate product categories, sorts them according to the similarity level, and selects the candidate product category with the highest similarity as the third output result of the third AI model.

[0038] Step S4: Perform a step-by-step judgment operation on the third output result to obtain the target product category of the product on the target platform. The step-by-step judgment operation includes: Step S41: Execute the first judgment logic. The first judgment logic includes judging whether all the third output results are the same. If so, use the same candidate product category as the target product category. If not, execute step S42; Step S42: Execute the second judgment logic. The second judgment logic is to judge whether there are any same third output results. If so, use the candidate product category with the majority of the same as the target product category. If not, execute step S43; Step S43: Execute the third judgment logic, which is to determine whether all the third output results are different. If so, use the candidate product category output by a trusted third AI model as the target product category.

[0039] The above hierarchical judgment operation is used to determine the target product category from the third output results of each third AI model. It can also be: when all the third output results are the same candidate product category in the candidate product category set, use this candidate product category as the target product category; when there are the same candidate product categories in the third output results, use the candidate product category with the most similarities as the target product category; when all the third output results are different candidate product categories, define a trusted third AI model and use the third output result of the third AI model as the target product category.

[0040] In an embodiment, the candidate product category set output by the second AI model includes candidate product categories A, B, C, D, and E. The three third AI models are AI1, AI2, and AI3 respectively. The original product category and the candidate product categories are structurally processed to obtain a structural processing result that can be directly input into each third AI model: which candidate product category in the original product category and the candidate product category set has the highest similarity. Input the structural processing result into each third AI model, and get: the third output result of AI1 is that the similarity between the original product category and candidate product category A is the highest; the third output result of AI2 is that the similarity between the original product category and candidate product category A is the highest; the third output result of AI3 is that the similarity between the original product category and candidate product category B is the highest. Perform a hierarchical judgment operation based on the above third output results. Specifically, execute the first judgment logic to determine whether all the third output results of the third AI models are the same. The judgment result is no, so execute the second judgment logic to determine whether there are the same third output results among all the third AI models. The judgment result is yes, and use the candidate product category with the most similarities as the target product category, that is, use candidate product category A as the target product category.

[0041] For all the above steps of calculating similarity, vectorization processing can be performed first and then vector similarity can be calculated. The multiple AI models with different functions used in this solution are technically related to each other.

[0042] In one embodiment, the method for determining a product category based on an AI model further includes step S6: receiving a user's publishing operation, associating a product with a target product category, displaying the association relationship on the user's publishing interface, and receiving an editing operation on the user's publishing interface to adjust the target product category. This solution provides a modification function on the user's publishing interface, that is, when receiving the user's product publishing operation, each step of the above method for determining a product category based on an AI model is executed, and the target product category is displayed on the user's publishing interface to prompt the user to determine the target product category and modify the target product category of the product according to the user's editing operation.

[0043] In addition, in addition to the above verification method, the target product category can also be verified a second time. For example, the method of calibrating the original product category based on the dynamic threshold and confidence probability value is used to verify the target product category, or sample products under the target product category of the product are obtained. The sample products are products that characterize the classification requirements of each product category in the product category table, and the similarity between the sample products and the product is calculated to determine whether the similarity is greater than a preset threshold. If so, it is confirmed that the target product category is correctly matched. If not, it is confirmed that the target product category is incorrectly matched, and the user is prompted to re-verify and perform an editing operation on the target product category.

[0044] The above are only the preferred embodiments of the present application, and do not limit the patent scope of the present application. Any equivalent transformation made using the content of the specification and drawings of the present application under the inventive concept of the present application, or direct / indirect application in other related technical fields, is included in the patent protection scope of the present application.

Claims

1. A method for determining commodity categories based on an AI model, characterized in that: The method comprises: Step S1: determine whether there is product category information in the product details information of the product. If so, use the corresponding product category as the original product category; If it does not exist, a supplement operation is performed through the first AI model, wherein the supplement operation generates the original product category based on the product detail information; Step S2: constructing a category database according to the commodity categories of the target platform, and performing similarity search on the original commodity categories through the second AI model, and extracting a set of commodity categories to be selected from the category database; Step S3: performing structured processing on the original product category and the set of product categories to be selected, and inputting the structured processing results into multiple third AI models to obtain multiple third output results, wherein the structured processing is processing the original product category and the set of product categories to be selected into query instructions that can be directly input into the third AI model; Step S4: performing a step-by-step judgment operation on the third output result to obtain a target commodity category of the commodity in the target platform, wherein the step-by-step judgment operation includes: Step S41: executing a first judgment logic, wherein the first judgment logic includes judging whether all third output results are the same, if so, taking the same to-be-selected commodity category as the target commodity category, otherwise, executing step S42; Step S42: Execute the second judgment logic, wherein the second judgment logic is to determine whether all the third output results are identical, if so, take the majority of identical selected commodity categories as the target commodity categories, if not, execute step S43; Step S43: Execute the third judgment logic, wherein the third judgment logic is to judge whether all the third output results are different. If so, the selected product category output by a trusted third AI model is used as the target product category.

2. The method according to claim 1, characterized in that: The first AI model outputs the original product category based on the product details information of the input product. The second AI model outputs N candidate product categories that are most similar to the original product category based on the input original product category and the category database, 5≤N≤10. The third AI model outputs a temporary product category based on the input structured processing result. The temporary product category is a candidate product category in the candidate product category set that has the highest similarity to the original product category, and the number of third AI models is at least 3.

3. The method according to claim 1, characterized in that: In step S1, the product details are obtained from the original platform where the product is located, and the fields are decomposed to obtain the field information of the product, where the field information includes at least one of the product title, product description, product picture, product category, and other information; Determine whether the commodity category field in the field information is empty; If not, extract the field value of the product category field to obtain the original product category of the product.

4. The method according to claim 2, characterized in that: Step S3 includes: Performing vectorization processing on all commodity categories on the target platform to obtain a category database, and saving a mapping relationship between each commodity category and a corresponding vector, wherein the category database contains vectorized data corresponding to each commodity category on the target platform; Performing vectorization processing on the original product category to obtain original product vector data; Performing a similarity search between the original product vector data and the category database through a second AI model, and extracting N candidate vector data that are most similar to the vector data from the category database; The commodity categories corresponding to the to-be-selected vector data are acquired based on the mapping relationship to obtain a to-be-selected commodity category set.

5. The method according to claim 4, characterized in that: Step S3 also includes: The similarity between the original product vector data and each vector in the category database is calculated, and the vectors in the category database are sorted based on the similarity to extract N candidate vector data that are most similar to the vector data. The calculated similarity is one of the Euclidean distance, cosine distance or vector inner product between the calculated vectors.

6. The method according to claim 2, characterized in that: Step S4 includes: The original product category and the set of product categories to be selected are used as input variables, and a structured model is constructed according to the retrieval mode of each third AI model, and the output result of the structured model is used as the structured processing result. The construction of the structured model includes determining the corresponding query instruction template based on the retrieval mode of each third AI model, filling the original product category and the set of product categories to be selected into the query instruction template, and the structured processing result is the query instruction of each third AI model.

7. The method according to claim 6, characterized in that In step S4, the retrieval mode of the third AI model includes a text retrieval mode, a multimodal retrieval mode and a structured retrieval mode. When the search mode is a text search mode, the query instruction template is a natural language template, and the structured model converts the input variables into a text combination; When the retrieval mode is a multimodal retrieval mode, the query instruction template is a multimodal template, and the structured model converts the input variable into a combination of an image feature vector and a text description; When the search mode is a structured search mode, the query instruction template is a structured query template, and the structured model converts input variables into a field format.

8. The method according to claim 1, characterized in that: Also includes: Step S6: receiving a user's publishing operation, associating the product with the target product category, displaying the association relationship on the user publishing interface, and receiving an editing operation on the user publishing interface to adjust the target product category.

9. The method according to claim 2, characterized in that: In step S3, the similarity corresponding to each to-be-selected commodity category is monitored, and when the similarity retrieved is lower than a preset value, the confidence probability value of the original commodity category output by the first AI model is obtained, and the confidence probability value indicates the probability value that the commodity output by the first AI model belongs to the original commodity category; Determine whether the confidence probability value is lower than a dynamic threshold, where the dynamic threshold is determined by a basic threshold and a heat index, where the heat index represents the release frequency of products under the product category; If yes, the category adjustment logic is triggered. The category adjustment logic is to determine whether there is a conflict between the original product category and the product description of the product according to the product category constraint rules. The product category constraint rules are the requirements for the classification of products in each product category. If it exists, then the sub-product category is associated according to the product description in the product details information, and the sub-product category is the product category with the smallest category level in the product category; If it does not exist, associate the sub-product category based on the product description in the product details and the original product category; The sub-commodity category is confirmed according to the hierarchical relationship between the historical commodity category and the sub-commodity category, and the sub-commodity category is used as the original commodity category. The historical commodity category is the commodity category to which the commodity in the historical commodity release data belongs.

10. A commodity category determination system, characterized in that: The commodity category determination system executes the operation instructions contained in the commodity category determination method based on the AI ​​model described in any one of claims 1-9.

Images