A system for extracting and identifying characteristic information of online or system products

Abstract

The present invention discloses a system for extracting and identifying online or system commodity feature information, wherein the commodity feature information identification is to extract the attribute field describing a specific commodity, and relates to the field of commodity information feature identification. The feature extraction module comprises: a title module 301, a text description module 302, a picture description module 303, and a symbol separation module 304. The title module 301 obtains the commodity title text, performs word segmentation on the title, calculates the weight of the phrase, and outputs the brand name and commodity feature attributes. The present invention can predict and describe commodity feature information according to the category by setting a deep neural network model, enhance the weight of the commodity feature information calculation, and improve the accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of commodity information feature recognition, and in particular relates to a system for extracting and identifying online or system commodity feature information. Background Art

[0002] The characteristic information of a product can distinguish the differences between products in different fields. The characteristic attributes of a product are a collection of the properties of the product and a collection of the differences of the product. However, the characteristic information of products on e-commerce platforms is irregular and has not been unified. When we want to use the characteristic information of a product to determine its uniqueness, the original data is completely unusable, which is not conducive to the subsequent work.

[0003] After we extract and process the characteristic information of online products, we can confirm the product itself by describing the characteristic information fields of the product. We need a system that can identify the characteristic information of online products to complete this task. Summary of the invention

[0004] In view of the above technical problems, the present invention provides a system for extracting and identifying characteristic information of online or system commodities, which can extract characteristic information describing online commodities and has high working efficiency.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a system for extracting and identifying online or system commodity feature information, characterized in that: the commodity feature information identification is to extract the attribute field describing a specific commodity, and the specific steps are:

[0006] Step S10: Collect product feature information from the e-commerce platform, and manually proofread and annotate the data to make it in dictionary form. The dictionary format is: {category, feature attribute}. Put the product feature information into a recurrent neural network model for training, calculate the correlation between the category and the feature attribute, and save the network model parameters.

[0007] Step S20: Input the product address or push the product information system, the feature extraction module obtains the product details status, determines the existence status of the text description information and the picture description information, and selects different working modes:

[0008] 1) Only text description information exists;

[0009] 2) Both text description information and picture description information exist;

[0010] Step S30: Based on the working mode of step S20, the feature extraction module is used to identify and extract the feature information of the commodity, and the extracted feature information is saved in the original information database;

[0011] Step S40: Process the characteristic information in the original information database, convert the characteristic information in the original information database into characteristic information in the standard information database through word segmentation, cleaning, prediction, sorting and other steps, and format the characteristic information of the product;

[0012] Step S50: Format and combine the extracted feature information: first calculate the correlation between each extracted feature information and the feature information of known products, and obtain the confidence that the feature information can describe the product. Then, combine the high-confidence feature information in an orderly manner so that the feature information can fully describe the product. This feature information contains all the fields that describe the product.

[0013] Furthermore, the specific steps of training the recurrent neural network model in step S10 are:

[0014] Step S101: First, crawl the public data of the commodity sales platform, extract commodity categories and various descriptive features of the commodities, and store them in a table form;

[0015] Step S102: manually proofread and annotate the data so that the data is in the form of {category, characteristic attribute}, and a vocabulary is established, where the category is the upper classification of the product on the e-commerce platform, and the characteristic attribute is a plurality of attributes describing the product;

[0016] Step S103: Input the manually labeled data into the long short-term memory network model, establish the loss function of features, shapes and categories, optimize the loss function, learn and predict the category of the goods by inputting the feature information of the goods, and form an encoder; the network model updates the parameters through back propagation;

[0017] Step S104: Train the long short-term memory network model as a decoder for output: when a product category is input, the characteristic attributes of the product are calculated by updating the parameters through back propagation; combine the encoder and decoder to construct a neural network system that can predict product characteristic information.

[0018] Furthermore, the feature extraction module includes: a title module 301, a text description module 302, a picture description module 303, and a symbol separation module 304. The title module 301 obtains the product title text or the product system recommendation information, segments the title, calculates the phrase weight, and outputs the brand name and product feature attributes; determines whether the picture description module is working by judging whether there is picture description information, and the picture description module extracts the text information of the picture description by performing OCR recognition on the picture; the text obtained by the text description module and the picture description module is segmented by paragraphs, special symbols and other segmentation information through the action of the character segmentation module, and the feature information is divided into paragraphs to make the data format neat; the product feature information is completed by comparing the information extracted by the text description module and the picture description module.

[0019] Furthermore, the specific steps of data processing in step S40 are:

[0020] Step S401: first, the title is segmented, the segmentation result is saved in the form of a word list, and then the marketing title hot words are removed to eliminate high-frequency useless information in the product description;

[0021] Step S402: The cleaned data is passed into the long short-term memory network model, and the model is used to calculate the characteristic information describing the product with high relevance under the product category, and the confidence between the known product characteristic information and the cleaned data is calculated, so as to increase the weight of the product characteristic information of the cleaned data when calculating the characteristic information weight through the data;

[0022] Step S403: outputting feature information with high weight, wherein the feature information includes all feature information fields describing the product, and the product itself can be identified based on the feature information fields.

[0023] Furthermore, the output feature information can be stored in a database as data for expanding the training network model.

[0024] Furthermore, the system includes a deep neural network model and a database for storing data.

[0025] Compared with the prior art, the present invention has the following advantages: (1) the present invention extracts the feature information of online products based on the feature extraction module, and uses the product description to complete the product title features, so that the product feature information is extracted completely; (2) the present invention can predict and describe the product feature information according to the category by setting a deep neural network model, thereby enhancing the weight of the product feature information calculation and improving the accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is the overall workflow diagram of the present invention.

[0027] Figure 2 This is a flow chart of the steps of training a neural network according to the present invention.

[0028] Figure 3 This is a structural diagram of the feature extraction module of the present invention.

[0029] Figure 4 The present invention is a flowchart for judging a picture description.

[0030] Figure 5 Output feature information calculation graph for the present invention. DETAILED DESCRIPTION

[0031] The present invention will be further described below in conjunction with specific embodiments. The illustrative embodiments and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0032] Embodiment: A system for extracting and identifying online or system commodity feature information, wherein the commodity feature information identification is to extract attribute fields describing a specific commodity, and the specific steps are as follows:

[0033] Step S10: Collect product feature information from the e-commerce platform, and manually proofread and annotate the data to make it in dictionary form. The dictionary format is: {category, feature attribute}. Put the product feature information into a recurrent neural network model for training, calculate the correlation between the category and the feature attribute, and save the network model parameters.

[0034] Step S20: Input the product address, the feature extraction module obtains the product details status, determines whether there is text description information and picture description information, and then selects different working modes:

[0035] 1) Only text description information exists;

[0036] 2) Both text description information and picture description information exist;

[0037] Step S30: Based on the working mode of step S20, the feature extraction module is used to identify and extract the feature information of the commodity, and the extracted feature information is saved in the original information database;

[0038] Step S40: Process the characteristic information in the original information database, and convert the characteristic information in the original information database into characteristic information in the standard information database through word segmentation, cleaning, prediction, and sorting, and format the characteristic information of the product;

[0039] Step S50: Format and combine the extracted feature information: first calculate the correlation between each extracted feature information and the feature information of known products, and obtain the confidence that the feature information can describe the product. Then, combine the high-confidence feature information in an orderly manner so that the feature information can fully describe the product. This feature information contains all the fields that describe the product.

[0040] Furthermore, the specific steps of training the neural network model in step S10 are:

[0041] Step S101: First, crawl the public data of the commodity sales platform, extract the commodity categories and commodity description features; summarize the data into a table, and name the table according to the commodity category;

[0042] Step S102: manually proofread and annotate the data to form the form of {category, characteristic attribute}, and establish a vocabulary, where the category is the upper classification of the product on the e-commerce platform, and the characteristic attribute is a plurality of attributes describing the product; for example, pants and T-shirts are classified as "apparel" on the e-commerce platform, and the characteristic attributes describing pants include: "long pants", "nine-point pants", "waistline", "pants length", etc.;

[0043] Step S103: Input the manually labeled data into the long short-term memory network model. The feature attributes and categories are in a corresponding relationship. The feature attributes of the goods are converted into corresponding feature vectors. The classification corresponding to the feature attributes is output in the network model, and the loss is calculated and the parameters are updated.

[0044] Step S104: The long short-term memory network model sets the decoder for output: when the product category is input, the characteristic attributes of the product are calculated by updating the parameters through back propagation, and the loss is minimized through the cooperation of the encoder and decoder, making the network framework prediction result more accurate.

[0045] Furthermore, the feature extraction module includes: a title module 301, a text description module 302, a picture description module 303, and a symbol separation module 304. The title module 301 obtains the product title text or system recommendation information, segments the title, calculates the phrase weight, and outputs the brand name and product feature attributes; determines whether the picture description module is working by judging whether there is picture description information, and the picture description module extracts the text information of the picture description by performing OCR recognition on the picture; the text obtained by the text description module and the picture description module is segmented by paragraphs, special symbols and other segmentation information through the action of the character segmentation module, and the feature information is divided into paragraphs to make the data format neat; the product feature information is completed by comparing the information extracted by the text description module and the picture description module.

[0046] Furthermore, the specific steps of data processing in step S40 are:

[0047] Step S401: First, the title is segmented to remove marketing title hot words and eliminate high-frequency useless information in the product description;

[0048] Step S402: The cleaned data is passed into the long short-term memory network model, and the model is used to calculate the characteristic information describing the product with high relevance under the product category, and the confidence between the known product characteristic information and the cleaned data is calculated, so as to increase the weight of the product characteristic information of the cleaned data when calculating the characteristic information weight through the data;

[0049] Step S403: input feature information with a high weight, the feature information includes all feature information fields describing the product, and the product itself can be determined based on the feature information fields.

[0050] In the above steps, the output feature information can be stored in the database as data for expanding the training network model.

[0051] In this embodiment, the recognition system mentioned includes a deep neural network model and a database for storing data.

Claims

1. A system for extracting and identifying characteristic information of online or system products, characterized by: The commodity feature information recognition is to extract the attribute fields describing specific commodities, and the specific steps are as follows: Step S10: collect commodity feature information on the e-commerce platform, and make the data in a dictionary form through manual data proofreading and annotation. The dictionary format is: {category, feature attribute}, put the commodity feature information into a recurrent neural network model for training, calculate the correlation between the category and the feature attribute, and save the network model parameters; Step S20: Input the product address or push the product information system, the feature extraction module obtains the product details status, determines the existence status of the text description information and the picture description information, and selects different working modes: 1) Only text description information exists; 2) Both text description information and picture description information exist; Step S30: Based on the working mode of step S20, the feature extraction module is used to identify and extract the feature information of the commodity, and the extracted feature information is saved in the original information database; Step S40: Process the feature information in the original information database, convert the feature information in the original information database into feature information in the standard information database through word segmentation, cleaning, prediction, and sorting steps, and format the feature information of the product; Step S50: Format and combine the extracted feature information: First, calculate the correlation between each extracted feature information and the feature information of known products, and obtain the confidence that the feature information can describe the product. Then, combine the feature information with high confidence in an orderly manner so that the feature information can fully describe the product. This feature information contains all the fields that describe the product. The specific steps of training the recurrent neural network model in step S10 are: Step S101: First, crawl the public data of the commodity sales platform, extract commodity categories and various descriptive features of the commodities, and store them in a table form; Step S102: manually proofread and annotate the data so that the data is in the form of {category, characteristic attribute}, and a vocabulary is established, where the category is the upper classification of the product on the e-commerce platform, and the characteristic attribute is a plurality of attributes describing the product; Step S103: Input the manually labeled data into the long short-term memory network model, establish the loss function of features, shapes and categories, optimize the loss function, learn and predict the category of the goods by inputting the feature information of the goods, and form an encoder; the network model updates the parameters through back propagation; Step S104: training the long short-term memory network model as a decoder for outputting: when the commodity category is input, the characteristic attributes of the commodity are calculated by updating the parameters through back propagation; combining the encoder and the decoder to construct a neural network system that can predict the characteristic information of the commodity; the feature extraction module includes: a title module 301, a text description module 302, a picture description module 303, and a symbol separation module 304, the title module 301 obtains the commodity title text or system recommendation information, performs word segmentation on the title, calculates the weight of the phrase, and outputs the brand name and the characteristic attributes of the commodity; by judging whether there is picture description information to determine whether the picture description module is working, the picture description module extracts the text information of the picture description by performing OCR recognition on the picture; the text obtained by the text description module and the picture description module is segmented by paragraphs and special symbol segmentation information through the action of the character segmentation module, and the paragraphs are divided according to the characteristic information to align the data format; by comparing the information extracted by the text description module and the picture description module, the characteristic information of the commodity is completed; the specific steps of data processing in step S40 are: Step S401: first, the title is segmented, the segmentation result is saved in the form of a word list, and then the marketing title hot words are removed to eliminate high-frequency useless information in the product description; Step S402: The cleaned data is passed into the long short-term memory network model, and the model is used to calculate the characteristic information describing the product with high relevance under the product category, and the confidence between the known product characteristic information and the cleaned data is calculated, so as to increase the weight of the product characteristic information of the cleaned data when calculating the characteristic information weight through the data; Step S403: Output feature information with high weight, the feature information includes all feature information fields describing the product, and the product itself can be identified based on the feature information fields; the output feature information can be stored in a database as data for expanding the training network model; the system includes a deep neural network model and a database for storing data.

Images