A method and system for generating description information of home commodities

Abstract

The application discloses a kind of household commodity description information generation method and system, belong to information processing technical field, its method includes the following steps: extracting the characteristic representation information of each historical household commodity description information, generates characteristic representation information set;The historical household commodity description information set is input into description information generation model, and description information generation model is trained;The description information generation model after training is corrected, and the description information correction model is obtained;Description information correction model is used to generate the description information of the household commodity to be processed.The application constructs the description information generation model of stronger practicability and universality on one hand, and on the other hand, the more representative characteristic representation information is extracted to correct the description information generation model, so that the generated household commodity description information is closer to the commodity itself, and is suitable for various types of household commodities, and the generation content is more detailed, and the wording is also more perfect.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of information processing, and particularly relates to a household commodity description information generation method and system. BACKGROUND

[0002] The existing household app platform mainly inputs commodity descriptions and descriptions manually, which wastes a lot of human resources and cannot accurately describe household commodity information, so that the selling points cannot be grasped, the user interest cannot be well captured, and the business efficiency and commodity sales rate and other related indicators are poor. SUMMARY

[0003] The application discloses a household commodity description information generation method and system.

[0004] The technical scheme of the application is as follows:

[0005] Obtain a historical household commodity description information set, extract feature representation information of each historical household commodity description information, and generate a feature representation information set;

[0006] Construct a description information generation model, input the historical household commodity description information set into the description information generation model, and train the description information generation model;

[0007] Correct the trained description information generation model by using the feature representation information set to obtain a description information correction model;

[0008] Generate description information of a household commodity to be processed by using the description information correction model.

[0009] Further, obtaining a historical household commodity description information set, extracting feature representation information of each historical household commodity description information, and generating a feature representation information set include the following steps:

[0010] Split the historical household commodity description information into a plurality of words, and extract word vectors of each word in the historical household commodity description information;

[0011] According to the word vectors of each word in the historical household commodity description information, calculate semantic description distances of each historical household commodity description information;

[0012] Sort all semantic description distances from large to small, extract keywords of historical household commodity description information corresponding to the top N semantic description distances, take the keywords as feature representation information of each historical household commodity description information, and generate a feature representation information set; wherein N represents the number of extracted semantic distances.

[0013] The beneficial effect of the further scheme is that in the present application, among all historical home commodity description information, a large number of words describing the characteristics of home commodities are contained, such as specifications, colors and materials and the like, but not all descriptions are effective descriptions, so the present application uses semantic description distance to select more representative historical home commodity description information, and uses the keywords of the more representative historical home commodity description information as characteristic representation, which can objectively and accurately reflect the required keywords for describing home commodities, and facilitate further correction of the description information generation model.

[0014] Further, the calculation formula of the semantic description distance S of the historical home commodity description information is:

[0015]

[0016] In the formula, x m represents the word vector of the mth word in the historical home commodity description information, x m+1 represents the word vector of the m+1th word in the historical home commodity description information, c represents a constant, M represents the number of words in the historical home commodity description information, and C m represents the word frequency of the mth word in the historical home commodity description information.

[0017] Further, the calculation formula of the number N of extracted semantic distances is:

[0018]

[0019] In the formula, Q represents the number of all historical home commodity description information, represents the upward rounding operation, q represents the number of semantic description distances greater than the average of all semantic description distances, and ε represents a minimum value.

[0020] Further, the description information generation model comprises an information input layer, an information matrix extraction layer, an information matrix transposition layer, a first convolutional layer, a second convolutional layer, a normalization layer, a fusion layer and an information output layer.

[0021] The input end of the information input layer is connected with the output end of the information matrix extraction layer, the first output end of the information matrix extraction layer is connected with the input end of the information matrix transposition layer, the second output end of the information matrix extraction layer is connected with the input end of the first convolutional layer, the output end of the information matrix transposition layer is connected with the input end of the second convolutional layer, the first output end of the first convolutional layer is connected with the first input end of the normalization layer, the second output end of the first convolutional layer is connected with the second input end of the normalization layer, the first output end of the second convolutional layer is connected with the first input end of the fusion layer, the second output end of the second convolutional layer is connected with the second input end of the fusion layer, the output end of the fusion layer is connected with the input end of the information output layer, and the output end of the information output layer is the output end of the description information generation model.

[0022] The information matrix extraction layer fills the word vectors of all words in the historical home commodity description information into the matrix template, fills the vacancies with 1, and outputs several matrices.

[0023] Further, the loss function F of the description information generation model is expressed as:

[0024]

[0025]

[0026]

[0027]

[0028] In the formula, LeakyRelu(·) represents an activation function, w0 represents the weight of the fusion layer, X1 represents the output of the first convolutional layer, w1 represents the weight of the first convolutional layer, X2 represents the output of the second convolutional layer, w2 represents the weight of the second convolutional layer, X3 represents the output of the normalization layer, w3 represents the weight of the normalization layer, max(·) represents a maximum value function, min(·) represents a minimum value function, e represents an index, c represents a constant, Y k represents the kth output matrix of the information matrix extraction layer, D k represents the rank of the kth output matrix of the information matrix extraction layer, Z g represents the gth output matrix of the information matrix transposition layer, d gRank of the gth output matrix of the information matrix transposition layer.

[0029] Further, the description information generation model is corrected by using the feature representation information set to obtain a description information correction model, including the following steps:

[0030] Constructing a feature association rule;

[0031] Extracting feature representation information in the feature representation information set that meets the feature association rule;

[0032] The description information generation model is corrected by using the feature representation information that meets the feature association rule to obtain a description information correction model.

[0033] The beneficial effects of the above further scheme are: in the present application, the description information generation model contains a large number of matrix operations representing semantic vectors, so the present application also corrects the description information generation model, so that the description information generated by the description information correction model is more accurate.

[0034] Further, the expression of the feature association rule is:

[0035] H j ≤H ave +1

[0036] In the formula, H j represents the average word frequency of all keywords in the jth feature representation information, H ave represents the average of all keywords in the feature representation information set.

[0037] Further, the expression of the description information correction model F' is:

[0038]

[0039] In the formula, LeakyRelu(·) represents an activation function, w0 represents the weight of the fusion layer, X1 represents the output of the first convolutional layer, w1 represents the weight of the first convolutional layer, X2 represents the output of the second convolutional layer, w2 represents the weight of the second convolutional layer, X3 represents the output of the normalization layer, w3 represents the weight of the normalization layer, and P0 represents the average word frequency of all keywords in the feature representation information that meets the feature association rule.

[0040] The beneficial effects of this invention are as follows: In order to help home furnishing merchants reduce labor costs and improve the actual product introduction effect, this invention discloses a method for generating home furnishing product description information. On the one hand, it constructs a description information generation model with stronger practicality and universality. On the other hand, it corrects the description information generation model by extracting more representative feature representation information, so that the generated home furnishing product description information is closer to the product itself and is applicable to various types of home furnishing products. The generated content is more detailed and the wording is more perfect.

[0041] Based on the above methods, the present invention also proposes a home furnishing product description information generation system, including an information set generation unit, an information model training unit, an information model correction unit, and a description information generation unit;

[0042] The information set generation unit is used to obtain a set of historical home furnishing product description information, extract the feature representation information of each historical home furnishing product description information, and generate a set of feature representation information.

[0043] The information model training unit is used to build a description information generation model and inputs a set of historical home furnishing product description information into the description information generation model to train the description information generation model;

[0044] The information model correction unit is used to correct the trained description information generation model using the feature representation information set, resulting in a description information correction model.

[0045] The description information generation unit is used to generate description information for the home furnishing products to be processed by modifying the model using description information.

[0046] The beneficial effects of this invention are: This invention discloses a home furnishing product description information generation system, which generates home furnishing product description information that accurately, objectively and vividly describes the characteristics of home furnishing products, effectively improving the image of home furnishing products. Attached Figure Description

[0047] Figure 1 A flowchart illustrating the method for generating product description information for home furnishings;

[0048] Figure 2 This is a schematic diagram illustrating the structure of the information generation model.

[0049] Figure 3 A schematic diagram of the system for generating description information for home furnishing products. Detailed Implementation

[0050] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0051] like Figure 1 As shown, the present invention provides a method for generating description information of home furnishing products, including the following steps:

[0052] obtain a set of historical home commodity description information, and extract feature representation information of each historical home commodity description information to generate a set of feature representation information;

[0053] construct a description information generation model, input the set of historical home commodity description information into the description information generation model, and train the description information generation model;

[0054] correct the trained description information generation model by using the set of feature representation information to obtain a description information correction model;

[0055] generate description information of a to-be-processed home commodity by using the description information correction model.

[0056] In the embodiment of the present application, a set of historical home commodity description information is obtained, and feature representation information of each historical home commodity description information is extracted to generate a set of feature representation information, including the following steps:

[0057] The historical home commodity description information is split into a plurality of words, and a word vector of each word in the historical home commodity description information is extracted;

[0058] According to the word vector of each word in the historical home commodity description information, the semantic description distance of each historical home commodity description information is calculated;

[0059] All semantic description distances are sorted from large to small, and the keywords of the historical home commodity description information corresponding to the top N semantic description distances are extracted, the keywords are taken as the feature representation information of each historical home commodity description information, and a set of feature representation information is generated; wherein N represents the number of extracted semantic distances.

[0060] In the present application, among all historical home commodity description information, a large number of words describing the characteristics of home commodities are contained, such as specifications, colors and materials, etc., but not all descriptions are effective descriptions, so the present application uses semantic description distance to select more representative historical home commodity description information, and the keywords of these more representative historical home commodity description information are taken as feature representation, which can objectively and accurately reflect the key words required for describing home commodities, and facilitate further correction of the description information generation model.

[0061] In the embodiment of the present application, the calculation formula of the semantic description distance S of the historical home commodity description information is:

[0062]

[0063] In the formula, x m represents the word vector of the mth word in the historical home commodity description information, x m+1Let C represent the word vector of the (m+1)th word in the historical home furnishing product description information, where c represents a constant, M represents the number of words in the historical home furnishing product description information, and C... m This represents the word frequency of the m-th word in the description information of historical home furnishing products.

[0064] In this embodiment of the invention, the formula for calculating the number of extracted semantic distances N is as follows:

[0065]

[0066] In the formula, Q represents the number of historical home furnishing product descriptions. This indicates the rounding up operation, q represents the number of semantic description distances greater than the average of all semantic description distances, and ε represents the minimum value.

[0067] In embodiments of the present invention, such as Figure 2 As shown, the information generation model includes an information input layer, an information matrix extraction layer, an information matrix transpose layer, a first convolutional layer, a second convolutional layer, a normalization layer, a fusion layer, and an information output layer.

[0068] The input terminal of the information input layer serves as the input terminal of the descriptive information generation model, and its output terminal is connected to the input terminal of the information matrix extraction layer. The first output terminal of the information matrix extraction layer is connected to the input terminal of the information matrix transpose layer, and its second output terminal is connected to the input terminal of the first convolutional layer. The output terminal of the information matrix transpose layer is connected to the input terminal of the second convolutional layer. The first output terminal of the first convolutional layer and the first output terminal of the second convolutional layer are respectively connected to the first and second input terminals of the normalization layer. The second output terminal of the first convolutional layer, the second output terminal of the second convolutional layer, and the output terminal of the normalization layer are respectively connected to the first, second, and third input terminals of the fusion layer. The output terminal of the fusion layer is connected to the input terminal of the information output layer. The output terminal of the information output layer serves as the output terminal of the descriptive information generation model.

[0069] In this invention, the information matrix extraction layer can fill the matrix template with the word vectors of all words in the historical home furnishing product description information, fill the gaps with 1, and output several matrices; the information matrix transpose layer is used to transpose the matrix output by the information matrix extraction layer; the first convolutional layer and the second convolutional layer perform mathematical operations on the matrices generated by the information matrix extraction layer and the information matrix transpose layer, respectively, and use the operation results as the input of the normalization layer to obtain another operation result; the fusion layer uses the activation function to process the three operation results and outputs them by the information output layer.

[0070] In this embodiment of the invention, the loss function F of the descriptive information generation model is expressed as:

[0071]

[0072]

[0073]

[0074]

[0075] In the formula, LeakyRelu(·) represents an activation function, w0 represents a weight of a fusion layer, X1 represents an output of a first convolutional layer, w1 represents a weight of the first convolutional layer, X2 represents an output of a second convolutional layer, w2 represents a weight of the second convolutional layer, X3 represents an output of a normalization layer, w3 represents a weight of the normalization layer, max(·) represents a maximum value function, min(·) represents a minimum value function, e represents an exponential, c represents a constant, Y k represents the kth output matrix of the information matrix extraction layer, D k represents the rank of the kth output matrix of the information matrix extraction layer, Z g represents the gth output matrix of the information matrix transposition layer, d g represents the rank of the gth output matrix of the information matrix transposition layer.

[0076] In the embodiment of the present application, the feature representation information set is used to correct the description information generation model to obtain a description information correction model, including the following steps:

[0077] Constructing a feature association rule;

[0078] Extracting feature representation information in the feature representation information set that meets the feature association rule;

[0079] Using the feature representation information that meets the feature association rule to correct the description information generation model to obtain a description information correction model.

[0080] In the present application, the description information generation model contains a large number of matrix operations representing semantic vectors, so the present application also corrects the description information generation model to make the description information generated by the description information correction model more accurate.

[0081] In the embodiment of the present application, the expression of the feature association rule is:

[0082] H j ≤H ave +1

[0083] In the formula, H j represents the average of the word frequencies of all keywords in the jth feature representation information, H ave represents the average of all keywords in the feature representation information set.

[0084] In the embodiment of the present application, the expression of the description information correction model F' is:

[0085]

[0086] In the formula, LeakyRelu(·) represents an activation function, w0 represents a weight of the fusion layer, X1 represents an output of the first convolutional layer, w1 represents a weight of the first convolutional layer, X2 represents an output of the second convolutional layer, w2 represents a weight of the second convolutional layer, X3 represents an output of the normalization layer, w3 represents a weight of the normalization layer, and P0 represents a mean value of term frequencies of all keywords in the feature representation information conforming to the feature association rule.

[0087] Based on the above method, the application further provides a household commodity description information generation system, as shown in the accompanying drawings, comprising an information set generation unit, an information model training unit, an information model correction unit and a description information generation unit. Figure 3

[0088] The information set generation unit is configured to obtain a historical household commodity description information set, extract feature representation information of each historical household commodity description information, and generate a feature representation information set.

[0089] The information model training unit is configured to construct a description information generation model, input the historical household commodity description information set into the description information generation model, and train the description information generation model.

[0090] The information model correction unit is configured to correct the trained description information generation model by using the feature representation information set to obtain a description information correction model.

[0091] The description information generation unit is configured to generate description information of a household commodity to be processed by using the description information correction model.

[0092] Those skilled in the art will appreciate that the embodiments described herein are presented for the purpose of helping the reader to understand the principles of the application and should be understood as not limiting the scope of protection of the application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations of the embodiments according to the technical inspiration provided by the disclosure of the application without departing from the spirit of the application, and these modifications and combinations are still within the scope of protection of the application.​

Claims

1. A method for generating description information for home furnishing products, characterized in that, Includes the following steps: Obtain a set of historical home furnishing product description information, extract the feature representation information of each historical home furnishing product description information, and generate a set of feature representation information; A description information generation model is constructed, and a set of historical home furnishing product description information is input into the description information generation model to train the model; The trained description information generation model is corrected by using the feature representation information set to obtain the description information correction model; The descriptive information is used to modify the model and generate descriptive information for the home furnishing products to be processed; The process of obtaining a set of historical home furnishing product description information and extracting feature representation information from each historical home furnishing product description information to generate a set of feature representation information includes the following steps: The description information of historical home furnishing products is broken down into several words, and the word vectors of each word in the description information of historical home furnishing products are extracted. Based on the word vectors of each word in the description information of historical home furnishing products, calculate the semantic description distance of each historical home furnishing product description information; Sort all semantic description distances from largest to smallest, extract keywords from the historical home furnishing product description information corresponding to the top N semantic description distances, and use the keywords as feature representation information for each historical home furnishing product description information to generate a feature representation information set; where N represents the number of extracted semantic distances; The process of modifying the descriptive information generation model using a set of feature representation information to obtain a modified descriptive information model includes the following steps: Construct feature association rules; Extract feature representation information from the feature representation information set that conforms to feature association rules; The descriptive information generation model is modified by using feature representation information that conforms to feature association rules, resulting in a modified descriptive information model.

2. The method for generating description information of home furnishing products according to claim 1, characterized in that, The formula for calculating the semantic description distance S of the historical home furnishing product description information is as follows: ； In the formula, x m Let x represent the word vector of the m-th word in the description information of historical home furnishing products. m+1 Let C represent the word vector of the (m+1)th word in the historical home furnishing product description information, where c represents a constant, M represents the number of words in the historical home furnishing product description information, and C... m This represents the word frequency of the m-th word in the description information of historical home furnishing products.

3. The method for generating description information of home furnishing products according to claim 1, characterized in that, The formula for calculating the number of extracted semantic distances N is as follows: ； In the formula, Q represents the number of historical home furnishing product descriptions. This indicates the rounding up operation, q represents the number of semantic description distances greater than the average of all semantic description distances, and ε represents the minimum value.

4. The method for generating description information of home furnishing products according to claim 1, characterized in that, The descriptive information generation model includes an information input layer, an information matrix extraction layer, an information matrix transpose layer, a first convolutional layer, a second convolutional layer, a normalization layer, a fusion layer, and an information output layer. The input terminal of the information input layer serves as the input terminal of the descriptive information generation model, and its output terminal is connected to the input terminal of the information matrix extraction layer; the first output terminal of the information matrix extraction layer is connected to the input terminal of the information matrix transpose layer, and its second output terminal is connected to the input terminal of the first convolutional layer. The output of the transpose layer of the information matrix is ​​connected to the input of the second convolutional layer; The first output of the first convolutional layer and the first output of the second convolutional layer are respectively connected to the first input and the second input of the normalization layer; the second output of the first convolutional layer, the second output of the second convolutional layer, and the output of the normalization layer are respectively connected to the first input, the second input, and the third input of the fusion layer; the output of the fusion layer is connected to the input of the information output layer; the output of the information output layer serves as the output of the descriptive information generation model.

5. The method for generating description information of home furnishing products according to claim 4, characterized in that, The loss function F of the description information generation model is expressed as follows: ； ； ； ； In the formula, LeakyRelu(·) represents the activation function, w0 represents the weights of the fusion layer, X1 represents the output of the first convolutional layer, w1 represents the weights of the first convolutional layer, X2 represents the output of the second convolutional layer, w2 represents the weights of the second convolutional layer, X3 represents the output of the normalization layer, w3 represents the weights of the normalization layer, max(·) represents the maximum value function, min(·) represents the minimum value function, e represents the exponent, c represents the constant, and Y... k D represents the k-th output matrix of the information matrix extraction layer. k Z represents the rank of the k-th output matrix of the information matrix extraction layer. g d represents the g-th output matrix of the information matrix transpose layer. g This represents the rank of the g-th output matrix of the transpose layer of the information matrix.

6. The method for generating description information of home furnishing products according to claim 1, characterized in that, The expression for the feature association rule is: ； In the formula, H j H represents the mean frequency of all keywords in the information represented by the j-th feature. ave The feature represents the mean of all keywords in the information set.

7. The method for generating description information of home furnishing products according to claim 1, characterized in that, The expression for the description information correction model F' is: ; In the formula, LeakyRelu(·) represents the activation function, w0 represents the weight of the fusion layer, X1 represents the output of the first convolutional layer, w1 represents the weight of the first convolutional layer, X2 represents the output of the second convolutional layer, w2 represents the weight of the second convolutional layer, X3 represents the output of the normalization layer, w3 represents the weight of the normalization layer, and P0 represents the average word frequency of all keywords in the feature representation information that conforms to the feature association rule.

8. A system for generating description information for home furnishing products, characterized in that, The home furnishing product description information generation system is used to execute the home furnishing product description information generation method as described in claim 1. The system includes an information set generation unit, an information model training unit, an information model correction unit, and a description information generation unit. The information set generation unit is used to obtain a set of historical home furnishing product description information, and extract the feature representation information of each historical home furnishing product description information to generate a set of feature representation information. The information model training unit is used to construct a description information generation model and input the set of historical home furnishing product description information into the description information generation model to train the description information generation model; The information model correction unit is used to correct the trained description information generation model using the feature representation information set to obtain the description information correction model. The description information generation unit is used to generate description information for the home furnishing products to be processed using the description information correction model.

Images