JP6884116B2 - Information processing equipment, information processing methods, and programs - Google Patents

Description

The present invention relates to an information processing device, an information processing method, and a program.

Generally, when a product or service is sold on a shopping site or the like, a hierarchical category is assigned to the product or service. In this regard, if the feature words in the text and the related words associated with the category are weighted, based on the match score between the feature word weights and the related word weights, A technique for determining a product category is known (see, for example, Patent Document 1).

Japanese Unexamined Patent Publication No. 2010-445884

However, in the prior art, the goods or services may be classified into inappropriate categories, resulting in reduced search accuracy of the goods or services.

The present invention has been made in view of the above problems, and provides an information processing device, an information processing method, and a program capable of classifying products or services into more appropriate categories and improving search accuracy. The purpose is.

One aspect of the present invention is an acquisition unit that acquires product data related to products or services sold via a network, and a product input using product data of products or services classified into hierarchical categories as learning data. Based on the output result of the classifier learned to output the classification result of the hierarchical category for the product or service indicated by the data, the category of the product or service indicated by the product data acquired by the acquisition unit. It is an information processing apparatus including a determination unit for determining.
Is.

According to one aspect of the present invention, goods or services can be classified into more appropriate categories to improve search accuracy.

It is a figure which shows an example of the information processing system 1 including the information processing apparatus 100 in 1st Embodiment. It is a figure which shows an example of the web page provided by the service providing apparatus 20. It is a figure which shows an example of a hierarchical category. It is a figure which shows an example of the structure of the information processing apparatus 100 in 1st Embodiment. It is a flowchart which shows the flow of a series of processing at the time of operation by the information processing apparatus 100 in 1st Embodiment. It is a figure which shows an example of the product data 134 handled on the sales site. It is a figure which shows an example of the classifier 200 in 1st Embodiment. It is a figure which developed the processing contents of RNN250 according to the lapse of time in parallel. It is a figure which shows an example of the classification result of a category. It is a figure which shows an example of the category error candidate list 136. It is a figure which shows an example of the web page of the sales site where products are listed in the order of ranking. It is a figure which shows an example of the web page of the sales site where products are listed in the order of ranking. It is a figure which shows an example of the product registration screen at the time of registering a product on a sales site. It is a figure which shows another example of the product registration screen at the time of registering a product on a sales site. It is a flowchart which shows the flow of a series of processing at the time of learning by the information processing apparatus 100 in 1st Embodiment. It is a figure for demonstrating the label smoothing process. It is a figure for demonstrating the label smoothing process. It is a figure which shows an example of the classifier 200 in 2nd Embodiment. It is a figure which shows an example of the classifier 200 in 3rd Embodiment. It is a figure which shows an example of the hardware composition of the information processing apparatus 100 of an embodiment.

Hereinafter, an information processing apparatus, an information processing method, and a program to which the present invention is applied will be described with reference to the drawings.

[Overview]
The information processing device is realized by one or more processors. The information processing device acquires product data related to products or services sold via a network. The information processing device determines the category to which the product or service indicated by the acquired product data is classified based on the classifier learned in advance. The categories to be determined are defined by a hierarchical structure. The classifier is trained to output the classification result of the hierarchical category for the product or service indicated by the input product data, using the product data of the product or service classified into the hierarchical category as training data. To. By utilizing such a classifier, the information processing apparatus can classify the goods or services into more appropriate categories, and as a result, the search accuracy of the goods or services can be improved. Hereinafter, the goods or services will be collectively referred to as "goods" and will be described.

<First Embodiment>
[overall structure]
FIG. 1 is a diagram showing an example of an information processing system 1 including an information processing device 100 according to the first embodiment. The information processing system 1 in the first embodiment includes, for example, one or more terminal devices 10, a service providing device 20, and an information processing device 100. These devices are connected via the network NW. Further, a part of these devices may be included in another device as a virtual device. For example, a part or all of the functions of the service providing device 20 are realized by the functions of the information processing device 100. It may be a virtual machine.

Each device shown in FIG. 1 transmits and receives various information via the network NW. The network NW includes, for example, the Internet, a WAN (Wide Area Network), a LAN (Local Area Network), a provider terminal, a wireless communication network, a wireless base station, a dedicated line, and the like. It should be noted that not all combinations of the devices shown in FIG. 1 need to be able to communicate with each other, and the network NW may include a local network in part.

The terminal device 10 is a terminal device including an input device, a display device, a communication device, a storage device, and an arithmetic device, such as a mobile phone such as a smartphone, a tablet terminal, and various personal computers. The communication device includes a network card such as a NIC (Network Interface Card), a wireless communication module, and the like. In the terminal device 10, a UA (User Agent) such as a web browser or an application program is activated, and a request according to the content input by the user is transmitted to the service providing device 20. Further, the terminal device 10 in which the UA is activated causes the display device to display various images based on the information acquired from the service providing device 20.

The service providing device 20 is, for example, a web server that provides a web page to the terminal device 10 in response to a request from a web browser started as a UA. The web page may be, for example, a web page constituting a website such as a shopping site, an auction site, or a free market site that sells products on the Internet (hereinafter, collectively referred to as a "sales site"). Further, the service providing device 20 may provide the terminal device 10 with a web page that provides various services such as a search site, an SNS (Social Networking Service), and a mail service. Further, the service providing device 20 is an application server that provides the same service as various websites such as a sales site by providing the content to the terminal device 10 in response to a request from an application started as a UA. May be good.

FIG. 2 is a diagram showing an example of a web page provided by the service providing device 20. In the illustrated example, the web page of the sales site is schematically shown. For example, on the web page of the sales site, an image of the product (R1 in the figure), a sentence indicating the title of the product (R2 in the figure), a sentence indicating the category of the product (R3), and a sentence indicating the outline of the product (in the figure). Contents such as R4) are included. For example, a seller who opens a store on a sales site and sells a product registers the product on the web page of the store where the store is opened, such as an image of the product, a sentence indicating the title of the product, and a sentence indicating an outline of the product. Enter at times. In addition, the seller sets a hierarchical category for the registered product at the same time when the product is registered on the web page.

FIG. 3 is a diagram showing an example of hierarchical categories. For example, on a sales site, the highest level generally has the most abstract category such as "fashion", and the second highest level has a more specific "fashion" than "fashion". A category such as "ladies'fashion" is set, a category such as "ladies'shoes" that is more specific than "ladies' fashion" is set in the third highest hierarchy, and a category such as "ladies' shoes" is set in the fourth highest hierarchy. A category such as "moccasin shoes" that is more specific than "ladies' shoes" is set, and a brand name that is more specific than "moccasin shoes" is set as a category in the fifth higher hierarchy.

These hierarchical categories can be set independently by each store. Therefore, as illustrated in FIG. 2, the higher category of "moccasin shoes" should be a category such as "ladies' shoes", but the seller mistakenly or intentionally made "ice cream". It may happen that a category that is not related to fashion such as "soft serve ice cream" is given.

If such a category setting error occurs, when searching across the products of all stores that open on the sales site, the products that should be searched are not searched or should not be searched. The product may be searched. For example, if the category "ice cream, soft serve" is set for a product related to "ladies'shoes", the product will be posted to users who are looking for ladies' shoes using the sales site. You will not be able to serve web pages.

In addition, if the category "ice cream, soft serve" is set for products related to "ladies'shoes", it is related to "ladies' shoes" to users who are looking for ice cream using the sales site. You will mistakenly provide a web page that contains the product you have purchased. In this way, when a category setting error occurs, the convenience of the sales site tends to decrease.

The information processing device 100 acquires product data related to products registered on the sales site, and machine-learns the acquired product data to detect an error in setting a category set by the seller for the product. The product data includes a plurality of different types of contents such as the above-mentioned image of the product, the title of the product, the category of the product, and the outline of the product. That is, product data is a collection (data set) of a plurality of data of different types.

[Configuration of information processing device]
FIG. 4 is a diagram showing an example of the configuration of the information processing apparatus 100 according to the first embodiment. As shown in the figure, the information processing device 100 includes, for example, a communication unit 102, a control unit 110, and a storage unit 130.

The communication unit 102 includes, for example, a communication interface such as a NIC. The communication unit 102 communicates with the terminal device 10, the service providing device 20, and the like via the network NW.

The control unit 110 includes, for example, an acquisition unit 112, a morphological analysis unit 114, a category determination unit 116, a list generation unit 118, an information providing unit 120, and a learning processing unit 122. The list generation unit 118 is an example of the “extraction unit”.

The components of the control unit 110 are realized, for example, by a processor such as a CPU (Central Processing Unit) executing a program stored in the storage unit 130. In addition, some or all of the components of the control unit 110 are hardware such as LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), or GPU (Graphics Processing Unit). It may be realized by (circuit unit; circuitry), or it may be realized by the cooperation of software and hardware.

The storage unit 130 is realized by, for example, a storage device such as an HDD (Hard Disc Drive), a flash memory, an EEPROM (Electrically Erasable Programmable Read Only Memory), a ROM (Read Only Memory), and a RAM (Random Access Memory). In addition to various programs such as firmware and application programs, the storage unit 130 stores classifier information 132, sales site handling product data 134, and category error candidate list 136. Details of these will be described later.

[Processing flow during operation]
Hereinafter, a series of processing flows during operation by the information processing apparatus 100 in the first embodiment will be described with reference to a flowchart. The operation time is a state in which the already learned classifier 200 is used. FIG. 5 is a flowchart showing a series of processing flows during operation by the information processing apparatus 100 according to the first embodiment. The processing of this flowchart may be repeated, for example, at a predetermined cycle.

First, the acquisition unit 112 causes the communication unit 102 to communicate with the service providing device 20, and acquires the product data of each product sold on the sales site from the service providing device 20 (S100). Then, the acquisition unit 112 integrates the acquired plurality of product data and stores the acquired product data in the storage unit 130 as the sales site handling product data 134.

FIG. 6 is a diagram showing an example of product data 134 handled on the sales site. As shown in the figure, the product data 134 handled by the sales site includes a category C _ij set by the seller of each store, a product title, and a product ID for the product ID indicating the identification information of each product sold on the sales site. It is data in which the outline of the product and the image of the product are associated with each other. The record of the product data 134 handled by the sales site represents each product data. “I” in category C _ij represents the depth of the hierarchy (what number of hierarchy is counted from the highest hierarchy), and “j” represents the type of category in each hierarchy. For example, "C ₁₁ " indicates that the category of the highest hierarchy is set to the category "1", and "C ₂₅ " is the second hierarchy (one step deeper than the highest hierarchy). Hierarchy) category is set to the category "5".

Next, the morphological analysis unit 114 performs morphological analysis on the sentences (title and summary) included in each product data included in the product data 134 handled by the sales site (S102). For example, the morphological analysis unit 114 performs morphological analysis on the title and summary of each product data, and divides each of the title and summary into a plurality of morphemes. The morphological analysis unit 114 vectorizes each of a plurality of words (words) or phrases (phrases) divided as morphemes.

For example, the morphological analysis unit 114 may vectorize words and phrases by using a technique called one-hot expression, and is called a distributed expression such as word2vec or GloVe or a word embedding (s). Words and phrases may be vectorized using techniques.

One-hot expression is to compare a list (dictionary) of words or phrases prepared in advance with words or phrases obtained by morphological analysis, and appear in the list of words or phrases obtained by morphological analysis. This is a method of generating a multidimensional vector such that the element value of a word or phrase is 1 and the element value of another word or phrase is 0.

Further, the distributed expression is a method of vectorizing a word or phrase based on the co-occurrence of the word or phrase obtained by morphological analysis and the word or phrase appearing before or after the word or phrase, for example. , Based on a corpus containing a plurality of words or phrases prepared in advance, the appearance probability of a word or phrase that appears before or after the word or phrase obtained by morphological analysis is obtained, and the appearance probability is used as an element value. It is a method to generate a vector of.

For example, the morphological analysis unit 114 disperses the words or phrases by inputting the words or phrases obtained by the morphological analysis into the neural network to which both the Countuous Bag-of-Words and Skip-gram learning models are applied. You may generate the represented vector.

The morphological analysis unit 114 generates a matrix of second-order tensors (hereinafter referred to as a word matrix) in which a vectorized word or phrase (hereinafter referred to as a word vector) is used as a row vector. For example, when the title is divided into 10 words by morphological analysis, the morphological analysis unit 114 uses a word matrix in which each of the 10 word vectors is a row vector (matrix of 10 rows × n columns: n is a word vector). The number of dimensions of) is generated.

Next, the category determination unit 116 constructs (generates) the classifier 200 with reference to the classifier information 132, and inputs the word matrix generated by the morphological analysis unit 114 into the constructed classifier 200 (S104). ).

The classifier information 132 is information (program or data structure) that defines a classifier. The classifier 200 is realized by various neural networks such as a convolutional neural network (CNN) and a recurrent neural network (RNN), for example.

In the classifier information 132, for example, how the neurons (units) contained in the input layer, one or more hidden layers (intermediate layers), and the output layers constituting each neural network are connected to each other. It includes various information such as connection information and connection coefficients given to data input / output between connected neurons. The connection information includes, for example, the number of neurons contained in each layer, information that specifies the type of neuron to which each neuron is connected, an activation function that realizes each neuron, a gate provided between neurons in the hidden layer, and the like. Contains information. The activation function that realizes the neuron may be, for example, a rectified linear function (ReLU function), a sigmoid function, a step function, or other function. The gate selectively passes or weights the data transmitted between neurons, for example, depending on the value returned by the activation function (eg 1 or 0). The coupling coefficient is a parameter of the activation function. For example, in a hidden layer of a neural network, when data is output from a neuron in a certain layer to a neuron in a deeper layer, the weight given to the output data is given. including. Further, the coupling coefficient may include a bias component peculiar to each layer.

FIG. 7 is a diagram showing an example of the classifier 200 according to the first embodiment. The classifier 200 in the first embodiment is realized by, for example, a first CNN 210, a second CNN 220, a third CNN 230, an FCNN (Fully-Connected Neural Network) 240, and an RNN 250.

The first CNN 210 and the second CNN 220 are convolutional neural networks learned to output a vector having a predetermined number of dimensions when a word matrix is input.

For example, when the word matrix is an image in which each element value is replaced with a pixel value, the first CNN210 and the second CNN220 convolve (obtain the inner product) the image corresponding to the input word matrix and a predetermined filter, and predetermined. The image area containing a plurality of pixels overlapping with the filter of is compressed into one unit area. A univariate value obtained by convolution is associated with this unit area as a feature quantity. The first CNN 210 and the second CNN 220 repeat the convolution process while shifting a predetermined filter on the image to generate a feature image in which the feature amount of the convolution called a convolutional layer formed by a plurality of unit areas is used as a pixel value. To do. The first CNN 210 and the second CNN 220 compress the size (number of dimensions) of each feature image by pooling (for example, Global Average Pooling) and output the result. Since each of the compressed feature images is associated with a univariate feature, a multidimensional vector (first-order tensor) having the convolution feature as an element is output for the number of convolution processes. ..

The third CNN 230 is a convolutional neural network learned to output a vector having a predetermined number of dimensions when an image is input. Like the first CNN 210 and the second CNN 220, the third CNN 230 has elements for the number of times of the convolution process, and outputs a multidimensional vector (first-order tensor) in which these elements are features of the convolution.

For example, the category determination unit 116 inputs the word matrix obtained from the product title into the first CNN 210, inputs the word matrix obtained from the product outline into the second CNN 220, and inputs the product image into the third CNN 230. As a result, each CNN converts the input data into a vector having a predetermined number of dimensions and outputs the data to the FCNN 240.

FCNN240 is a neural network in which each of the units of each layer is connected to each of the units of the next layer having a deeper layer. The FCNN240 may be two layers in which each of the input layer units having no intermediate layer and each of the output layer units are connected to each other. For example, in FCNN240, when the number of dimensions of the vector output from the first CNN210 is 500, the number of dimensions of the vector output from the second CNN220 is 500, and the number of dimensions of the vector output from the third CNN230 is 200, all of them. Outputs one vector of 1200 dimensions, which is the sum of the number of dimensions of the vector of.

The RNN250 corresponds to, for example, an RNN in which the middle layer of the network is an RSTM (Long short-term memory), and when a vector that is a first-order tensor is input, the title, outline, and image that are the basis of the vector are input. It is a neural network learned to output a score indicating the certainty (probability) of the category to which the product is classified for each layer. For example, LSTMs calculate the inner product of weights and state variables. Formula (1) represents the formula for calculating gates and neurons in LSTMs. A state variable is a vector obtained by each gate.

t represents the processing cycle (processing time) of the recursive processing repeatedly performed by the RNN 250, x _t represents the vector input from the FCNN 240 in the processing cycle t (hereinafter, input vector), and h _t represents the processing. It represents a vector (hereinafter referred to as an output vector) output by the RNN 250 in the period t. Recursive processing is to use the output vector obtained in the past processing cycle to derive the output vector in the current processing cycle.

i _t is the vector output by the input gates included in LSTM (hereinafter, input gate vector) represents the, sigma represents the possible activation function of the gate is a sigmoid function, W _i is The weight for linearly converting the input vector x _{t is represented, R i represents the weight for linearly transforming the output vector h t-1} of the previous processing cycle t-1, and Q _i is the internal calculation vector c _t-1 of the processing cycle of the previous stored in the memory cell t-1 represents the weight for linear conversion, b _i represents the bias component of the input gate. The memory cell is a storage area (LSTM block) for temporarily storing the internal calculation vector c.

For example, RNN250 includes inner product value of the weight _{W i} and the input vector _{x t,} weight _{R i} and the inner product value between the previous output vector _{h t-1,} weight _{Q i} and the previous internal operation vector _{c t-1} and Hadamard product (the product of each element) and, by solving the sigmoid function σ that the sum of the bias component b _i as a variable, to derive the input gate vector i _t of the processing cycle t.

f _t is a vector output by FORGET gates included in LSTM (hereinafter, forget gate vector) represents the, W _f, like weight W _i, for linear transformation of the input vector x _t It represents a weight, _{R f,} like weight _{R i,} the output vector _{h t-1} of the previous processing cycle t-1 represents the weight for linear conversion, _{Q f} is the weight _{Q i} Similarly, the weight for linearly converting _{the internal calculation vector c t-1} of the previous processing cycle t-1 stored in the memory cell _{is represented, and b f} represents the bias component of the forget gate. There is.

For example, the RNN 250 has _{an inner product value of the weight W f} and the input vector x _t , an inner product value of the weight R _f and the previous output vector h _t-1 , and the weight Q _f and the previous internal operation vector c _t-1. By solving the sigmoid function σ with the sum of the Hadamard product and the bias component b _f as a variable, the forget gate vector ft of the processing period _t is derived.

c _t denotes the internal operation vectors to be stored in the memory cells included in LSTM.

For example, RNN250 is forget gate vector processing cycle t _{f t} and a Hadamard product of the vector _{z t} to be described later, processing period t-1 for internal calculation vector _{c t-1} and the processing cycle t of Forget gate vector _{f t} the sum of the Hadamard product of the derives as an internal operation vector c _t processing cycle t.

z _t is a hyperpolic whose variable is the sum of the inner product value of the weight W _z and the input vector x _t , the inner product value of the weight R _z and the previous output vector h _t-1 , and the bias component b _z. It is derived by solving the tangent function tanh. W _z represents a weight for linearly transforming the input vector x _{t like the weights Wi i} and W _{f, and R f} is the previous processing cycle t − like the weights R _i and R _f. The weight for linearly transforming the output vector h _{t-1 of 1 is represented, and b z} represents the bias component of the memory cell.

o _t are vectors output by the output gates included in LSTM (hereinafter, output gate vector) represents the, W _o, like the like the weight W _i, to linear transformation of the input vector x _t 's represents the weight, R _o, as with such weight R _i, represents the weight for linear conversion of the output vector h _t-1 of the previous processing cycle t-1, Q _o is memory represents a weight for linear conversion the internal operation vector c _t of the current processing cycle t stored in the cell, b _o represents a bias component of the output gate.

For example, RNN250 is the inner product value of the weight _{W o} and the input vector _{x t,} and the inner product value of the output vector _{h t-1} weight _{R o} before, the weight _{Q o} and the current internal operation vector _{c t} and Hadamard product, by solving the sigmoid function σ that the sum of the bias component b _o as variables, to derive the output gate vector o _t of the processing cycle t.

RNN250, when deriving the output gate vector o _t, the output vector of the output and the gate vector o _t, the Hadamard product of the solution of the internal calculation vector hyper Po Rick tangent function tanh with a c _t as a variable, processing period t derived as h _t.

Each element value of the output vector h _t represents the score of the category for each layer. For _{example, h t = {e1, e2} , e3, ... e9, e10} if it was, element e1 is product represents a score (probability) indicating the certainty of being the category 1, element e2 is commodity Represents a score indicating the certainty that the product is in category 2, and element e3 represents a score indicating the certainty that the product is in category 3. Hereinafter, similarly, each element of e4 to e10 corresponds to the score. It represents the score of the category. The RNN 250 repeatedly derives the output vector by repeating the calculation based on the above-mentioned mathematical formula.

FIG. 8 is a diagram in which the processing contents of the RNN 250 according to the passage of time are developed in parallel. For example, _{in the first processing cycle t1 in which the input vector x t1} is input from the FCNN 240, the RNN 250 outputs the output vector h _t1 . _{The output vector h t1 of} t1 at the time of initial processing includes the score of the highest hierarchy category as an element, and each element includes each category such as fashion, food, home appliances, interior, sports, healthcare, and automobile. Is the score of.

In the next processing cycle t2, the RNN 250 outputs an output vector h _{t2 based on the input vector x t1 input in the first processing cycle t1 and the output vector h t1 in} the previous processing cycle t1. Output vector h _t2 processing cycle t2 includes the scores of the second hierarchical category as an element. For example, of the score of the top-level hierarchy category, if the score of "fashion" is the largest, because the score of "fashion" is most reflected in the output vector h _t2, each element of the output vector h _t2 is, coat Ya It is the score of the lower category of the category "fashion" such as jackets, tops, bottoms, ladies'shoes, and ladies' bags.

In the next processing cycle t3, the RNN 250 outputs the output vector h _{t3 based on the input vector x t1 input in the first processing cycle t1 and the output vector h t2 in} the previous processing cycle t2. Output vector h _t3 processing cycle t3 includes a score for the third hierarchical category as an element. For example, among the second score of the hierarchical categories, when the score of "Women's Shoes" is the largest, because the score of the "Women's Shoes" is most reflected in the output vector h _t3, each element of the output vector h _t3 is , Slip-ons, pumps, sneakers, moccasins, sandals, etc., which is the score of the lower category of the category "ladies'shoes".

In this way, the RNN 250 derives an output vector having a score of a category with a deeper hierarchy as an element by repeating the recursive process. It should be noted that the RNN 250 repeats recursively deriving the output vector until the output vector having the score of the lowest layer category as an element is derived. For example, when the product data to which predetermined additional information (hereinafter referred to as an end label) is given to the lowest layer category is used during the learning of the RNN 250, the RNN 250 is the category to which the end label is given. The recursive process is repeated until the output vector whose element is the score is derived. As a result, for example, even if there are categories "ladies'shoes" and "ladies' bags" in the subcategories of the "fashion" category, and the number of layers existing from those categories to the lowest category is different from each other. Since the recursive process is repeated as many times as the number of layers up to the layer of the category to which the end label is attached, the type of each category can be selected without adjusting the hyper parameter of the number of recursive processes of RNN250 (k in FIG. 8). The score can be obtained while dynamically changing the number of recursive processes accordingly.

Returning to FIG. 5, next, the category determination unit 116 acquires the category classification result from the classifier 200, and determines the product category based on the classification result (S106). The category classification result includes the output vector output by RNN250 for each recursive process, and the output vector includes the result of the score of each layer category from the top layer to the bottom layer as an element. There is.

FIG. 9 is a diagram showing an example of the classification result of the category. As shown in the figure, information associated with each hierarchy category is acquired as a classification result for each product ID. As described above, C _ij in the figure represents the category of the type j of the i-th layer, and the category of the type j represents the category corresponding to the element having the largest value among each element of the output vector. There is. That is, the category of type j has the highest score and represents the most probable category as a product classification destination.

_{For example, the output vector h t} obtained when the product data related to the product whose product ID is "item_A" is input to the classifier 200 is {e1, e2, e3 ... e9, e10}, and the elements e1 to e10. It is assumed that each corresponds to the score of each category of "1" to "10". In such a case, when the element e1 is the largest, the category determination unit 116 is most likely to have the category "1" as the category of the highest hierarchy, so that the category of the highest hierarchy of the product is "1". To decide. Further, when the output vector h _{t + 1} of the second recursive processing is {e1, e2, e3, ... e9, e10} and the element e5 is the largest among them, the category determination unit 116 is the category of the second layer. Since the category "5" is the most probable, the category of the second layer of the product is determined to be "5".

Returning to FIG. 5, the list generation unit 118 compares each product data included in the product data 134 handled by the sales site with the classification result of the category by the classifier 200, and a plurality of products sold on the sales site. From the products, products whose categories set by the seller are different from the categories classified by the classifier 200 (hereinafter referred to as category error product candidates) are extracted (S108).

Next, the list generation unit 118 generates a list of extracted category error product candidates as a category error candidate list 136 and stores it in the storage unit 130 (S110).

FIG. 10 is a diagram showing an example of the category error candidate list 136. For example, when comparing the product data 134 handled by the sales site illustrated in FIG. 6 with the classification result of the category by the classifier 200 illustrated in FIG. 9, the category of the second layer is the product whose product ID is "item_B". Different (C ₂₁ ≠ C ₂₇ ), the category of the first layer is different for the product with the product ID "item_C" (C ₁₃ ≠ C ₁₆ ), and the category of the second layer is also different (C ₂₂ ≠ C _29). ). In such a case, the list generation unit 118 extracts a product having a product ID of "item_B" and a product having a product ID of "item_C" from a plurality of products sold on the sales site, and illustrates them in FIG. Generate a category error candidate list 136 that does.

Returning to FIG. 5, the information providing unit 120 controls the communication unit 102 to transmit the category error candidate list 136 generated by the list generation unit 118 to the predetermined terminal device 10, and the predetermined terminal device 10 The user who can use the above is requested to determine whether or not the product extracted as the category error product candidate is truly a category error (S112). At this time, the information providing unit 120 may transmit the product data of the category error product candidate together with the category error candidate list 136.

The predetermined terminal device 10 may be, for example, a computer of a crowdsourcing participant. The user who is requested to judge the correctness of the category error refers to the product data (product image, product outline, etc.) of the category error product candidate sent together with the category error candidate list 136, and the seller makes the product. It is determined whether the set category is wrong or the category classified by the classifier 200 is wrong, and the determination result is transmitted to the information processing device 100 using the terminal device 10.

Next, the information providing unit 120 waits until the communication unit 102 acquires the correct / incorrect determination result of the category error candidate list 136 from the predetermined terminal device 10 (S114), and the communication unit 102 waits until the communication unit 102 obtains the correct / incorrect determination result of the category error candidate list 136. When the correct / incorrect judgment result is acquired, among one or more category incorrect product candidates included in the category error candidate list 136, the category that is the classification result of the classifier 200 is correct, and the category set by the seller is incorrect. Predetermined information is provided (transmitted) to the service providing device 20 based on the product determined to be (hereinafter, the product in the wrong category) (S116).

For example, the information providing unit 120 includes a category error candidate list 136 including only category error products for which it is determined that the category that is the classification result of the classifier 200 is correct and the category set by the seller is incorrect as predetermined information. May be transmitted to the service providing device 20. In response to this, the service providing device 20 may lower the search ranking of the wrong category product on the sales site, for example. The service providing device 20 is an example of a “search processing unit”.

11 and 12 are diagrams showing an example of a web page of a sales site in which products are listed in order of ranking. In the example of FIG. 11, the products A to Z are listed in the order of ranking from the top to the bottom of the page. Among these products, for example, when the product A is included in the category error candidate list 136 as a category error product, the service providing device 20 lowers the search ranking of the product A and illustrates it in FIG. In addition, the posting position of the product A may be changed from the top of the page to the bottom of the page.

Further, for example, the information providing unit 120 urges the seller who sells the wrong category product on the sales site to change the category currently set for the product to the category classified by the classifier 200 as predetermined information. The information may be transmitted to the service providing device 20 or the terminal device 10 of the seller who sells the wrong category product. When the service providing device 20 receives information prompting the category change as predetermined information from the information processing device 100, the service providing device 20 transmits the information to the terminal device 10 of the seller. This allows the seller to change the category of the product to a more appropriate one.

In the above description, it is assumed that after the seller sets a category for a product on the sales site, an error in setting the category of the product is detected, but the present invention is not limited to this. For example, at the timing when the seller sets a category for a product on the sales site, an error in setting the category of the product may be detected.

FIG. 13 is a diagram showing an example of a product registration screen when registering a product on a sales site. The product registration screen as shown in the figure may be displayed on the display device of the terminal device 10. For example, when the seller inputs a category of a product to be sold on the product registration screen, the service providing device 20 transmits a sentence (text) representing the input category to the information processing device 100. At this time, the service providing device 20 has already registered (uploaded) a part or all of the product image, the text representing the product outline, and the text representing the product title via the product registration screen before the category is input. ), This (these) information is transmitted to the information processing apparatus 100 as product data together with the text indicating the category.

When the information processing device 100 receives the product data from the service providing device 20, the information processing device 100 causes the classifier 200 to classify the product categories, and returns the classification result to the service providing device 20. Upon receiving the classification result of the category by the classifier 200, the service providing device 20 determines whether or not the input category and the category classified by the classifier 200 are different, and if the categories are different, the seller is notified. In order to prompt the user to change the category, as shown in R1 in the figure, the candidate of the category presumed to be correct is displayed on the product registration screen together with the characters and images indicating that the category entered by the seller is incorrect. .. For example, the service providing device 20 displays category candidates on the product registration screen in descending order of score based on the category scores included in the category classification results. In the illustrated example, the category "ice cream" is entered on the product registration screen, whereas "ladies'shoes", "men's shoes", "ladies'bags", "ladies' accessories", "wallets," Along with the candidates of the categories belonging to the same hierarchy of "fashion accessories", the scores of the candidates of those categories are presented as the correct answer probability. In this way, by displaying the categories classified by the classifier 200 together with the score (correct answer probability) on the product registration screen, it is possible to urge the seller to change the category at the time of inputting the category.

Also, at the timing when the seller sets the category for the product on the sales site, in addition to or instead of detecting the setting error of the category of the product, before the seller sets the category for the product, You may present (suggest) the category of the product.

FIG. 14 is a diagram showing another example of the product registration screen when registering the product on the sales site. For example, on the product registration screen, before the seller enters the product category, some or all of the product image, the text representing the product outline, and the text representing the product title have already been entered. If so, the service providing device 20 transmits the product data including the product image, the product outline, and the product title to the information processing device 100.

When the information processing device 100 receives the product data from the service providing device 20, the information processing device 100 causes the classifier 200 to classify the product categories, and returns the classification result to the service providing device 20. When the service providing device 20 receives the classification result of the category by the classifier 200, the service providing device 20 displays the category of the product included in the classification result in the input field of the product category on the product registration screen. At this time, as shown in R2 in the figure, the information processing device 100 may display characters and images indicating that the title of the product is automatically determined by using the classifier 200 on the product registration screen. In this way, before the seller inputs the category, the category is predicted from the product data of the product, and the category that is the prediction result is suggested, which is convenient when registering the product on the sales site. Can be improved.

Further, in the above description, the category determination unit 116 refers to the classification result of the classifier 200 and determines that the category having the highest score in each layer is the category in that layer, but the present invention is not limited to this. For example, the category determination unit 116 causes the RNN 250 to perform the previous recursive processing again when all the scores of the categories in each layer or the scores of a predetermined number or more are less than the threshold value, and the score of the category is the threshold value. The output vector of the hierarchy below may be derived again. For example, RNN250 is, among the plurality of elements e of the output vector _{h t1} derived in processing cycle t1, to derive the output vector _{h t2} most value based on the output vector _{h t1} leaving only a large factor in the processing cycle t2 And. In such a case, when the element e of the output vector h _t2 in the processing cycle t2 were both less than the threshold value, i.e., if the score of the second hierarchy categories were all less than the threshold value, the category determining unit 116 causes the RNN 250 to perform the recursive processing of the processing cycle t1 again. At this time, the category determination unit 116 does not leave the element having the largest value among the plurality of elements e of _{the output vector h t1 derived in the processing cycle t1, for example, the element having the second largest value, or 3} The RNN 250 is made to derive the _{output vector h t2 based on the output vector h t1 in} which the element having the fourth or fourth largest value is left. In this way, if the classification result of the category deeper than the highest level is ambiguous, it is judged that there is a high possibility that the classification of the highest level category is incorrect, and the score of the highest level category is the second. After that, based on the score of the probable category, the score of the category of the next hierarchy is recalculated. As a result, it is possible to suppress the occurrence of an event peculiar to recursive processing in which the classification of the highest-level category is incorrect and the category classification of all subsequent layers is incorrect.

[Processing flow during learning]
Hereinafter, a series of processing flows during learning of the information processing apparatus 100 according to the first embodiment will be described with reference to a flowchart. The learning time is a state in which the classifier 200 used at the time of operation is trained. FIG. 15 is a flowchart showing a series of processing flows during learning by the information processing apparatus 100 according to the first embodiment.

First, the learning processing unit 122 performs sampling processing for extracting high-quality learning data that can improve the performance of the classifier 200 from a plurality of learning data candidates prepared in advance for training the classifier 200. Carry out (S200). The training data candidate is product data in which the correct product category is associated with a teacher label (teacher data), and may be, for example, product data included in the product data 134 handled by the sales site. Product data prepared separately from this may be used. Further, the above-mentioned end label may be given as a teacher label to the lowermost category which is the correct answer.

For example, the learning processing unit 122 extracts the product data that the user has converted from among the plurality of product data treated as candidates for the learning data as high-quality learning data. Conversion means, for example, that a product listed in a list format on a sales site is selected by a click operation or a touch operation, and the screen transitions to the product purchase page, or the product is purchased on the product purchase page. It means that the user has performed a predetermined action expected by the seller, such as staying on the purchase page of the product for a predetermined time or longer.

More specifically, the learning processing unit 122 extracts product data related to products purchased by the user more than a predetermined number of times as high-quality learning data. In addition, the learning processing unit 122 extracts product data related to a product that the user selects by a click operation or a touch operation, accesses the purchase page of the product, and browses the product a predetermined number of times or more, as high-quality learning data. You may. Further, the learning processing unit 122 may extract product data relating to a product having a page stay time of a predetermined time or more after the user accesses the product purchase page as high-quality learning data.

In addition, the learning processing unit 122 may exclude product data to which a predetermined category is assigned as a teacher label from a plurality of product data treated as candidates for learning data as sampling processing. The predetermined category is, for example, the "other" category.

Next, the learning processing unit 122 performs label smoothing processing on a plurality of product data extracted as high-quality learning data (S202).

16 and 17 are diagrams for explaining the label smoothing process. For example, when identification information (a numerical value of 1 to 10 in the illustrated example) indicating the type of the correct category is given to the product data as a teacher label, the score of the correct category is 100. Scores in other categories that are treated as [%] and are not correct are treated as 0 [%]. In the example of FIG. 16, the category of "4" is given as the correct answer category, the category of "4" is 100 [%], and the other scores are 0 [%]. In such a case, as a label smoothing process, the learning processing unit 122 sets the score of the correct answer category to less than 100 [%] and sets the score of the other non-correct category to 0 [%] as illustrated in FIG. ] To be larger than. For example, the learning processing unit 122 uniformly increases the scores of the other categories that are not correct so that the total increase in the scores is the same as the decrease in the scores in the correct category. By such a process, the learning accuracy can be improved.

Next, the learning processing unit 122 selects one product data of interest (hereinafter referred to as attention learning data) from a plurality of product data extracted as high-quality learning data (S204).

Next, the learning processing unit 122 inputs the selected learning data of interest to the classifier 200, causes the classifier 200 to derive the score of the product category indicated by the learning data of interest, and the derived score and label smoothing. The parameters of the classifier 200 are learned so that the error from the score adjusted by the process (that is, the teacher label) is minimized (S206). The parameters of the classifier 200 are, for example, the weights and bias components of the first CNN210, the second CNN220, the third CNN230, the FCNN240, and the RNN250 (LSTM). Further, when an end label is given to the lowermost category as a teacher label, it is learned as a parameter of the classifier 200 so that the number of recursive processes of the RNN 250 becomes the number of layers of the category.

Next, the learning processing unit 122 increments the number of learnings (S208) and determines whether or not the number of learnings has reached a predetermined number (S210).

The learning processing unit 122 performs the bootstrapping process when the number of learnings reaches a predetermined number (S212), and omits the bootstrapping process when the number of learnings does not reach the predetermined number of times. The bootstrapping process is a process of updating (updating) the score determined as the teacher label of the learning data to the score of the category output by the classifier 200. By such processing, when the number of learnings is performed more than a predetermined number of times and the classifier 200 is sufficiently learned, the teacher label of the learning data is updated, so that the training data contains a category that is originally wrong. Learning accuracy can be improved even when the product data given as a teacher label is included.

Next, the learning processing unit 122 determines whether or not all of the plurality of product data extracted as high-quality learning data has been selected as the focus learning data (S214), and still uses all the product data as the focus learning data. If it is not selected, the process returns to S204, new learning data of interest is selected, and the parameters of the classifier 200 are repeatedly learned.

On the other hand, when all the product data are selected as the learning data of interest, the learning processing unit 122 ends the processing of this flowchart.

According to the first embodiment described above, the input product data indicates that the product data related to the products sold via the network is acquired and the product data of the products classified into the hierarchical categories is used as learning data. In order to determine the product category indicated by the acquired product data based on the output result of the classifier 200 learned in advance so as to output the classification result of the hierarchical category for the product, the product (including the service) Can be classified into more appropriate categories to improve search accuracy.

Further, according to the first embodiment described above, the classifier 200 is provided with product data to which information (end label) indicating that the category is the lowest layer is added as learning data together with a category that is a correct answer as a teacher label. In order to learn, even if the number of layers up to the lowest layer category is different for each product, the recursive process is repeated up to the layer of the category with the end label, so it is not necessary to adjust the hyperparameter of the number of recursive processes of RNN250. , Scores can be obtained while dynamically changing the number of recursive processes according to the type of each category, and even if the depth of the category hierarchy is not known in advance, the categories of each hierarchy up to the bottom layer are classified. can do.

Further, according to the first embodiment described above, in order to extract high-quality learning data by a predetermined sampling process, the classifier 200 is trained more accurately than in the case of randomly sampling the training data. be able to.

In general, a product that the seller has classified as "Other" is really correct as a "Other" category, whether it is a product that is misclassified even though other categories are more suitable. It is difficult to determine whether it is a product. Therefore, even if the product data related to the products classified as "Other" is used as the learning data, the reliability of the teacher label becomes poor and it is difficult to improve the learning accuracy of the classifier 200. On the other hand, in the present embodiment, when extracting high-quality learning data, the product data related to the products classified as "other" is excluded from the learning data, so that the product data is currently classified as "other". , Products that should normally be classified in other categories can be classified in the correct category. In general, the number of products that can only be classified as "Other" is less than the total number [%], so even if products that can only be classified as "Other" cannot be correctly classified, Since the majority of other products can be classified into the correct category, the total accuracy of the classifier can be improved.

<Second Embodiment>
Hereinafter, the second embodiment will be described. In the first embodiment described above, it has been described that the classifier 200 includes an RNN 250 that performs recursive processing. On the other hand, the second embodiment differs from the first embodiment described above in that the classifier 200 does not include the RNN 250. Hereinafter, the differences from the first embodiment will be mainly described, and the points common to the first embodiment will be omitted. In the description of the second embodiment, the same parts as those of the first embodiment will be described with the same reference numerals.

FIG. 18 is a diagram showing an example of the classifier 200 in the second embodiment. The classifier 200 in the second embodiment is realized by, for example, the first CNN210, the second CNN220, the FCNN240, and the word embedding model 260.

The product data in the second embodiment includes, for example, content such as a product image, a product title, and a product outline, as well as a store ID representing identification information of each store that sells the product on the sales site. Is done.

The morphological analysis unit 114 in the second embodiment inputs the store ID included in the product data into the word embedding model 260. The word embedding model 260 is, for example, a neural network learned to convert a store ID into a vector by a word embedding expression when a store ID is input and output a matrix using the vector of the store ID as a row vector. is there.

The FCNN240 in the second embodiment combines and outputs the matrices output from each of the first CNN210, the second CNN220, and the word embedding model 260, for example.

For example, the category determination unit 116 in the second embodiment regards each element of the output vector of the FCNN 240 as a category score, and determines the category having the highest score as the product category.

According to the second embodiment described above, the word matrix generated from the product title is input to the first CNN210, the word matrix generated from the product outline is input to the second CNN220, and the matrix generated from the store ID is embedded in words. Input to model 260, integrate the vectors output by each of these models, and determine the product category of the product data, including the product title, product overview, and store ID, based on each element of the output vector. Therefore, as in the first embodiment, the products (including services) can be classified into more appropriate categories to improve the search accuracy of the products.

Further, according to the second embodiment described above, since the store ID is used for classifying the categories, the product categories can be determined based on the tendency of the types of products that are easy to sell for each store.

<Third Embodiment>
Hereinafter, the third embodiment will be described. The third embodiment differs from the first and second embodiments described above in that the classifier 200 includes, in addition to the RNN 250, a dedicated classifier that classifies the categories of the highest hierarchy. Hereinafter, the differences from the first and second embodiments will be mainly described, and the points common to the first and second embodiments will be omitted. In the description of the third embodiment, the same parts as those of the first and second embodiments will be described with the same reference numerals.

FIG. 19 is a diagram showing an example of the classifier 200 according to the third embodiment. The classifier 200 in the third embodiment is realized by, for example, a first CNN210, a second CNN220, a third CNN230, an FCNN240, a plurality of RNN250s, and a top category classifier 270.

The top-level category classifier 270 is, for example, a neural network pre-learned to output a vector having a score of the category of the top-level hierarchy as an element when a vector is input from FCNN240. The vector output by the top-level category classifier 270 may be a vector in which only one element is 1 and the other elements are 0, as in the one-hot expression, for example.

Each of the plurality of RNN 240s is prepared in the same number as the number of categories in the uppermost layer (the same number as the number of dimensions of the output vector of the uppermost category classifier 270), and each RNN 250 is a category in the second and subsequent layers. The correct answer of is learned in advance based on the product data given as the teacher label. As a result, each RNN 250 classifies the categories of the second and subsequent layers excluding the highest layer, so that one RNN 250 classifies the categories of each layer from the highest level to the lowest level. The categories of each layer can be classified more accurately.

According to the third embodiment described above, after classifying the categories of the highest hierarchy by the dedicated top category classifier 270 for classifying the categories of the highest hierarchy, each category of the highest hierarchy is specified. Since the categories of the second and subsequent hierarchies excluding the highest hierarchy are classified by the RNN250 learned by converting, the products can be classified into more appropriate categories, and as a result, the search accuracy of the products can be improved. Can be done.

<Fourth Embodiment>
Hereinafter, the fourth embodiment will be described. In the first to third embodiments described above, the RNN 250 has described that the intermediate layer of the network is an LSTM. On the other hand, in the fourth embodiment, the RNN 250 is a neural network (hereinafter referred to as CONVLSTM) in which a convolutional neural network and an RNN whose intermediate layer of the network is LSTM are combined. It is different from the third embodiment. Hereinafter, the differences from the first to third embodiments will be mainly described, and the points common to the first to third embodiments will be omitted. In the description of the fourth embodiment, the same parts as those of the first to third embodiments will be described with the same reference numerals.

In the fourth embodiment, for example, various neural networks provided in the upper stage of the classifier 200 can be obtained by omitting the pooling process for each of the first CNN210, the second CNN220, and the third CNN230, or omitting the FCNN240. Then, the third-order tensor is input to the RNN250.

The RNN 250 in the fourth embodiment is, for example, a neural network (hereinafter referred to as CONVLSTM) in which a convolutional neural network and an RNN whose intermediate layer of the network is an LSTM are combined, and when a third-order tensor is input. , It is a neural network learned to output the score indicating the certainty (probability) of the title and outline of the tensor, and the category to which the product corresponding to the image is classified, for each layer. In the CONVLSTM, the inner product calculation of the weight and the state variable is changed to the convolution calculation in the above-mentioned calculation formula of each gate of the LSTM. Formula (2) represents the formula for calculating the gates and neurons of CONVLSTM.

_{X t} is in the fourth embodiment, the third floor tensor inputted in the processing cycle t (hereinafter, input tensor) represents, _{h t} is 3 order tensor output by RNN250 in the processing cycle t (hereinafter, the output tensor) Represents.

For example, RNN250 in the fourth embodiment, the convolution values the weight _{W i} and the input tensor _{x t,} the convolution values the output tensor _{h t-1} weight _{R i} and the previous weights _{Q i} and the previous internal operation tensor and (3 tensor) Hadamard product of c _t-1 (a product of each element), by solving the sigmoid function σ that the sum of the bias component b _i as a variable, an input gate tensor processing cycle t (3 to derive the floor tensor) _{i t.}

Further, the RNN 250 has _{a convolution value of the weight W f} and the input tensor x _t , a convolution value of the weight R _f and the previous output vector h _t-1 , and a weight Q _f and the previous internal calculation tensor c _t-1. By solving the sigmoid function σ with the sum of the Hadamard product of and the bias component b _f as a variable, the forget gate tensor ft of the processing period _t is derived.

Further, RNN250 includes a Hadamard product of Forget gate tensor _{f t} and tensor _{z t} of the processing cycle t, Hadamard the forget gate tensor _{f t} of the internal processing cycle t-1 arithmetic tensor _{c t-1} and the processing period t the sum of the product, is derived as an internal operation tensor c _t processing cycle t.

z _t is a hyperpolic with the sum of the convolution value of the weight W _z and the input tensor x _t , the convolution value of the weight R _z and the previous output tensor h _t-1 , and the bias component b _{z as variables.} It is derived by solving the tangent function tanh.

Further, RNN250 is of the convolution values the weight _{W o} and the input tensor _{x t,} the convolution values the output tensor _{h t-1} weight _{R o} before, the weight _{Q o} and the current internal operation tensor _{c t} and Hadamard product, by solving the sigmoid function σ that the sum of the bias component b _o as variables, to derive the output gate tensor (3-order tensor) o _t of the processing cycle t.

RNN250, when deriving the output gate tensor o _t, the output tensor outputs and gate tensor o _t, the Hadamard product of the solution of the hyper Po Rick tangent function tanh with the internal operation tensor c _t as a variable, processing period t derived as h _t.

Each element value of the output tensor h _t represents the score of the category for each layer, as in the output vector of the above-described embodiment. The RNN 250 repeatedly derives the output tensor by repeating the calculation based on the above-mentioned mathematical formula.

According to the fourth embodiment described above, since the RNN 250 of the classifier 200 is an RNN whose intermediate layer is CONVLSTM, the goods (including services) are compared with the first to third embodiments described above. It can be further classified into appropriate categories, and as a result, the search accuracy can be further improved.

<Hardware configuration>
The information processing device 100 of the above-described embodiment is realized by, for example, a hardware configuration as shown in FIG. FIG. 20 is a diagram showing an example of the hardware configuration of the information processing apparatus 100 of the embodiment.

The information processing device 100 includes NIC100-1, CPU100-2, RAM100-3, ROM100-4, secondary storage devices 100-5 such as flash memory and HDD, and drive devices 100-6, which are internal buses or dedicated communication lines. The configuration is interconnected by. A portable storage medium such as an optical disk is mounted on the drive device 100-6. A program stored in a portable storage medium mounted on the secondary storage device 100-5 or the drive device 100-6 is expanded into the RAM 100-3 by a DMA controller (not shown) or the like, and executed by the CPU 100-2. As a result, the control unit 110 is realized. The program referred to by the control unit 110 may be downloaded from another device via the network NW.

Although the embodiments for carrying out the present invention have been described above using the embodiments, the present invention is not limited to these embodiments, and various modifications and substitutions are made without departing from the gist of the present invention. Can be added.

1 ... Information processing system, 10 ... Terminal device, 20 ... Service providing device, 100 ... Information processing device, 102 ... Communication unit, 110 ... Control unit, 112 ... Acquisition unit, 114 ... Morphological analysis unit, 116 ... Category determination unit, 118 ... list generation unit, 120 ... information providing unit, 122 ... learning processing unit, 130 ... storage unit, 200 ... classifier

Claims (15)

An acquisition department that acquires product data related to products or services sold via the network,
A classifier trained to output the classification result of a hierarchical category for the product or service indicated by the input product data, using the product data of the product or service classified into the hierarchical category as training data. Based on the output result, a determination unit that determines the category of the product or service indicated by the product data acquired by the acquisition unit, and
With
When the input product data contains a plurality of contents of different types, the classifier classifies the product or service of the input product data based on the characteristics of each of the plurality of contents. A score representing the certainty of the category to be is output, and the input product data, the feature derived when the product data was input in the past, and the input product data based on the score are output. Contains a recursive neural network that repeats a recursive process that outputs a score for a product or service category.
The recurrent neural network outputs a score of a category having a deeper hierarchy each time the recursive processing is repeated, and derives a score for each of a plurality of categories for each processing cycle of the recursive processing. Based on the highest score derived for each category, the score for the next processing cycle is derived.
The decision unit
Based on the score output by the classifier, the category of the product or service indicated by the product data acquired by the acquisition unit is determined.
When the recursive neural network outputs the scores of a plurality of categories in a certain processing cycle, it is determined whether or not some or all of the scores of the plurality of categories are below the threshold value. And
When it is determined that a part or all of the scores of the plurality of categories is less than the threshold value, the plurality of categories output by the recurrent neural network in the processing cycle before the processing cycle. Of the respective scores, the score based on the second and subsequent largest scores is derived by the recurrent neural network as the score in the processing cycle.
Information processing device. The product data includes at least two or more contents of the title, summary, and image of the product or service.
The information processing device according to claim 1. The recurrent neural network repeats the recursive process as many times as the number of layers of the product or service category indicated by the product data given as the training data at the time of training.
The information processing device according to claim 1 or 2. An extraction unit that extracts products or services whose categories determined by the determination unit and the categories preset by the seller are different from the products or services sold via the network.
The terminal device of the seller who sells the goods or services extracted by the extraction unit is further provided with an information providing unit that provides information prompting the seller to change the category.
The information processing device according to any one of claims 1 to 3. An extraction unit that extracts products or services whose categories determined by the determination unit and the categories preset by the seller are different from the products or services sold via the network.
A search processing unit that lowers the search ranking of the product or service extracted by the extraction unit is further provided.
The information processing device according to any one of claims 1 to 4. When the seller sets a category at the time of selling a product or service, if the category set by the seller and the category determined by the determination unit are different, the category set by the seller is selected as the determination unit. Further provided with an information providing unit that provides the terminal device of the seller with information prompting the change to the category determined by the seller.
The information processing device according to any one of claims 1 to 5. Further provided with an information providing unit that presents the category determined by the determination unit as the category of the product or service to be sold at the time of selling the product or service.
The information processing device according to any one of claims 1 to 6. The classifier is trained so as to output the classification result of the hierarchical category for the product or service indicated by the input product data using the product data of the product or service classified into the hierarchical category as training data. Further equipped with a learning processing unit to make
The information processing device according to any one of claims 1 to 7. The learning processing unit trains the classifier using the product data related to the product or service that the user has reached a predetermined action as the learning data among the plurality of product data that are candidates for the learning data.
The information processing device according to claim 8. The prescribed actions include purchasing goods or services.
The learning processing unit learns the classifier using product data related to products or services purchased by the user more than a predetermined number of times as the learning data.
The information processing device according to claim 9. The prescribed actions include browsing goods or services.
The learning processing unit learns the classifier using product data related to products or services that have been viewed by the user a predetermined number of times or more as the learning data.
The information processing device according to claim 9 or 10. The predetermined action includes the time when the product or service is browsed for a predetermined time or more.
The learning processing unit learns the classifier using product data related to products or services that have been viewed by the user for a predetermined time or longer as the learning data.
The information processing device according to any one of claims 9 to 11. Learning to learn the classifier using the plurality of product data excluding the product data of products or services classified into a predetermined category by the seller from the plurality of product data candidates for the training data as the training data. Further equipped with a processing unit,
The information processing device according to any one of claims 1 to 12. The computer
Get product data about products or services sold over the network
A classifier trained to output the classification result of a hierarchical category for the product or service indicated by the input product data, using the product data of the product or service classified into the hierarchical category as training data. Based on the output result, the category of the product or service indicated by the acquired product data is determined.
When the input product data contains a plurality of contents of different types, the classifier classifies the product or service of the input product data based on the characteristics of each of the plurality of contents. A score representing the certainty of the category to be is output, and the input product data, the feature derived when the product data was input in the past, and the input product data based on the score are output. Contains a recursive neural network that repeats a recursive process that outputs a score for a product or service category.
The recurrent neural network outputs a score of a category having a deeper hierarchy each time the recursive processing is repeated, and derives a score for each of a plurality of categories for each processing cycle of the recursive processing. Based on the highest score derived for each category, the score for the next processing cycle is derived.
In making the above decision
Based on the score output by the classifier, the category of the product or service indicated by the acquired product data is determined.
When the recursive neural network outputs the scores of a plurality of categories in a certain processing cycle, it is determined whether or not some or all of the scores of the plurality of categories are below the threshold value. And
When it is determined that a part or all of the scores of the plurality of categories is less than the threshold value, the plurality of categories output by the recurrent neural network in the processing cycle before the processing cycle. Of the respective scores, the score based on the second and subsequent largest scores is derived by the recurrent neural network as the score in the processing cycle.
Information processing method. On the computer
The process of retrieving product data about products or services sold over the network,
A classifier trained to output the classification result of a hierarchical category for the product or service indicated by the input product data, using the product data of the product or service classified into the hierarchical category as training data. Based on the output result, the process of determining the category of the product or service indicated by the acquired product data, and
It is a program to execute
When the input product data contains a plurality of contents of different types, the classifier classifies the product or service of the input product data based on the characteristics of each of the plurality of contents. A score representing the certainty of the category to be is output, and the input product data, the feature derived when the product data was input in the past, and the input product data based on the score are output. Contains a recursive neural network that repeats a recursive process that outputs a score for a product or service category.
The recurrent neural network outputs a score of a category having a deeper hierarchy each time the recursive processing is repeated, and derives a score for each of a plurality of categories for each processing cycle of the recursive processing. Based on the highest score derived for each category, the score for the next processing cycle is derived.
In the process of determining, the computer
Based on the score output by the classifier, the category of the product or service indicated by the acquired product data is determined.
When the recursive neural network outputs the scores of a plurality of categories in a certain processing cycle, it is determined whether or not some or all of the scores of the plurality of categories are below the threshold value. Let me
When it is determined that a part or all of the scores of the plurality of categories is less than the threshold value, the plurality of categories output by the recurrent neural network in the processing cycle before the processing cycle. Among the respective scores, the score based on the second and subsequent largest scores is derived by the recurrent neural network as the score in the processing cycle.
program.

Images