Information processing device, information processing method, and recording medium

Abstract

To efficiently execute a classification task targetting enormous kinds of tags.SOLUTION: An information processing device includes a processor and a memory. The processor is configured to: reference a list indicative of correspondence between a genre of products in electronic commerce and tags to be attached to the products, thereby determining one or more tags not corresponding to the genre; and perform a process of training a machine learning model including at least one classifiers that each determine an output value of each of the one or more tags related to the products on the basis of an embedded expression in product information on the products, the training process being performed on the basis of a loss function of the one or more output values excluding the output value related to each of the one or more tags not corresponding to the genre of the products.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to an information processing device, an information processing method, and a recording medium. [Background technology]

[0002] A technology aimed at adding annotations to input text information related to a product is disclosed in Patent Document 1. In Patent Document 1, the user is prompted to specify whether the extraction of the character string (tag information) to which the annotation is to be added is appropriate or not, and machine learning is performed based on the result of the specification. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2020-115287 A Summary of the Invention [Problem to be solved by the invention]

[0004] The number of products that are the subject of electronic commerce is increasing and becoming more diverse, and the number of types of tags that need to be attached to products is also becoming enormous. For this reason, it is necessary to efficiently execute classification tasks that target a huge number of types of tags, for example, in the process of determining tags to be attached to products based on text and images (including video images) that describe the products that are the subject of electronic commerce. [Means for solving the problem]

[0005] One aspect of the present invention that solves the problems of the conventional examples is an information processing device including a processor and a memory, in which the processor determines one or more tags that do not correspond to the genre by referring to a list showing the correspondence between a product genre in electronic commerce and tags to be assigned to the product, and performs a learning process of a machine learning model that includes at least a classifier that determines an output value for each of the one or more tags associated with the product based on an embedded representation of product information of the product, based on a loss function of the one or more output values ​​excluding the output value for each of the one or more tags that do not correspond to the product genre. [Brief description of the drawings]

[0006] [Figure 1] 1 is a block diagram illustrating an example of a configuration of an information processing device according to an embodiment of the present invention. [Diagram 2] 1 is a functional block diagram illustrating an example of an information processing device according to an embodiment of the present invention. [Diagram 3] 5 is an explanatory diagram illustrating an example of the contents of a correspondence relationship list used in the information processing device according to the embodiment of the present invention. FIG. [Figure 4] 1 is a functional block diagram of an example of an information processing device according to an embodiment of the present invention; [Diagram 5] FIG. 4 is a flowchart illustrating an example of the operation of the information processing device according to the embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0007] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described with reference to the drawings. In the following description, the term "embedded representation" refers to tensor information (a set of numerical values) corresponding to input data such as a word or an image.

[0008] An information processing device 1 according to an embodiment of the present invention includes a processor 11, a memory 12, and an input / output unit 13, as shown in FIG.

[0009] The processor 11 includes at least one program control device such as a CPU. The processor 11 may include a GPU (Graphic Processing Unit) or other processor unit, or may include a plurality of CPUs or a combination of CPUs and GPUs. The processor 11 operates according to a program stored in the memory 12. In the example of the present embodiment, the processor 11 executes a process of determining tags related to a product that is a target of electronic commerce. That is, the processor 11 determines one or more tags that do not correspond to the product genre by referring to a list indicating the correspondence between the product genre in electronic commerce and the tags to be assigned to the product. The processor 11 also executes a learning process of a machine learning model that includes at least a classifier that determines an output value related to each of one or more tags related to the product based on a predetermined embedded expression related to the product information of the product. In this learning process, the processor 11 executes the learning process based on a loss function of one or more output values ​​excluding the output value related to each of one or more tags that do not correspond to the product genre. The detailed operation of the processor 11 will be described later.

[0010] The memory 12 is a storage element, a disk device, or the like, and stores a program executed by the processor 11. This program may be provided by being stored in a computer-readable and non-transitory recording medium, and may be copied into the memory 12.

[0011] The input / output unit 13 includes, for example, a USB (Universal Serial Bus) interface. The input / output unit 13 is connected to a keyboard, a mouse, and the like, and receives information such as text input by a user. The input / output unit 13 may also include a network interface, and receive various types of information such as text information, image information, and even audio information, which are product information, from other information processing devices. The input / output unit 13 may also be connected to a display, and output information to the display according to instructions input from the processor 11.

[0012] Next, a description will be given of the operation of the processor 11. The processor 11 executes a program stored in the memory 12, thereby achieving a configuration that functionally includes a learning processing unit 110 and an inference processing unit 210, as shown in FIG.

[0013] Here, the learning processing unit 110 includes an input receiving unit 111, a model learning unit 112, an estimated tag output unit 113, a correspondence list acquiring unit 114, a masking unit 115, and a loss calculating unit .

[0014] The inference processing unit 210 includes an input receiving unit 211 , an inference processing unit 212 , and an estimated tag output unit 213 .

[0015] The input receiving unit 111 of the learning processing unit 110 receives information on the genre of a product in electronic commerce, information related to the product (hereinafter referred to as product information), and information that identifies at least one tag that is the correct answer (hereinafter referred to as correct answer information). Here, the product information may be text information such as a product name and product description, or may be video / audio information such as an image of the product, a video explaining the product, or a sound explaining the product.

[0016] The model learning unit 112 executes machine learning processing of a machine learning model that is a target of the machine learning processing. In the example of the present embodiment, the machine learning model that is machine learned by the model learning unit 112 uses a Transformer network, and specifically, is BERT (Bidirectional Encoder Representations from Transformers; J. Devlin, et al., BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, arXiv:1810.04805v2).

[0017] Here, it is assumed that BERT has undergone a predetermined machine learning process in advance and is capable of outputting an embedded expression of product information based on the product information.

[0018] However, the machine learning model used in the information processing device 1 of the present embodiment does not have to be BERT. The machine learning model used by the information processing device 1 of the present embodiment may be another one as long as it is machine-learned so as to be capable of outputting an embedded expression of product information and includes at least one classifier that determines a predetermined output value (probability of matching each tag to product information) for each of one or more tags related to a product based on the embedded expression.

[0019] For example, the machine learning models may include ELECTRA (https: / / arxiv.org / abs / 2003.10555) and ViT (Vision Transformer: https: / / openreview.net / forum?id=YicbFdNTTy).

[0020] The model learning unit 112 uses the product information accepted by the input acceptance unit 111 as input to the machine learning model, obtains an embedded representation of the product information, which is one of the outputs, and outputs it to the estimated tag output unit 113. When the machine learning model is BERT, a CLS token (classification token) may be used as the embedded representation here. However, this is only one example, and the embedded representation here may be an average of the embedded representations of each token (word) as long as it appropriately represents the input product information. In addition, based on information of a loss function output by a loss calculation unit 116 described later, the model learning unit 112 trains the machine learning model to input product information and output a matching probability for each tag for the input product information. This machine learning process is performed by updating parameter information included in the machine learning model by, for example, backpropagation.

[0021] The estimated tag output unit 113 outputs, as an output value, a matching probability (probability that each tag is determined to be assigned) for the input product information for each preset tag based on the embedded representation of the product information output by the model learning unit 112. Specifically, the estimated tag output unit 113 uses a neural network and an activation function such as a sigmoid function included in the machine learning model to calculate the matching probability of tags close to the embedded representation generated by the machine learning model. When the machine learning model includes a feedforward network as a classifier, this calculation is performed by using the feedforward network.

[0022] When the machine learning model includes a fully connected layer instead of a feedforward network, the estimated tag output unit 113 inputs the embedded representation of the product information output by the model learning unit 112 to the fully connected layer, and obtains the matching probability (the probability that each tag should be assigned) for the input product information. In this example, the estimated tag output unit 113 can determine the output value (matching probability) without applying an activation function.

[0023] The correspondence list acquisition unit 114 acquires a correspondence list which is a list showing the correspondence between the genre of a product in electronic commerce and the tag to be assigned to the product. In one example of the present embodiment, a correspondence list (L) in which the genre (G) of a product is associated with the tag (T) to be assigned is prepared in advance as shown in FIG. 3 and stored in, for example, the memory 12. As an example, this correspondence list records information such as colors such as red, blue, and green and sizes such as S, M, and L as tags in association with the product genre "clothing". For this product genre of "clothing", tags such as TV screen sizes such as 32 inches, 40 inches, and 43 inches and drinking water volumes such as 350 mL, 500 mL, and 1.5 L are incompatible tags.

[0024] In an example in which such a correspondence list is used, the correspondence list acquisition unit 114 acquires tag information associated with information on the genre of the product accepted by the input acceptance unit 111 from the correspondence list.

[0025] The masking unit 115 outputs to the loss calculation unit 116 information on the matching probability, excluding the matching probability for tags other than the tags specified by the information acquired by the correspondence list acquisition unit 114, out of the matching probability for each tag calculated by the estimated tag output unit 113. In other words, this masking unit 115 sets the output value (tag matching probability) of the estimated tag output unit 113 for tags (mismatching tags) that do not correspond to the input product genre, to "0" (the result of applying a large negative value to the activation function). In this way, the masking unit 115 selectively outputs to the loss calculation unit 116 output values ​​(generally values ​​greater than "0") for one or more tags not included in the mismatching tags, excluding the output value for the mismatching tags.

[0026] The loss calculation unit 116 applies a predetermined activation function (for example, a nonlinear function such as a sigmoid function or a softmax function) to the output value for each tag output by the masking unit 115. The loss calculation unit 116 then calculates a loss function using the value obtained by applying the activation function and the correct answer information accepted by the input acceptance unit 111. Note that the loss function calculated by the loss calculation unit 116 here may be a sum of squares error or a cross-entropy error, and an appropriate loss function is adopted depending on the task settings. The value of the loss function calculated by the loss calculation unit 116 is provided to the machine learning process of the machine learning model in the model learning unit 112.

[0027] The model learning unit 112 uses the value of this loss function to fine-tune the machine learning model.

[0028] The processor 11 also operates as an inference processing unit 210 in response to an instruction from a user. An input receiving unit 211 of the inference processing unit 210 receives input of product information and outputs it to an inference processing unit 212.

[0029] The inference processing unit 212 inputs the product information accepted by the input accepting unit 211 to the machine learning model trained by the learning processing unit 110. The inference processing unit 212 obtains information on the matching probability of each tag to the product information, which is output by the machine learning model.

[0030] When the machine learning model includes a feedforward network, the estimated tag output unit 213 uses information on the matching probability for each tag obtained by the inference processing unit 212 by applying the output of the feedforward network to an activation function to identify a tag related to the input product information, and outputs information identifying the tag. As an example, the estimated tag output unit 213 refers to information on the matching probability for each tag obtained by the inference processing unit 212, and outputs information identifying a tag whose matching probability exceeds a predetermined value. The number of tags identified by the information output here may be one or more. Furthermore, when the matching probability of all tags is below a predetermined value, the number of tags identified may be zero.

[0031] As already described, the inference processing unit 212 may use a fully connected layer instead of a feedforward network to obtain the matching probability of each tag (the probability that each tag is determined to be assigned).

[0032] [Operation] The information processing device 1 of this embodiment basically has the above configuration and operates as in the following example. Note that in the following example, the machine learning model used by the information processing device 1 will be described as BERT.

[0033] 4, the information processing device 1 functionally includes a model learning unit 112, a masking unit 115, a nonlinear function unit 1161, and a tag master 200 which is a correspondence relationship list. The model learning unit 112 also includes a machine learning model 1121, a CLS token 1122 which is the output thereof, tokens 1123a, 1123b, . . . relating to words, and a network unit 1124.

[0034] First, the user performs machine learning processing (fine tuning) on ​​the machine learning model of the information processing device 1. Specifically, the user inputs data for machine learning to the information processing device 1, which data is a set of information on the genre of a plurality of products in electronic commerce, product information, and corresponding correct answer information.

[0035] 5, the processor 11 of the information processing device 1 sequentially accepts information on the genre of each product, the product information, and the corresponding correct answer information contained in the input data (S11). Then, the processor 11, for example, uses the accepted product information as an input to BERT, which is a machine learning model 1121, and obtains the output, CLS token 1122, as an embedded representation of the product information (S12).

[0036] Processor 11 inputs the embedded representation obtained in step S12 to network unit 1124, which is a feedforward network, and obtains as its output a vector of matching probabilities (probabilities that each tag should be determined to be assigned) for the input product information for each preset tag (S13).

[0037] On the other hand, the processor 11 acquires, from the received information, tag information that is listed in advance as tags to be attached to products of the genre represented by the information, using the genre information (S14). In one example of the present embodiment, the processor 11 acquires tag information corresponding to the received genre information, by referring to the tag master 200, which is a correspondence list that holds the product genres and the tags to be attached in association with each other, as already described.

[0038] Processor 11 removes the matching probability of non-matching tags other than the tag represented by the information acquired in step S14 from the components of the vector obtained in step S13 by masking them using masking unit 115 (S15). However, processor 11 outputs the value of "0" for components whose value is "0". Processor 11 calculates the value of a loss function such as a square sum error or cross entropy error between the matching probability of each tag and information in which the component corresponding to the tag of the correct answer information input in step S11 is set to "1" and the component corresponding to the tag not included in the correct answer information is set to "0" (S16) using nonlinear function unit 1161. At this time, processor 11 ignores the matching probability of the masked tag during the calculation (it does not obtain the difference from the correct answer, and does not accumulate).

[0039] The processor 11 further performs machine learning processing (fine tuning) of the machine learning model based on the value of this loss function, and updates each parameter in the machine learning model (S17).

[0040] Processor 11 then repeatedly executes the processes of steps S11 to S17 for each product included in the input data, and when the processes for the products related to the input data are completed, ends this machine learning process.

[0041] [Example of batch processing] Note that, although an example of repeatedly performing machine learning processing for each product has been shown here, the processing may also be performed by so-called batch processing, in which a single update is performed based on data (batch) relating to multiple products.

[0042] In this case, in step S13, processor 11 obtains a tensor in which vectors corresponding to each product included in the batch are arranged. In step S14, processor 11 obtains tag information corresponding to the genre of the products included in the batch. In step S15, processor 11 masks, by setting to "0", components of vectors included in the tensor obtained in step S13, corresponding to each product in the batch, that correspond to incompatible tags that are not included in the tag information obtained in step S14 corresponding to the genre of the product. Hereinafter, a widely known method for batch processing can be used for the calculation of the loss function.

[0043] [Create correspondence list] Furthermore, the correspondence list used by the processor 11 in the above example may be generated based on the relationship between tags previously assigned to a product that is the subject of electronic commerce and the genre of the product.

[0044] For example, the processor 11 generates a correspondence list using data that associates information on the genre of a product that was the subject of electronic commerce in the past with information on tags attached to the product. As an example, the processor 11 finds, based on the data, information on tags that have been attached to products of the genre more than a predetermined threshold number of times. The processor 11 then associates the found tag information with the genre of the product and includes it in the correspondence list.

[0045] [Other examples of masking] In the above description, it has been assumed that the processor 11 prepares a correspondence list that associates product genres with tags to be added in order to obtain information on tags to be masked (information on non-compliant tags). However, the present embodiment is not limited to the example in which this correspondence list is used.

[0046] For example, the processor 11 may cluster and classify pre-listed tag candidates by product genre, and generate a list that associates product genres with tags to be assigned.

[0047] In this example, the processor 11 obtains a vector representation for each tag. For example, this vector representation may be set as follows. First, a permutation G1, G2, ... of product genres is set, and for tags previously assigned to products, a vector representation is obtained by setting the number of times the tag has been assigned to products of the genre Gi as the value of the i-th component corresponding to the genre Gi. For example, the permutation of genres is set as "clothing", "shoes", "bags", ..., and a vector representation Vj of a certain tag Tj (j=1, 2, ...) is set as T1j, the number of times the tag Tj has been assigned to products of the "clothing" category, T2j, the number of times the tag Tj has been assigned to products of the "shoes" category, ..., to obtain a vector representation Vj=(T1j, T2j, T3j, ...).

[0048] The processor 11 divides the vector representation obtained for each tag into a plurality of clusters by a predetermined clustering process such as the k-means method. The processor 11 then associates information on the genre of the product with each cluster. This association may be performed artificially by setting genre information for each cluster while referring to the tags belonging to the cluster. The processor 11 may also include vector representations for genre information (in the above example, for example, the vector representation of genre Gi may be set as a one-hot vector with only the i-th component set to "1" and the other components set to "0") as targets for clustering processing, and perform clustering processing together with the vector representations of tags. In this case, each cluster is associated with genre information included in the cluster. If genre information is not included in a cluster, the cluster may be associated with genre information having a vector representation closest to the center of the cluster.

[0049] The processor 11 associates the information of the tag that belongs to each cluster with the information of the genre associated with the cluster. This sets the tag corresponding to the genre information. The processor 11 may use the information of the correspondence relationship of the tag corresponding to the genre information obtained by this method instead of a correspondence relationship list prepared in advance, or may record and use the information of the correspondence relationship obtained by this method as a correspondence relationship list.

[0050] [Tag inference behavior] In addition, the information processing device 1 of this embodiment uses the machine learning model that has been machine-learned through the above-mentioned processing to select a tag to be assigned to a product as follows, and outputs information identifying the selected tag.

[0051] The user inputs product information of a product to be tagged to the information processing device 1. As already described, this product information may be text information such as a product name and product description, or may be video / audio information such as an image of the product, a video explaining the product, or audio explaining the product, and is assumed to be information of the same type (text, image, video, audio, or a combination thereof, etc.) as that input during machine learning processing.

[0052] The processor 11 of the information processing device 1 uses the input product information as an input to the machine learning model, and obtains an embedded representation of the product information, which is one of the outputs. Specifically, when the machine learning model includes BERT, a CLS token (classification token) is used as the embedded representation.

[0053] Processor 11 inputs the obtained embedded representation into a feedforward network that constitutes the machine learning model, and obtains as its output a vector of matching probabilities (probabilities that each tag should be assigned) for the input product information for each pre-set tag.

[0054] When the machine learning model includes a fully connected layer instead of a feedforward network, the processor 11 may input the obtained embedded representation to the fully connected layer and obtain, as the output value, a vector of matching probability (probability that each tag is determined to be assigned) for the input product information for each preset tag. At this time, the processor 11 converts the number of dimensions of the embedded representation to a number of dimensions according to the number of tags via the fully connected layer. Note that when the fully connected layer is used instead of the feedforward network, the processor 11 determines the output value (matching probability) without applying an activation function.

[0055] Then, the processor 11 refers to the matching probability information for each tag, and outputs information that identifies a tag whose matching probability exceeds a predetermined value as information of the tag to be assigned. The tag identified by the output information here may be only one tag or may be multiple tags.

[0056] Furthermore, when a tag specified by the information output as information on a tag to be assigned is assigned to a product, the processor 11 may receive information to that effect, associate the genre of the product with the assigned tag, include the tag in the correspondence list, and update the correspondence list. In this way, the correspondence list is updated every time a tag to be assigned to a product is determined. [Explanation of symbols]

[0057] 1 Information processing device, 11 Processor, 12 Memory, 13 Input / output unit, 110 Learning processing unit, 111 Input acceptance unit, 112 Model learning unit, 113 Estimated tag output unit, 114 Correspondence list acquisition unit, 115 Masking unit, 116 Loss calculation unit, 200 Tag master, 210 Inference processing unit, 211 Input acceptance unit, 212 Inference processing unit, 213 Estimated tag output unit, 1121 Machine learning model, 1124 Network unit, 1161 Nonlinear function unit.

Claims

1. An information processing device including a processor and a memory, The processor, A vector representation is obtained for each of the pre-listed tag candidates, the vector representation obtained for each tag is classified into a plurality of clusters, and each cluster is associated with a product genre in electronic commerce, thereby associating the product genre with the tag to be assigned; A learning process of a machine learning model including at least a classifier that determines an output value for each of the one or more tags related to the product based on an embedded representation of the product information of the product is performed based on a loss function of the one or more output values ​​excluding the output value for each of the one or more tags that does not correspond to the genre of the input product. Information processing device.

2. An information processing device including a processor and a memory, The processor, A vector representation is obtained for each of the pre-listed tag candidates, the vector representation obtained for each tag is classified into a plurality of clusters, and each cluster is associated with a product genre in electronic commerce, thereby associating the product genre with the tag to be assigned; A learning process of a machine learning model including at least a classifier that determines an output value for each of the one or more tags related to the product based on an embedded representation of product information of the product is performed based on a loss function of the one or more output values ​​excluding the output value for each of the one or more tags that are not classified into a cluster corresponding to a predetermined product genre. Information processing device.

3. 3. The information processing device according to claim 1, The information processing device wherein the product information is text information.

4. 3. The information processing device according to claim 1, When the processor obtains a vector representation for each of the tag candidates, the information processing device sets the vector representation for tags that have been previously assigned to products for each product genre Gi (i = 1, 2, ...) with the number of times that the tag has been assigned as the i-th component.

5. 3. The information processing device according to claim 1, When the processor associates the product genre with the tag to be assigned, the information processing device obtains a vector representation for each of the pre-listed tag candidates and the product genre in e-commerce, classifies the vector representations of the tags and product genre into multiple clusters, and associates tags belonging to a common cluster with the product genre, thereby associating the product genre with the tag to be assigned.

6. Using a processor, A vector representation is obtained for each of the pre-listed tag candidates, the vector representation obtained for each tag is classified into a plurality of clusters, and each cluster is associated with a product genre in electronic commerce, thereby associating the product genre with the tag to be assigned; A learning process of a machine learning model including at least a classifier that determines an output value for each of the one or more tags related to the product based on an embedded representation of the product information of the product is performed based on a loss function of the one or more output values ​​excluding the output value for each of the one or more tags that does not correspond to the genre of the input product. Information processing methods.

7. The processor, a means for associating a product genre with a tag to be assigned by obtaining a vector representation for each of pre-listed tag candidates, classifying the vector representation obtained for each tag into a plurality of clusters, and associating each cluster with a product genre in electronic commerce; A learning process of a machine learning model including at least a classifier that determines an output value for each of the one or more tags related to the product based on an embedded representation of product information of the product, based on a loss function of the one or more output values ​​excluding the output value for each of the one or more tags that does not correspond to the genre of the input product; A program that functions as a

Images