JP7660845B2 - Attribute value extraction system, attribute value extraction method, and program - Google Patents

Description

This disclosure relates to an attribute value extraction system, an attribute value extraction method, and a program.

Conventionally, there is known a technique for assigning attribute values to items such as products or content using a machine learning technique. Non-Patent Document 1 and Non-Patent Document 2 describe a technique in which second item data related to a second item and a second query, which is an attribute related to the second item, are input to a model equivalent to an attribute value extraction model that has learned the relationship between first item data and a first query related to a first item and an attribute value related to the first item, and information corresponding to a named entity such as a second attribute value output from the model equivalent to the attribute value extraction model is assigned to the second item.

Qifan Wang, Li Yang, Bhargav Kanagal, Sumit Sanghai, D.Sivakumar, Bin Shu, Zac Yu, and Jon Elsas. 2020. Learning to extract attribute value from product via question answering: A multi-task approach. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 47-55, Online. ACM. Xiaoya Li, Jingrong Feng, Yuxian Meng, Qinghong Han, Fei Wu, and Jiwei Li. 2020. A unified MRC framework for named entity recognition. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5849-5859, Online. Association for Computational Linguistics.

However, in the techniques of Non-Patent Document 1 and Non-Patent Document 2, for example, the accuracy of the second attribute value output from a model equivalent to the attribute value extraction model may decrease depending on the query. For example, the query input to the attribute value extraction model was originally expected to be a relatively long interrogative sentence, but when a relatively simple word is input as the query, the accuracy of the second attribute value output from the model equivalent to the attribute value extraction model may decrease.

One of the objectives of this disclosure is to improve the accuracy of the output from the attribute value extraction model.

The attribute value extraction system according to the present disclosure includes a storage unit that stores an attribute value extraction model in which a relationship between first item data and a first query related to a first item and a first attribute value related to the first item is learned, a second attribute value acquisition unit that acquires a second attribute value associated with a second attribute related to a second item, and a third attribute value extraction unit that extracts a third attribute value from the second item data based on the second item data related to the second item, a second query corresponding to the second attribute value, and the attribute value extraction model.

This disclosure improves the accuracy of output from question answering models.

FIG. 1 illustrates an example of an overall configuration of an attribute value extraction system. 11A and 11B are diagrams illustrating an example of attributes and attribute values acquired from product data. FIG. 1 is a diagram illustrating an example of a conventional question answering model. FIG. 2 is a diagram illustrating an example of a question answering model in the embodiment. FIG. 2 is a functional block diagram showing an example of functions realized by the attribute value extraction system. FIG. 2 is a diagram illustrating an example of a training database. FIG. 4 is a diagram illustrating an example of an attribute database. FIG. 2 is a flow diagram illustrating an example of a process executed in the attribute value extraction system. FIG. 11 is a diagram illustrating an example of a process for calculating an average of feature amounts relating to each of a plurality of second attribute values. FIG. 11 is a diagram illustrating an example of a process for calculating an average of feature amounts relating to each of a plurality of second attribute values. FIG. 11 is a diagram illustrating an example of a process for calculating an average of feature amounts relating to each of a plurality of second attribute values.

[1. Overall configuration of the attribute value extraction system]
An example of an embodiment of an attribute value extraction system according to the present disclosure will be described. Fig. 1 is a diagram showing an example of the overall configuration of the attribute value extraction system. A network N is any network such as the Internet or a LAN. The attribute value extraction system S is not limited to the example of Fig. 1 as long as it includes at least one computer.

The server 10 is a server computer. The control unit 11 includes at least one processor. The storage unit 12 includes a volatile memory such as a RAM and a non-volatile memory such as a hard disk. The communication unit 13 includes at least one of a communication interface for wired communication and a communication interface for wireless communication.

The user terminal 20 is a computer operated by a user. For example, the user terminal 20 is a personal computer, a smartphone, a tablet terminal, or a wearable terminal. The physical configurations of the control unit 21, the memory unit 22, and the communication unit 23 are similar to those of the control unit 11, the memory unit 12, and the communication unit 13, respectively. The operation unit 24 is an input device such as a touch panel or a mouse. The display unit 25 is a liquid crystal display or an organic EL display.

The programs or data stored in the storage units 12 and 22 may be supplied via the network N. Each computer may also include at least one of a reading unit (e.g., a memory card slot) that reads a computer-readable information storage medium and an input/output unit (e.g., a USB port) for inputting and outputting data to and from an external device. For example, the programs or data stored in the information storage medium may be supplied via at least one of the reading unit and the input/output unit.

2. Overview of the Attribute Value Extraction System
In this embodiment, the process executed by the attribute value extraction system S will be described by taking as an example a case where an attribute value corresponding to an attribute of a product is acquired from product data related to the product. A product is an object that is the subject of commercial transaction. In this embodiment, a product in an online shopping mall is taken as an example, but the product itself may be any product, for example, a product bought and sold in an internet auction or a flea market app.

Product data is information about the details of a product. For example, product data includes letters, numbers, other symbols, images, audio, video, or a combination of these. In this embodiment, products in an online shopping mall are used as an example, so any text (product description) entered by a store staff member corresponds to the product data. For example, the product data can be viewed from the website or application of the online shopping mall. In this embodiment, a case is described in which a store staff member corresponds to the user, but the user may be anyone who uses the attribute value extraction system S, and is not limited to a store staff member. For example, the user may be an administrator of the online shopping mall.

Attributes are categories for classifying products. Attributes are sometimes called categories or genres. Attributes may be defined hierarchically. When attributes are defined hierarchically, the higher the attribute, the more abstract the meaning of the attribute, and the lower the attribute, the more specific the meaning of the attribute. Hereinafter, when simply describing an attribute, it means the name of the attribute. Attributes are expressed by letters, numbers, other symbols, or combinations of these. Attributes can be defined from any perspective as long as they can classify products from a specific perspective. For example, attributes are information such as brand, color, material, size, function, pattern, or place of origin.

An attribute value is a value that indicates a specific classification. Multiple attribute values are prepared in advance for an attribute. At least one of the multiple prepared attribute values is assigned to a product. The multiple prepared attribute values are candidates for the attribute values to be assigned to a product. Attribute values are expressed by letters, numbers, other symbols, or combinations of these. For example, when uploading product data for a product, a user specifies an attribute and at least one of the multiple attribute values prepared for that attribute. The specified attribute and attribute value are assigned to the product.

For example, various brand names are prepared as attribute values for the attribute "brand." Products of a certain brand are assigned an attribute value indicating the brand name of the brand of this product from among the multiple attribute values prepared for the attribute "brand." For example, various colors such as black, white, and yellow are prepared in advance as attribute values for the attribute "color." Products that look black are assigned an attribute value indicating "black" from among the multiple attribute values prepared for the attribute "color." The same is true for other attributes; any attribute value can be prepared in advance and assigned to products.

Figure 2 is a diagram showing an example of attributes and attribute values obtained from product data. For example, when a user operates the operation unit 24 to access a website of an online shopping mall, a product page P including product data is displayed on the display unit 25. For example, product data such as the store name, product title, product image, and detailed product description are displayed on the product page P. If the user specifies attributes and attribute values when uploading product data, the specified attributes and attribute values are also displayed as product data on the product page P. The product attributes and attribute values are used as indexes during searches.

If appropriate attribute values according to the attributes of a product can be obtained from product data, various effects can be achieved. For example, the burden on the user can be reduced because the user does not need to specify attribute values. For example, even if the user mistakenly specifies an inappropriate attribute value, it can be replaced with an appropriate attribute value obtained from the product data. For example, new attribute values can be obtained from the product data even if they are not currently available as attribute values. For example, it can be possible to gain a deeper understanding of user preferences or the marketplace and use this for marketing.

In the example of FIG. 2, if the attribute value "BBB bag" of the attribute "brand", the attribute value "black" of the attribute "color", the attribute value "leather" of the attribute "material", the attribute value "10''H x 13''L x 6''D" of the attribute "size", and the attribute value "waterproof" of the attribute "function" can be obtained from the product data, the various effects described above can be achieved. In this embodiment, a question answering model (QA model) is used as the attribute value extraction model to obtain these attribute values. In this embodiment, an embodiment in which a question answering model is used as the attribute value extraction model is illustrated, but there is no restriction on the type as long as it is a model that extracts attribute values by inputting entities such as product data and their attributes. In the following, a question answering model will be described as an example of an attribute value extraction model, so that in the following description, the description of the question answering model can be replaced with the description of the attribute value extraction model.

The question answering model is a model that uses machine learning. Various methods can be used for the machine learning itself. For example, the question answering model may be any model of supervised learning, semi-supervised learning, or unsupervised learning. In this embodiment, a question answering model that uses a model called BERT is given as an example, but the question answering model itself may be any model. For example, the question answering model may be a so-called Transformer-based model other than BERT, or may be a neural network model that was used before the advent of the Transformer.

Figure 3 is a diagram showing an example of a conventional question answering model. Below, an overview of conventional question answering models M1 and M2 will be explained. For details of question answering models M1 and M2, please refer to Non-Patent Document 1 and Non-Patent Document 2 listed as prior art documents, or the references described therein.

The question-answering model M1 is a model that, when a news article and a query are input, outputs a response corresponding to the query. The question-answering model M1 has learned training data that indicates the relationship between the news article and the query, and the response corresponding to the query. The query input to the question-answering model M1 is a question about the content of the news article. The response output from the question-answering model M1 is an answer to the question, which is the query.

The question-answering model M1 estimates which parts of a news article are appropriate as a response, and outputs the parts as a response. In the example of FIG. 3, a news article about the lifting of the state of emergency in Japan and a query asking when the state of emergency will be lifted are input to the question-answering model M1. The question-answering model M1 estimates that the part of the news article that reads "September 30th" is appropriate as the date when the state of emergency will be lifted, and outputs this part, "September 30th," as a response. Such a question-answering model M1 can also be used to obtain attribute values from product data.

The question-answering model M2 is a model that, when product data and a query are input, outputs a response corresponding to the query. The question-answering model M2 has learned training data that indicates the relationship between the product data and the query, and the response corresponding to the query. The query input to the question-answering model M2 is an attribute of the product indicated by the product data. The query can be said to be a question that asks about a part of the product data that contains an attribute value corresponding to the attribute. The response output from the question-answering model M2 is an attribute value corresponding to the attribute that is the query.

The question-answering model M2 estimates an appropriate part of the product data as a response, and outputs the estimated result as a response. In the example of FIG. 3, product data related to the characteristics of the bag and the attribute "brand" which is a query are input to the question-answering model M2. In FIG. 3, an image of the product is also input to the question-answering model M2, but in reality, only the text part of the product data is input to the question-answering model M2. Here, the image part of the product data may be converted into a combination of numerical values such as a feature vector via a feature extractor based on a machine learning model such as a neural network, and then input to the question-answering model M2. Here, the attribute used as the query is specified by the user.

For example, before uploading product data, a user inputs the product data in FIG. 3 and the attribute "brand" into question-answering model M2. Question-answering model M2 estimates that the portion "BBB bag" in the product data is appropriate as an attribute value (brand name) corresponding to the attribute "brand", and outputs this portion "BBB bag" as a response. For example, other attributes such as the attribute "color", the attribute "material", the attribute "size", or the attribute "function" can also be input as a query to question-answering model M2.

However, since the queries input to the question-answering model M2 are often shorter and less specific than those input to the question-answering model M1, the question-answering model M2 may not be able to recognize the meaning of the query. If the question-answering model M2 cannot recognize the meaning of the query, the accuracy of the output from the question-answering model M2 will also decrease. In particular, in the case of an online shopping mall, when training data is created based on product data, there tend to be many attributes of popular products, and the training data for other products tends to be sparse. The sparseness of the training data is also one of the factors that can reduce the accuracy of the output from the question-answering model M2.

Taking the attribute "function" as an example, a product such as an air conditioner may have functions such as "timer" and "energy saving." A product such as a bag may have attributes such as "waterproof," "mildew-proof," or "stain-resistant." For this reason, the same attribute "function" may have a variety of meanings. The question-answering model M2 may not be able to recognize the meaning of a short, abstract query such as the attribute "function," and may not be able to output an appropriate response.

The same is true for other attributes. If a short and abstract query is input to the question-answering model M2, the question-answering model M2 may not be able to fully recognize the meaning of the query, and the accuracy of the output from the question-answering model M2 may decrease. That is, the quality of the query may decrease the accuracy of the output from the question-answering model M2. Therefore, in this embodiment, by inputting not only the product attributes but also pre-prepared attribute values for the attributes as the query, the query is expanded and the quality of the query is improved. That is, the problem of sparseness of the training data mentioned above can also be solved.

Figure 4 is a diagram showing an example of a question-answering model in an embodiment. Question-answering model M3 is similar to question-answering model M2 in that it is a model that outputs a response according to the query when product data and a query are input. However, as described above, the query is expanded so that the meaning of the query is easier to recognize. In this embodiment, the query included in the training data for question-answering model M3 also includes attribute values in addition to attributes, but the training data for question-answering model M3 may be similar to that of question-answering model M2.

For example, suppose the product indicated by the product data is a bag, and the product attribute input as the query is "function." As mentioned above, the attribute "function" is a short and abstract attribute, and has various meanings related to products other than bags. For this reason, a query is input that includes the attribute "function" and the attribute value "mold prevention" that is prepared in advance for the bag-related attribute "function," so that the question-answering model M3 can recognize that the meaning is a bag-related attribute.

The product in FIG. 4 is not a product that has a function for preventing mold, but the attribute value "mold prevention" included in the query allows the question-answering model M3 to recognize the meaning of the attribute "function". As a result, the question-answering model M3 can output, as a response, an attribute value such as "waterproof", which is a function name that has an expression that is semantically or typographically similar to the attribute value "mold prevention" included in the query, from the product data. Similarly, when another expression that is semantically or typographically similar to the attribute value "waterproof" to be extracted, such as the attribute value "stain prevention" (function for preventing stains), is given as a query, the question-answering model M3 can output the attribute value "waterproof" from the product data as a response. This point is also the same when a query that includes a short and abstract attribute, such as the attribute "type", is input to the question-answering model M3, not limited to the attribute "function".

In this embodiment, one of a number of pre-prepared attribute values is selected at random. For this reason, in the example of FIG. 4, the attribute "function" specified by the user and the randomly selected attribute value "anti-mold" are input as a query to the question answering model M3. If an attribute value indicating the function of an air conditioner is selected for a bag product, there is a possibility that the question answering model M3 will not be able to output an appropriate response. Therefore, it is assumed that the meaning of the bag-related attribute "function" can be specified in some way, such as by the user. The meaning of the bag-related attribute "function" may be specified from other attributes and attribute values specified by the user, or the user himself may specify the attribute value "anti-mold" of the bag-related attribute "function".

For example, by inputting not only other attributes such as the attribute "brand", the attribute "color", the attribute "material", the attribute "size", or the attribute "function" as well as pre-prepared attribute values as queries, the question answering model M3 can more accurately recognize the meaning of the query. This makes it possible to improve the accuracy of the output from the question answering model M3. The attribute value extraction system S will be described in detail below.

[3. Functions Realized by the Attribute Value Extraction System]
5 is a functional block diagram showing an example of functions realized by the attribute value extraction system S. In this embodiment, a case will be described in which the main functions are realized by the server 10. The data storage unit 100 is mainly realized by the storage unit 12. The learning unit 101, the second attribute acquisition unit 102, the second attribute value acquisition unit 103, the third attribute value extraction unit 104, and the third attribute value assignment unit 105 are mainly realized by the control unit 11.

Hereinafter, the product data, query, and attribute value included in the training data of the question-answering model M3 are referred to as the first product data, first query, and first attribute value, respectively. The product indicated by the first product data is referred to as the first product. The product data and query input to the question-answering model M3 are referred to as the second product data and second query, respectively. The attribute and attribute value included in the second query are referred to as the second attribute and second attribute value, respectively. The product indicated by the second product data is referred to as the second product. The attribute value output from the trained question-answering model M3 is referred to as the third attribute value.

[3-1. Data storage unit]
The data storage unit 100 stores data necessary for assigning a third attribute value to the second product. For example, the data storage unit 100 stores a question answering model M3 in which the relationship between the first product data and the first query regarding the first product and the first attribute value regarding the first product is learned. The first product is an example of a first item. The first product data is an example of first item data. Therefore, the first product can be read as the first item. The first product data can be read as the first item data.

The first item is an item for training the question answering model M3. An item is an object to which attributes and attribute values are assigned. An item may be a tangible object such as a product, or an intangible object such as a service or data. An item may be any object and is not limited to a product. For example, an item may be content related to accommodation, content related to a restaurant, an e-book, a video, a song, a website, or other content. For example, an item may be an introduction to a service such as a financial service or a communication service, or a post on a social networking site.

The first item data is item data that includes some content related to the first item. The item data is data related to details of the item. The item data may be data according to the item. For example, if the item is an accommodation facility, the title or description of the accommodation facility or room may correspond to the item data. For example, if the item is an e-book, the actual data portion of the e-book may correspond to the item data. The first item data is item data related to the first item.

For example, the data storage unit 100 stores a trained question-answering model M3. The question-answering model M3 includes a program portion for executing processing on input data and a parameter portion that is adjusted by learning. The data storage unit 100 stores the question-answering model M3 before learning, and the parameters of the question-answering model M3 before learning are adjusted by learning. When learning is completed, the question-answering model M3 before learning is updated to the question-answering model M3 after learning. For example, in addition to the question-answering model M3, the data storage unit 100 also stores a training database DB1 and an attribute database DB2.

Figure 6 is a diagram showing an example of the training database DB1. The training database DB1 is a database in which training data for the question-answering model M3 is stored. The training data is data relating to the relationship between the first product data and the first query, and the first attribute value. The training data includes pairs of an input portion that is input to the question-answering model M3 during learning, and an output portion that is to be output from the question-answering model M3.

The input portion of the training data has the same format as the data input to the trained question-answering model M3. Therefore, the format of the first product data and the first query is the same as the format of the second product data and the second query. In this embodiment, the first product data and the first query, and the second product data and the second query are both in character format, but these may be in any format, such as numbers, other symbols, or a combination of these and characters. For example, if character features are input to the question-answering model M3 rather than characters, the features of the first product data and the first query may be included in the training data as the input portion of the training data.

The output portion of the training data has the same format as the data output from the trained question-answering model M3. Therefore, the format of the first attribute value and the format of the third attribute value are the same. In this embodiment, the first attribute value and the third attribute value are both in character format, but they may be in any format, such as numbers, other symbols, or a combination of these and characters. For example, if the question-answering model M3 does not output a character string indicating the attribute value, but rather outputs an ID or number that can identify the attribute value, the ID or number that can identify the first attribute value may be included in the training data as the output portion of the training data.

In this embodiment, a case will be described in which a product actually sold at an online shopping mall corresponds to the first product. That is, product data actually uploaded to the online shopping mall corresponds to the first product data. For example, the first product data is the title, description, or a combination of these of a product sold at the online shopping mall. The first product data includes characters entered by the user. Note that the training data may be created manually by an administrator of the attribute value extraction system S, rather than being created based on products actually sold at the online shopping mall.

The first query is a question for inquiring about an attribute value of the first product. For example, the first query is a combination of an attribute of the first product and an attribute value corresponding to the attribute. The attribute of the first product is an attribute associated with the first product. For example, the attribute included in the first query includes characters indicating an attribute specified by the user. The attribute value included in the first query may be an attribute value specified by the user, or may be an attribute value randomly selected from multiple attribute values prepared for the attribute included in the first query.

The first attribute value is an attribute value that is the correct answer for the first product. The first attribute value is any one of a plurality of attribute values prepared for the attribute included in the first query. The first attribute value may be the same as or different from the attribute value included in the first query. For example, the first attribute value includes characters that indicate the attribute value specified by the user. If the user specifies an incorrect attribute value, a correct attribute value may be specified by an administrator of the attribute value extraction system S, and the attribute value may be used as the first attribute value.

Figure 7 is a diagram showing an example of attribute database DB2. Attribute database DB2 is a database in which pre-prepared attribute values for each of a plurality of attributes are stored. For example, attribute database DB2 stores each of a plurality of attributes and a plurality of pre-prepared attribute values. Attribute database DB2 is used to obtain the second attribute value.

In this embodiment, the attribute database DB2 is created based on training data. For example, the server 10 associates an attribute included in the input portion of the training data with a first attribute value, which is the output portion corresponding to the input portion, and stores them in the attribute database DB2. The server 10 may store in the attribute database DB2 only first attribute values that appear in a predetermined number or more of training data as attribute values to be associated with a certain attribute.

The attribute database DB2 can be created by any method and does not have to be created based on training data. For example, the server 10 may create the attribute database DB2 based on a product database related to products sold in an online shopping mall. In this case, the server 10 may store in the attribute database DB2 pairs of attributes and attribute values of products that are not used as training data, or may store in the attribute database DB2 pairs of attributes and attribute values that appear in a predetermined number or more of products. For example, the attribute database DB2 may be manually created by an administrator of the attribute value extraction system S. For example, a list of attributes and attribute values that can be assigned to products sold in an online shopping mall may be used as the attribute database DB2.

[3-2. Learning section]
The learning unit 101 executes a learning process of the question answering model M3 based on the training data stored in the training database DB1. For example, the learning unit 101 executes the learning process of the question answering model M3 such that, when an input portion of training data is input, an output portion of training data is output. The learning process itself can use various algorithms according to the question answering model M3, and may take the form of fine tuning performed on a pre-trained Transformer-based model such as BERT, and an algorithm such as backpropagation or gradient descent may be used to find optimal parameters for other models.

For example, the learning unit 101 combines first product data included in the input portion of the training data with a first query included in the input portion of the training data. A word indicating that it is the first product data may be assigned to the first product data. A word indicating that it is the first query may be assigned to the first query. Furthermore, a word indicating that it is an attribute and an attribute value may be assigned to the first query. A word indicating a word separator may be placed between the first product data and the first query. The learning unit 101 executes a learning process such that, when the first product data and the first query are input, a first attribute value associated with the first product data and the first query is output.

[3-3. Second attribute acquisition part]
The second attribute acquisition unit 102 acquires a second attribute associated with second product data related to the second product. The second product is an example of a second item. The second product data is an example of second item data. Therefore, the section describing the second product can be read as the second item. The section describing the second product data can be read as the second item data. The second item is an item for estimation by the question answering model M3. The meaning of the item is as described above. The second item data is item data including some content related to the second item. The meaning of the item data is also as described above.

In this embodiment, the second attribute acquisition unit 102 acquires the second attribute specified by the user. For example, the second attribute acquisition unit 102 acquires the second attribute from the user terminal 20. The second attribute can be acquired by any method and does not have to be specified by the user. For example, if the attribute value extraction system S is used for the purpose of determining whether appropriate attributes and attribute values are assigned to a product, the second attribute acquisition unit 102 may refer to the product database and acquire an attribute associated with the product as the second attribute. In other words, the second attribute acquisition unit 102 may acquire an attribute assigned as an index of the product as the second attribute.

For another example, when a predetermined character is present in the second product data, the second attribute acquisition unit 102 may acquire the character as the second attribute. The second attribute acquisition unit 102 may determine whether or not there is a character in the second product data that matches the name of an attribute defined in the attribute database DB2. The second attribute acquisition unit 102 may acquire, from the second product data, a character that matches the name of an attribute defined in the attribute database DB2 as the second attribute.

[3-4. Second attribute value acquisition unit]
The second attribute value acquisition unit 103 acquires a second attribute value associated with a second attribute related to a second product. In this embodiment, since the relationship between the attribute and the attribute value is defined in the attribute database DB2, the second attribute value acquisition unit 103 refers to the attribute database DB2 to acquire the second attribute value. In this embodiment, since the second attribute is acquired by the second attribute acquisition unit 102, the second attribute value acquisition unit 103 acquires a second attribute value associated with the second attribute acquired by the second attribute acquisition unit 102. Note that the relationship between the attribute and the attribute value may be defined in data other than the attribute database DB2. For example, a list indicating the relationship between the attribute and the attribute value may be stored in the data storage unit 100.

The attribute database DB2 in this embodiment is created based on the training data, so referring to the attribute database DB2 to obtain the second attribute value corresponds to the second attribute value acquisition unit 103 obtaining the second attribute value based on the training data of the question answering model M3. The relationship between the attribute and the attribute value is also defined in the training database DB1, so the second attribute value acquisition unit 103 may refer to the training database DB1 to obtain the second attribute value associated with the second attribute.

The second attribute value acquisition unit 103 may acquire the second attribute value based on dictionary data on candidates for the second attribute value. This dictionary data is different from the training data. The dictionary data may be data in which the administrator of the attribute value extraction system S manually defines the relationship between the attribute and the attribute value, or product category data used in e-commerce services or the like may be used as the dictionary data. The second attribute value acquisition unit 103 may refer to the dictionary data and acquire the second attribute value associated with the second attribute. Whether or not the second attribute exists in the dictionary data is determined by string matching. The list described above may be an example of dictionary data.

In this embodiment, the second attribute value acquisition unit 103 randomly acquires a second attribute value from among a plurality of second attribute values associated with the second attribute. For example, the second attribute value acquisition unit 103 refers to the attribute database DB2 and identifies a plurality of second attribute values associated with the second attribute. The second attribute value acquisition unit 103 randomly acquires one of the identified plurality of second attribute values. Various methods can be used as the method for randomly selecting the second attribute value, and may be, for example, a method that uses random numbers.

[3-5. Third attribute value extraction unit]
The third attribute value extraction unit 104 extracts the third attribute value from the second product data based on the second product data related to the second product, the second query corresponding to the second attribute value, and the question answering model M3. In this embodiment, since the second query includes the second attribute and the second attribute value, the third attribute value extraction unit 104 extracts the third attribute value based on the second query corresponding to the second attribute and the second attribute value.

For example, the third attribute value extraction unit 104 combines the second product data and the second query. The second product data may be assigned a word indicating that it is the second product data. The second query may be assigned a word indicating that it is the second query. Furthermore, the second query may be assigned a word indicating that it is the second attribute and the second attribute value. A word indicating a word separator may be placed between the second product data and the second query. The question answering model M3 calculates the feature amounts of each of the input second product data and the second query, and outputs the third attribute value based on the calculated feature amounts. For the method of outputting the attribute value according to the feature amount of the input word, please refer to Non-Patent Document 1 and Non-Patent Document 2 listed as prior art documents, or reference documents described therein.

In this embodiment, the third attribute value extraction unit 104 extracts, as the third attribute value, a word included in the second product data estimated as the third attribute value by the question answering model M3. The question answering model M3 outputs information that identifies a portion of the second product data that is appropriate as the third attribute value. For example, the question answering model M3 outputs the start portion and the end portion of the portion that is appropriate as the third attribute value. If the 10th to 13th characters of the second product data are appropriate as the third attribute value, the question answering model M3 outputs information indicating that the 10th character is the start portion and information indicating that the 13th character is the end portion. The third attribute value extraction unit 104 acquires the portion from the 10th to 13th characters as the third attribute value.

In this embodiment, the question answering model M3 outputs the start and end parts of the third attribute value, and the third attribute value extraction unit 104 extracts the part between these start and end parts from the second product data as the third attribute value. The question answering model M3 may extract the third attribute value from the second product data and output it, rather than outputting the start and end parts. Note that the third attribute value does not have to be included directly as a wording in the second product data. An attribute value corresponding to the second attribute input as a query to the question answering model M3 does not have to be included in the second product data, which is an entity.

[3-6. Third attribute value assignment unit]
The third attribute value assigning unit 105 assigns the third attribute value extracted by the third attribute value extracting unit 104 to the second product. Assigning the third attribute value to the second product means associating the third attribute value with the second product data. The third attribute value assigning unit 105 assigns the third attribute value to the second product by associating the second product data input to the question answering model M3 with the third attribute value acquired by the third attribute value extracting unit 104 and recording them in the data storage unit 100. The third attribute value may be assigned to the second product after the user has confirmed whether it is appropriate.

[4. Processing Executed in the Attribute Value Extraction System]
Fig. 8 is a flow diagram showing an example of processing executed by the attribute value extraction system S. The processing in Fig. 8 is executed by the control unit 11 of the server 10 operating in accordance with a program stored in the storage unit 12. It is assumed that the training database DB1 and the attribute database DB2 have been created in advance before the processing in Fig. 8 is executed.

As shown in FIG. 8, the server 10 executes a learning process of the question-answering model M3 based on the training database DB1 (S1). In S1, the server 10 adjusts the parameters of the question-answering model M3 so that when an input portion of each of the multiple training data stored in the training database DB1 is input, an output portion of the training data is output. When the learning process of S1 is completed, the user becomes able to use the learned question-answering model M3.

The server 10 acquires the second product data and the second attribute input by the user from the user terminal 20 (S2). The server 10 acquires the second attribute value randomly from among the multiple attribute values associated with the second attribute acquired in S2 based on the attribute database DB2 (S3). The server 10 acquires the second query based on the second attribute acquired in S2 and the second attribute value acquired in S3 (S4).

The server 10 extracts a third attribute value from the second product data based on the second product data acquired in S2, the second query acquired in S4, and the question answering model M3 (S5). In S5, the server 10 inputs the second product data and the second query to the question answering model M3. The question answering model M3 calculates features based on the second product data and the second query, and outputs the start and end parts of the third attribute value according to the calculated features. The server 10 extracts the character string between the start and end parts as the third attribute value. The server 10 assigns the third attribute value extracted in S5 to the second product (S6), and the process ends.

According to the attribute value extraction system S of this embodiment, the third attribute value is extracted based on the second product data, the second query corresponding to the second attribute value, and the question-answering model M3. By using the second attribute value that allows the question-answering model M3 to easily recognize the meaning of the second query, the accuracy of the output from the question-answering model M3 is improved. For example, even if the second product has an abstract attribute such as "function", the question-answering model M3 can output an appropriate third attribute value. By improving the accuracy of the output from the question-answering model M3, a more appropriate third attribute value can be assigned to the second product. For example, since the conventional question-answering model M1 is created assuming that a relatively long query is input, the accuracy of the output may decrease when a relatively short query is input as in the conventional question-answering model M2. In this regard, by making the second query have a more specific meaning, such as the second attribute value, it is possible to obtain an output with higher accuracy than the conventional question-answering model M2.

The attribute value extraction system S also extracts a third attribute value based on the second attribute and a second query corresponding to the second attribute value. This further improves the accuracy of the output from the question answering model M3. By including the second attribute in the second query, it is easy to identify which attribute the third attribute value output from the question answering model M3 belongs to.

The attribute value extraction system S also acquires the second attribute value based on training data for making the question-answering model M3 learn relationships. This results in a second query that corresponds to the content that the question-answering model M3 has been made to learn, making it easier for the question-answering model M3 to recognize the meaning of the second query. This further increases the accuracy of the output from the question-answering model M3.

The attribute value extraction system S also acquires the second attribute value based on dictionary data related to candidates for the third attribute value. This results in a second query defined in the dictionary data, making it easier for the question-answering model M3 to recognize the meaning of the second query. This further increases the accuracy of the output from the question-answering model M3.

The attribute value extraction system S also randomly acquires the second attribute value from among a plurality of second attribute values associated with the second attribute. This simplifies the process of acquiring the second attribute value and shortens the time required to acquire the output from the question answering model M3. As a result, the processing speed of the server 10 can be increased and the processing load on the server 10 can be reduced.

The attribute value extraction system S also extracts, as the third attribute value, a word contained in the second product data that is estimated as the third attribute value by the question answering model M3. By setting the word contained in the second product data as the third attribute value, a more appropriate third attribute value can be assigned to the second product.

The attribute value extraction system S also acquires a second attribute associated with the second product data, and acquires a second attribute value associated with the acquired second attribute. This allows an appropriate third attribute value to be assigned according to the attribute of the second product.

The first item is a first product, and the first item data is data relating to the first product. The second item is a second product, and the second item data is data relating to the second product. This makes it possible to extract an appropriate third attribute value. For example, it is possible to eliminate the need for a user at an online shopping mall to specify a third attribute value. For example, even if an incorrect attribute value is assigned to a product sold at an online shopping mall, it is possible to assign an appropriate attribute value by treating this product as a second product and assigning the third attribute value.

5. Modifications
The present disclosure is not limited to the above-described embodiment, and may be modified as appropriate without departing from the spirit and scope of the present disclosure.

[5-1. Modification 1]
For example, the second attribute value acquisition unit 103 may acquire a second attribute value with a relatively high frequency from among a plurality of second attribute values associated with the second attribute. The frequency is the frequency with which an attribute value appears (occurs) in product data. For example, the server 10 counts the number of times each attribute and attribute value are associated with each other in the product data of products sold in an online shopping mall. The server 10 acquires the frequency of each second attribute value based on the counting result. The method of counting the frequency may be any method, and may be, for example, the frequency in the training database DB1 or the frequency in the product database of all products in the online shopping mall.

A second attribute value with a relatively high frequency is a second attribute value for which there is at least one other second attribute value with a lower frequency than itself. A second attribute value with a relatively high frequency can also be defined as a second attribute value other than the second attribute value with the lowest frequency. For example, if there are four second attribute values, the second attribute value with the highest frequency, the second attribute value with the second highest frequency, and the second attribute value with the third highest frequency correspond to second attribute values with a relatively high frequency.

In the first modification, a case is described in which the second attribute value acquisition unit 103 acquires the second attribute value with the highest frequency from among a plurality of second attribute values, but the second attribute value acquisition unit 103 only needs to acquire a second attribute value with a relatively high frequency, and may acquire a second attribute value with a frequency that is second or higher. By combining the embodiment and the first modification, the second attribute value acquisition unit 103 may randomly acquire a second attribute value from among the second attribute values with a relatively high frequency.

Assuming that there are three second attribute values with relatively high frequencies, as in the example of four second attribute values described above, the second attribute value acquisition unit 103 may randomly acquire a second attribute value from among the three second attribute values with relatively high frequencies. In this case, the second or third most frequently occurring second attribute value may be acquired. The processing after the second attribute value acquisition unit 103 acquires the second attribute value is the same as in the embodiment, and a second query corresponding to the second attribute value is acquired and estimation is performed by the question answering model M3.

According to the first variant example, a second attribute value with a relatively high frequency is obtained from among a plurality of second attribute values associated with the second attribute. A second attribute value with a relatively high frequency often appropriately expresses the meaning of the second attribute, and therefore there is a high probability that the second query is a word that makes it easy for the question-answering model M3 to recognize the meaning of the second query. By making the second query correspond to a second attribute value with a relatively high frequency, the question-answering model M3 becomes a second query whose meaning is easier to understand, and the accuracy of the output from the question-answering model M3 is further improved.

In addition, the most frequent second attribute value is obtained from among the multiple second attribute values. The most frequent second attribute value often more appropriately expresses the meaning of the second attribute, and therefore there is a higher probability that the second query is a word that makes it easier for the question-answering model M3 to recognize the meaning of the second query. By setting the second query according to the most frequent second attribute value, the question-answering model M3 becomes a second query whose meaning is easier to understand, and the accuracy of the output from the question-answering model M3 is further improved.

[5-2. Modification 2]
For example, the second attribute value acquisition unit 103 may acquire a plurality of second attribute values. In the second modification example, a second query corresponding to a plurality of second attribute values is input to the question answering model M3. For example, the second query includes the second attribute and a plurality of second attribute values. For example, the second query is a text in which the second attribute and the plurality of second attribute values are combined. As in the embodiment, each second attribute value may be given a special word indicating that it is a second attribute value. Furthermore, a special word indicating a separator may be placed between the second attribute values.

In the second modification, the second attribute value acquisition unit 103 acquires all of the second attribute values associated with the second attribute, but the second attribute value acquisition unit 103 may acquire a part of the second attribute values associated with the second attribute. For example, if there are four second attribute values, the second attribute value acquisition unit 103 may acquire two or three of the four second attribute values. In this case, by combining the embodiment and the second modification, the second attribute value acquisition unit 103 may randomly acquire two or three second attribute values from among the four second attribute values. As another example, by combining the embodiment and the first modification, the second attribute value acquisition unit 103 may acquire two or three second attribute values from among the four second attribute values in order of frequency.

The third attribute value extraction unit 104 extracts the third attribute value based on the second query corresponding to the multiple second attribute values. In the second modification, all of the multiple second attribute values associated with the second attribute are acquired, so the third attribute value extraction unit 104 extracts the third attribute value based on the second query corresponding to all of the multiple second attribute values. For example, the third attribute value extraction unit 104 inputs the second product data and the second query including all of the multiple second attribute values acquired by the second attribute value acquisition unit 103 to the question answering model M3. This differs from the embodiment in that the second query may be longer than in the embodiment, but the processing of the question answering model M3 itself is the same as in the embodiment. In the second modification, the input portion of the training data may include multiple attribute values. That is, in the second modification, a question answering model M3 in which multiple second attribute values are learned as the input portion may be used.

According to the second modification, a third attribute value is extracted based on a second query corresponding to a plurality of second attribute values. This increases the amount of information in the second query, making it possible to make the second query easier for the question-answering model M3 to recognize the meaning of. By utilizing a plurality of second attribute values, the second query can be effectively expanded, thereby further increasing the accuracy of the output from the question-answering model M3.

In addition, all of the multiple second attribute values associated with the attribute are obtained, and the third attribute value is extracted based on a second query corresponding to all of the multiple second attribute values. This increases the amount of information in the second query, making it possible to make the second query easier for the question-answering model M3 to recognize the meaning of. By utilizing all of the multiple second attribute values, the second query can be expanded more effectively, thereby improving the accuracy of the output from the question-answering model M3.

[5-3. Modification 3]
For example, when a plurality of second attribute values are acquired as in the second modification, the third attribute value extraction unit 104 may calculate an average of the feature quantities for each of the second attribute values, and extract the third attribute value based on the calculated average. The feature quantity is information relating to the features of a word. In the third modification, a case will be described in which the feature quantity is a so-called embedded vector, but the feature quantity is not limited to a vector format and can be expressed in any format. For example, the feature quantity may be in another format such as a single numerical value or an array.

Figures 9 to 11 are diagrams showing an example of a process for calculating the average of features for each of multiple second attribute values. As shown in Figure 9, second product data and a second query including a second attribute and multiple second attribute values are input to the question answering model M3. The embedding layer of the question answering model M3 calculates the features of the input words. The method of calculating the features by the embedding layer is the same as that of the conventional question answering model M1 and question answering model M2.

The third attribute value extraction unit 104 calculates the average of the feature quantities calculated by the embedding layer. For example, the third attribute value extraction unit 104 may calculate the average for each word included in each of the multiple second attribute values, and extract the third attribute value based on the average calculated for each word. In the example of FIG. 10, it is assumed that two second attribute values, the second attribute value "dark blue" and the second attribute value "brilliant yellow", are obtained for the second attribute "color".

The third attribute value extraction unit 104 calculates the average of the feature amount of "dark", the first word of the second attribute value "dark blue", and the feature amount of "brilliant", the first word of the second attribute value "brilliant yellow", as the average of the first word. The third attribute value extraction unit 104 calculates the average of the feature amount of "blue", the second word of the second attribute value "dark blue", and the feature amount of "yellow", the second word of the second attribute value "brilliant yellow", as the average of the second word.

The third attribute value extraction unit 104 calculates the average of the feature amounts for each word included in each of the second attribute values "dark blue" and "brilliant yellow", and extracts the third attribute value based on the average. When three or more second attribute values are input to the question answering model M3, the third attribute value extraction unit 104 may also calculate the average of the feature amounts for each word in the same manner as in FIG. 10, and obtain an output according to the average. When three or more second attribute values are input to the question answering model M3, the third attribute value extraction unit 104 may also calculate the average of the feature amounts for three or more second attribute values for each word in the same manner as in FIG. 10, and extract the third attribute value based on the average.

Note that the multiple second attribute values may include combinations of second attribute values with different numbers of words. For example, one second attribute value may include only one word, and another second attribute value may include two words. In this case, the output unit may calculate the average for each word based on padding for a second attribute value that has a relatively small number of words among the multiple second attribute values. The padding is a predetermined character. In the example of FIG. 11, it is assumed that two second attribute values, the second attribute value "dark blue" and the second attribute value "red", are acquired for the second attribute "color". The second attribute value "dark blue" is two words, while the second attribute value "red" is one word.

The third attribute value extraction unit 104 calculates the average of the feature amount of "dark", which is the first word of the second attribute value "dark blue", and the feature amount of "red", which is the first word of the second attribute value "red", as the average of the first word. Since there is no second word in the second attribute value "red", the third attribute value extraction unit 104 calculates the average of the feature amount of "blue", which is the second word of the second attribute value "dark blue", and the feature amount of "pad", which is a predetermined padding, as the average of the second word. In this way, the third attribute value extraction unit 104 calculates the average of the feature amounts for the second attribute value "red" after compensating for the padding that is insufficient to the number of words in the second attribute value "dark blue", which has the largest number of words, and extracts the third attribute value based on this average.

According to the third modification, the average of the feature quantities for each of the multiple second attribute values is calculated, and the third attribute value is extracted based on the second query corresponding to the calculated average. This makes it possible to effectively expand the second query by averaging the features of the multiple second attribute values, thereby further improving the accuracy of the output from the question-answering model M3.

In addition, an average is calculated for each word contained in each of the multiple second attribute values, and a third attribute value is extracted based on a second query corresponding to the average calculated for each word. This makes it possible to average the features for each word of the multiple second attribute values before obtaining an output from the question answering model M3, thereby further improving the accuracy of the output from the question answering model M3. For example, it becomes possible to assign a specific number of words of a third attribute value to a second product.

In addition, when the multiple second attribute values include a combination of second attribute values with different numbers of words, the average for each word is calculated based on the padding for the second attribute value with a relatively small number of words among the multiple second attribute values. This makes it possible to obtain output from the question answering model M3 even if the second attribute values have different numbers of words, thereby further improving the accuracy of the output from the question answering model M3. For example, it becomes possible to assign to the second product the third attribute value with the greatest number of words among the multiple second attribute values.

[5-4. Modification 4]
For example, when there are multiple second attributes associated with the second product, the second attribute value acquisition unit 103 may acquire second attribute values associated with each of the multiple second attributes. The process of acquiring second attribute values associated with each second attribute is as described in the embodiment. The second attribute value acquisition unit 103 executes the process described in the embodiment for each second attribute to acquire a second attribute value. The second attribute value acquisition unit 103 may execute the process described in any of the modified examples 1 to 3 for each second attribute value to acquire a second attribute value.

The third attribute value extraction unit 104 extracts a third attribute value for each second attribute. The process of extracting third attribute values according to each second attribute is as described in the embodiment. The third attribute value assignment unit 105 assigns a third attribute value to the second product for each second attribute. The process of assigning a third attribute value according to each second attribute is as described in the embodiment. For example, the third attribute value assignment unit 105 may assign a third attribute value to the second product for each second attribute such as "brand", "color", "material", "size" and "function".

According to the fourth modification, a second attribute value associated with each of a plurality of second attributes is obtained, and a third attribute value is extracted for each second attribute. This makes it possible to extract a third attribute value for each second attribute even if a second product has a plurality of second attributes associated therewith. As a result, for example, the burden on the user can be further reduced.

For example, the functions described as being realized by the server 10 may be realized by another computer such as the user terminal 20, or may be shared among multiple computers. For example, data stored in the data storage unit 100 may be stored in a database server.

S Attribute value extraction system, N Network, P Product page, 10 Server, 11, 21 Control unit, 12, 22 Memory unit, 13, 23 Communication unit, 20 User terminal, 24 Operation unit, 25 Display unit, M1, M2, M3 Question answering model, 100 Data memory unit, 101 Learning unit, 102 Second attribute acquisition unit, 103 Second attribute value acquisition unit, 104 Third attribute value extraction unit, 105 Third attribute value assignment unit, DB1 Training database, DB2 Attribute database.

Claims (17)

a storage unit that stores an attribute value extraction model in which a relationship between a first item data item and a first query item related to a first item and a first attribute value related to the first item is learned;
a second attribute value acquisition unit that acquires the second attribute value by referring to data in which a second attribute related to a second item and a second attribute value are associated with each other ;
a third attribute value extraction unit that extracts a third attribute value related to the second attribute from the second item data based on second item data related to the second item, a second query including the second attribute and the second attribute value, and the attribute value extraction model;
An attribute value extraction system comprising: The second attribute value acquisition unit acquires the second attribute value based on training data of the attribute value extraction model.
The attribute value extraction system according to claim 1 . the second attribute value acquisition unit acquires the second attribute value based on dictionary data regarding candidates for the second attribute value;
3. The attribute value extraction system according to claim 1 or 2 . the second attribute value acquisition unit randomly acquires the second attribute value from among a plurality of second attribute values associated with the second attribute;
4. The attribute value extraction system according to claim 1 . the second attribute value acquisition unit acquires the second attribute value having a relatively high frequency from among the plurality of second attribute values associated with the second attribute;
5. The attribute value extraction system according to claim 1 . the second attribute value acquisition unit acquires the second attribute value that occurs most frequently from among the plurality of second attribute values;
The attribute value extraction system according to claim 5 . the second attribute value acquisition unit acquires a plurality of the second attribute values;
the second query includes the plurality of second attribute values;
7. An attribute value extraction system according to claim 1 . the second attribute value acquisition unit acquires all of the second attribute values associated with the second attribute;
the second query includes all of the plurality of second attribute values;
The attribute value extraction system according to claim 7 . the third attribute value extraction unit calculates an average of the feature amounts calculated by the attribute value extraction model for each of the plurality of second attribute values, and extracts the third attribute value based on the calculated average;
9. The attribute value extraction system according to claim 7 or 8 . the third attribute value extraction unit calculates the average for each word included in each of the plurality of second attribute values, and extracts the third attribute value based on the average calculated for each word.
The attribute value extraction system according to claim 9 . The plurality of second attribute values include combinations of the second attribute values each having a different number of the words;
the third attribute value extraction unit calculates the average for each of the words based on padding for the second attribute value having a relatively small number of the words among the plurality of second attribute values;
The attribute value extraction system according to claim 10 . the third attribute value extraction unit extracts, as the third attribute value, a word included in the second item data estimated as the third attribute value by the attribute value extraction model;
An attribute value extraction system according to any one of claims 1 to 11 . the attribute value extraction system includes a second attribute acquisition unit that acquires the second attribute designated by a user ;
the second attribute value acquisition unit acquires the second attribute value associated with the second attribute acquired by the second attribute acquisition unit;
An attribute value extraction system according to any one of claims 1 to 12 . the second attribute value acquisition unit acquires the second attribute value associated with each of the plurality of second attributes;
the third attribute value extraction unit extracts the third attribute value for each of the second attributes;
An attribute value extraction system according to any one of claims 1 to 13 . the first item is a first commodity;
the first item data is first product data relating to the first product;
the second item is a second commodity;
The second item data is second product data related to the second product.
An attribute value extraction system according to any one of claims 1 to 14 . The computer
a second attribute value acquisition step of acquiring the second attribute value by referring to data in which a second attribute related to a second item and a second attribute value are associated with each other ;
a third attribute value extraction step of extracting the third attribute value related to the second attribute from the second item data based on second item data related to the second item, a second query including the second attribute and the second attribute value, and an attribute value extraction model in which a relationship between first item data related to a first item, the first query, and the first attribute value related to the first item is learned;
The attribute value extraction method to perform . a second attribute value acquisition unit that acquires the second attribute value by referring to data in which a second attribute related to a second item and a second attribute value are associated with each other ;
a third attribute value extraction unit that extracts the third attribute value related to the second attribute from the second item data based on an attribute value extraction model in which a relationship between first item data related to a first item and the first query and a first attribute value related to the first item is learned based on second item data related to the second item and a second query including the second attribute and the second attribute value;
A program that makes a computer function as a

Images