KR102640369B1 - Method, computer device, and computer program to search for products based on embedding similarity - Google Patents

Abstract

A product search method, computer device, and computer program based on embedding similarity are disclosed. The product search method includes generating an embedding-based search model that searches for a product based on embedding matching between a query and a product using text appearing on a product search service; And it may include providing a product search result that matches the embedding of the input query using the embedding-based search model for the input query.

Description

Product search method based on embedding similarity, computer device, and computer program {METHOD, COMPUTER DEVICE, AND COMPUTER PROGRAM TO SEARCH FOR PRODUCTS BASED ON EMBEDDING SIMILARITY}

The explanation below is about the technology for searching for products.

General product search analyzes the terms of the query and product and provides products that match the term of the query as search results.

As an example of product search technology, Korean Patent Publication No. 10-2007-0097720 (published on October 5, 2007) discloses a technology that provides product search results through query matching.

In the case of term matching-based product search, products that have the same meaning as the term of the query but are expressed in a different form of term cannot be provided as search results.

For queries with low search frequency or queries close to natural language, even if a related product exists, product search results may not be provided due to difficulty in term matching with the query.

Provides methods and devices that can be applied to matching and ranking by analyzing queries and products in embedding units rather than term units.

Provides a method and device for searching products based on embedding similarity between queries and products.

In a product search method executed on a computer device, the computer device includes at least one processor configured to execute computer-readable instructions included in a memory, and the product search method includes, by the at least one processor, a product Creating an embedding-based search model that searches for a product based on embedding matching between the query and the product using text appearing on the search service; and providing, by the at least one processor, a product search result that matches an embedding of the input query using the embedding-based search model for the input query.

According to one aspect, the step of providing the product search results includes linking a term matching search model that searches for products based on term matching between the query and the product with the embedding-based search model to create an embedding matching product and a term matching product. It can be provided as a result of the above product search.

According to another aspect, the step of generating the embedding-based search model includes a product database that maintains product information, a shopping document set (shopping corpus) consisting of text of product reviews or queries registered by users, and text output from the shopping domain. Extracting text from at least one of the shopping vocabulary consisting of and using it as learning data for creating the embedding-based search model, but in the case of a numeric field, converting it to text and using it as the learning data. there is.

According to another aspect, the step of generating the embedding-based search model includes generating query-product pair data according to the degree of relevance based on click logs between the query and the product as fine tuning data; and further learning the embedding-based search model using the fine tuning data.

According to another aspect, the step of generating the fine tuning data may include assigning a label corresponding to the degree of association between the query and the product to the query-product pair data.

According to another aspect, the step of generating the fine tuning data may include differentially applying the importance of the query-product pair data based on an indicator indicating the degree of association between the query and the product.

According to another aspect, the step of generating the fine tuning data involves negative sampling of products that do not match the term of the query among products that are similar to the product clicked on for the query, and generates a query with a low correlation between the query and the product. -May include steps for generating product pair data.

According to another aspect, the step of providing the product search results includes indexing the product with a centroid corresponding to the embedding of the product in a cluster centroid learned as a product embedding set; And it may include searching for a product of the same centroid as the centroid corresponding to the embedding of the input query among the indexed products.

According to another aspect, the step of providing the product search results involves searching for products with the same PQ code as the PQ code corresponding to the embedding of the input query using a PQ (product quantization) codebook learned as a product embedding set. Additional steps may be included.

According to another aspect, the step of searching for the product may include excluding an embedding matching product by the embedding-based search model from the product search result according to the quality of the centroid corresponding to the embedding of the input query. You can.

According to another aspect, the step of providing the product search results includes adapting the weight for the embedding matching product by the embedding-based search model and the weight for the term matching product by the term matching search model according to the input query. It may include the step of setting it as an enemy.

According to another aspect, the step of providing the product search results includes searching for the product using embedding similarity with the input query through the embedding-based search model and term similarity with the input query through the term matching search model. A step of providing rankings for search results, wherein the step of providing rankings for product search results includes adaptively setting a weight for the embedding similarity and a weight for the term similarity according to the input query. May include steps.

A computer program stored in a computer-readable recording medium is provided to execute the product search method on a computer device.

A computer device comprising: at least one processor configured to execute computer readable instructions included in a memory, wherein the at least one processor performs embedding matching between a query and a product using text appearing on a product search service. A computer device is provided that generates an embedding-based search model for searching products, and provides product search results that match the embedding of the input query using the embedding-based search model for an input query.

According to embodiments of the present invention, more accurate product search results for a query can be provided by providing products with embeddings similar to the embedding of the query as search results.

According to embodiments of the present invention, the quality of search ranking can be improved by using the embedding similarity between the query and the product as a feature for determining the ranking of the search results.

1 is a diagram illustrating an example of a network environment according to an embodiment of the present invention.
Figure 2 is a block diagram showing an example of a computer device according to an embodiment of the present invention.
FIG. 3 is a diagram illustrating an example of components that a processor of a computer device according to an embodiment of the present invention may include.
Figure 4 is a flowchart showing an example of a method that can be performed by a computer device according to an embodiment of the present invention.
Figure 5 is an example diagram illustrating the process of pre-learning a basic model for product search in one embodiment of the present invention.
Figure 6 is an example diagram illustrating the process of generating an embedding search model by fine-tuning a basic model in one embodiment of the present invention.
Figures 7 to 9 are example diagrams for explaining the process of searching for products based on embedding similarity in one embodiment of the present invention.
Figure 10 is an example diagram illustrating a process for providing search results through ensemble with other search models in one embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings.

Embodiments of the present invention relate to product search technology for matching and ranking products related to a query.

Embodiments including those specifically disclosed in this specification can perform matching and ranking operations to provide search results based on query and product embedding, thereby improving search efficiency, accuracy, and quality. Significant advantages can be achieved.

The product search device according to embodiments of the present invention may be implemented by at least one computer device, and the product search method according to embodiments of the present invention is performed through at least one computer device included in the product search device. It can be. At this time, the computer program according to an embodiment of the present invention may be installed and driven in the computer device, and the computer device may perform the product search method according to the embodiments of the present invention under the control of the driven computer program. . The above-described computer program can be combined with a computer device and stored in a computer-readable recording medium to execute the product search method on the computer.

1 is a diagram illustrating an example of a network environment according to an embodiment of the present invention. The network environment in FIG. 1 shows an example including a plurality of electronic devices 110, 120, 130, and 140, a plurality of servers 150 and 160, and a network 170. Figure 1 is an example for explaining the invention, and the number of electronic devices or servers is not limited as in Figure 1. In addition, the network environment in FIG. 1 only explains one example of environments applicable to the present embodiments, and the environment applicable to the present embodiments is not limited to the network environment in FIG. 1.

The plurality of electronic devices 110, 120, 130, and 140 may be fixed terminals or mobile terminals implemented as computer devices. Examples of the plurality of electronic devices 110, 120, 130, and 140 include smart phones, mobile phones, navigation devices, computers, laptops, digital broadcasting terminals, Personal Digital Assistants (PDAs), and Portable Multimedia Players (PMPs). ), tablet PC, etc. For example, in FIG. 1, the shape of a smartphone is shown as an example of the electronic device 110. However, in embodiments of the present invention, the electronic device 110 actually communicates with other devices through the network 170 using a wireless or wired communication method. It may refer to one of various physical computer devices capable of communicating with electronic devices 120, 130, 140 and/or servers 150, 160.

The communication method is not limited, and may include not only a communication method utilizing a communication network that the network 170 may include (for example, a mobile communication network, wired Internet, wireless Internet, and a broadcast network), but also short-range wireless communication between devices. For example, the network 170 may include a personal area network (PAN), a local area network (LAN), a campus area network (CAN), a metropolitan area network (MAN), a wide area network (WAN), and a broadband network (BBN). , may include one or more arbitrary networks such as the Internet. Additionally, the network 170 may include any one or more of network topologies including a bus network, star network, ring network, mesh network, star-bus network, tree or hierarchical network, etc. Not limited.

Each of the servers 150 and 160 is a computer device or a plurality of computers that communicate with a plurality of electronic devices 110, 120, 130, 140 and a network 170 to provide commands, codes, files, content, services, etc. It can be implemented with devices. For example, the server 150 may be a system that provides a service (for example, a product search service, etc.) to a plurality of electronic devices 110, 120, 130, and 140 connected through the network 170.

Figure 2 is a block diagram showing an example of a computer device according to an embodiment of the present invention. Each of the plurality of electronic devices 110, 120, 130, and 140 described above or each of the servers 150 and 160 may be implemented by the computer device 200 shown in FIG. 2.

As shown in FIG. 2, this computer device 200 may include a memory 210, a processor 220, a communication interface 230, and an input/output interface 240. The memory 210 is a computer-readable recording medium and may include a non-permanent mass storage device such as random access memory (RAM), read only memory (ROM), and a disk drive. Here, non-perishable large-capacity recording devices such as ROM and disk drives may be included in the computer device 200 as a separate permanent storage device that is distinct from the memory 210. Additionally, an operating system and at least one program code may be stored in the memory 210. These software components may be loaded into the memory 210 from a computer-readable recording medium separate from the memory 210. Such separate computer-readable recording media may include computer-readable recording media such as floppy drives, disks, tapes, DVD/CD-ROM drives, and memory cards. In another embodiment, software components may be loaded into the memory 210 through the communication interface 230 rather than a computer-readable recording medium. For example, software components may be loaded into memory 210 of computer device 200 based on computer programs installed by files received over network 170.

The processor 220 may be configured to process instructions of a computer program by performing basic arithmetic, logic, and input/output operations. Commands may be provided to the processor 220 by the memory 210 or the communication interface 230. For example, processor 220 may be configured to execute received instructions according to program code stored in a recording device such as memory 210.

The communication interface 230 may provide a function for the computer device 200 to communicate with other devices (eg, the storage devices described above) through the network 170. For example, a request, command, data, file, etc. generated by the processor 220 of the computer device 200 according to a program code stored in a recording device such as memory 210 is transmitted to the network ( 170) and can be transmitted to other devices. Conversely, signals, commands, data, files, etc. from other devices may be received by the computer device 200 through the communication interface 230 of the computer device 200 via the network 170. Signals, commands, data, etc. received through the communication interface 230 may be transmitted to the processor 220 or memory 210, and files, etc. may be stored in a storage medium (as described above) that the computer device 200 may further include. It can be stored as a permanent storage device).

The input/output interface 240 may be a means for interfacing with the input/output device 250. For example, input devices may include devices such as a microphone, keyboard, or mouse, and output devices may include devices such as displays and speakers. As another example, the input/output interface 240 may be a means for interfacing with a device that integrates input and output functions, such as a touch screen. The input/output device 250 may be configured as a single device with the computer device 200.

Additionally, in other embodiments, computer device 200 may include fewer or more components than those of FIG. 2 . However, there is no need to clearly show most prior art components. For example, the computer device 200 may be implemented to include at least some of the input/output devices 250 described above, or may further include other components such as a transceiver, a database, etc.

Hereinafter, specific embodiments of a method and device for searching products based on embedding similarity will be described.

In this embodiment, in order to expand the coverage of product search results, product matching and product ranking can be performed by analyzing queries and products in embedding units rather than term units. In other words, these embodiments can provide products with embeddings similar to the query embeddings as search results.

FIG. 3 is a block diagram showing an example of components that a processor of a computer device may include according to an embodiment of the present invention, and FIG. 4 is a method that the computer device may perform according to an embodiment of the present invention. This is a flowchart showing an example.

The computer device 200 according to this embodiment can provide a product search service to clients through a dedicated application installed on the client or through access to a web/mobile site related to the computer device 200. The computer device 200 may be configured with a computer-implemented product search device.

The processor 220 of the computer device 200 is a component for performing the product search method according to FIG. 4, and as shown in FIG. 3, the model learning unit 310, the model tuning unit 320, and the product search It may include unit 330. Depending on the embodiment, components of the processor 220 may be selectively included in or excluded from the processor 220. Additionally, depending on the embodiment, components of the processor 220 may be separated or merged to express the functions of the processor 220.

The processor 220 and the components of the processor 220 can control the computer device 200 to perform steps S410 to S440 included in the product search method of FIG. 3. For example, the processor 220 and its components may be implemented to execute instructions according to the code of an operating system included in the memory 210 and the code of at least one program.

Here, the components of the processor 220 may be expressions of different functions performed by the processor 220 according to instructions provided by program codes stored in the computer device 200. For example, the model learning unit 310 may be used as a functional representation of the processor 220 that controls the computer device 200 according to the above-described commands so that the computer device 200 learns a shopping model.

The processor 220 may read necessary instructions from the memory 210 where instructions related to controlling the computer device 200 are loaded. In this case, the read command may include a command for controlling the processor 220 to execute steps S410 to S440 that will be described later.

Steps S410 to S440 to be described later may be performed in an order different from the order shown in FIG. 4, and some of the steps S410 to S440 may be omitted or additional processes may be included.

Referring to FIG. 4, in step S410, the model learning unit 310 may pretrain a basic model for product search using text from a shopping domain that provides a product search service. For example, the model learning unit 310 may generate a model that well understands text in the shopping domain based on the BERT (Bidirectional Encoder Representations from Transformer) model. At this time, the model learning unit 310 may generate a shopping corpus by extracting text from a product database that maintains product information and product reviews or inquiries registered from product buyers. Additionally, the model learning unit 310 may generate a shopping vocabulary dictionary by extracting text such as words or phrases that frequently appear in the shopping domain. The model learning unit 310 may generate a basic model by pre-training the BERT model using at least one of a shopping document set and a shopping vocabulary dictionary. In addition to the BERT model, a basic model can be created using an arbitrary pretrained language model (PLM) that can extract embeddings and learn the meaning of the text.

In step S420, the model tuning unit 320 fine-tunes the basic model based on the query and the embedding of the product, thereby creating an embedding matching-based product search model (hereinafter referred to as 'embedding search model') that can be used for actual product search. ) can be created. For example, the model tuning unit 320 analyzes the click log between the query and the product (products clicked among products provided as search results for the query) and the product database to determine which products are highly relevant and which are low related to the query. Fine tuning data can be generated by distinguishing between . The model tuning unit 320 may input the query and product generated from fine tuning data into the basic model to generate an embedding vector of the query and an embedding vector of the product. At this time, products with a high degree of relevance to the query learn the basic model to increase the similarity (e.g., cosine similarity) between the embedding vectors, and conversely, products with a low degree of correlation learn the basic model to increase the similarity between the embedding vectors (e.g., cosine similarity). ) can be learned to reduce the basic model. The model tuning unit 320 may generate an embedding search model by going through a learning process using fine tuning data on the basic model.

In step S430, when a query is given in a real-time search process, the product search unit 330 may search for products having an embedding similar to the embedding of the given query. First, for product search, the product search unit 330 can extract embeddings for all products using an embedding search model, and apply the K-means clustering algorithm to the product embeddings to find an appropriate number of products. Cluster centroids can be learned. In the offline indexing step, the product search unit 330 may extract and index a product centroid term using the centroid ID closest to the product for each product. Thereafter, the product search unit 330 may search by matching products having the same product centroid term as the query centroid term corresponding to the embedding of the query entered in the real-time search step. In addition, the product search unit 330 can use a PQ (product quantization) codebook learned from a set of product embeddings to match and search products having the same PQ code as the PQ code of the query embedding.

In step S440, the product search unit 330 may rank search results using an embedding search model based on embedding similarity between the query and the product. In other words, the product search unit 330 can use the embedding similarity between the query and the product as a feature to determine the ranking of the search results.

Furthermore, the product search unit 330 may adaptively ensemble the embedding search model and at least one other search model according to the query in product matching (S430) and product ranking (S440). For example, the product search unit 330 may provide search results by linking an embedding search model for a given query with a term matching search model that searches for products having the same term as the term of the query. At this time, from the product matching perspective, the product search unit 330 may adaptively set the weight for the product searched through term matching and the weight for the product searched through embedding matching according to the query. In some cases, in order to maintain the accuracy of search results at a high level, logic can be applied to omit the search using the embedding search model depending on the quality of the centroid closest to the query embedding. Additionally, when using the term similarity and embedding similarity between the query and the product as ranking features, the product search unit 330 can adaptively set the weight for the term similarity and the weight for the embedding similarity according to the query.

Figure 5 is an example diagram to explain the process of pre-learning a basic model for product search in one embodiment of the present invention.

In actual services, it is necessary to lighten the model structure so that the task of extracting query and product embeddings can operate quickly. After setting the target real-time processing speed and throughput indicators required in the embedding extraction process, design a model structure at a level that can achieve the set indicators.

For example, referring to Figure 5, the parameters of a lightweight model that satisfies the target index are distilled from the BERT model 50, or the shopping vocabulary dictionary, which is a learning document, is added to the BERT model 50 consisting of a lightweight model with initialized parameters. A method of directly pre-learning the (501) and shopping document set (502) can be applied.

Downstream tasks are performed through the process of transfer learning a shopping vocabulary dictionary (501) and a shopping document set (502) consisting of text from the shopping domain to the lightweight BERT model (50) with multiple versions of the model structure. ), the model with the best performance can be selected as the basic model (510).

The model learning unit 310 may generate the basic model 510 by pre-training the lightweight BERT model 50 using the shopping vocabulary dictionary 501 and the shopping document set 502.

The model learning unit 310 may utilize the characteristics of a shopping domain rather than a general web document in the process of generating the shopping document set 502. The model learning unit 310 extracts the text of all fields that can be used for term matching with queries from the product database and product reviews, and then preprocesses and normalizes the extracted text using a morphological analyzer. For example, the model learning unit 310 can extract numerical fields that are difficult to use for term matching, such as product price, size, category ID, color code, and manufacturer serial number, as text and use them for learning. For example, if the price of a product is 11,000 won, it can be converted into text indicating the price range, such as '10,000 won range', and used for learning.

In the process of creating the shopping document set 502, not only the product database can be used, but also data obtained from reviews or queries entered by users can also be used. For example, by analyzing the distribution of price range information included in user queries, the section and interval for extracting price range text can be learned.

In this embodiment, pre-learning data for generating the basic model 510 can be configured by considering the characteristics of the shopping domain.

Figure 6 is an example diagram illustrating the process of generating an embedding search model by fine-tuning a basic model in one embodiment of the present invention.

Referring to FIG. 6, the model tuning unit 320 can apply a fine tuning process to the basic model 510 to create an embedding search model 620 that can be used to extract embeddings of queries and products.

In the fine tuning process according to this embodiment, product data clicked within the search results can mainly be utilized. By joining and analyzing product click logs and product databases clicked on in the search results of a query, products with a high relevance to a query are distinguished from products with a low relevance, and the query-product pair data with a high relevance and a query with a low correlation are classified. -Product pair data can be created as fine tuning data.

The model tuning unit 320 inputs fine tuning data into the basic model 510 and performs learning to improve the quality of the embedding vectors of queries and products. Referring to FIG. 6, for products that are suitable for the query, the basic model 510 is learned to increase the cosine similarity calculated from the embedding of the query and the product, and for products that are not suitable for the query, the embedding of the query and the product is learned. The basic model 510 can be learned in the direction of lowering the cosine similarity calculated from .

Likewise, the characteristics of the shopping domain can be utilized when extracting product embeddings during the fine tuning process. Unlike general document embedding, which is modeled as a list of sentences, shopping products can be composed of various text fields and numeric fields. In addition to utilizing major text fields used in general term matching (e.g., product name, category name, brand, store name, product review, etc.), numeric fields that are difficult to utilize in term matching such as price, size, and various codes Additionally, it can be converted to text format and used for product embedding.

As an example of a technology to improve the quality of the embedding search model 620, differential labels can be assigned to query-product pair data and used as fine tuning data. Rather than configuring learning data with labels according to the presence or absence of a dichotomous association, learning data can be configured by assigning differential labels to queries and products (shopping domain) to indicate the degree to which the product is related to the query. there is. You can quantify how the product's text field is related to the query and create a label based on this. For example, for the query 'xx laptop', laptop products of brand xx are labeled 'very related', laptop products of brand xx and other brands are labeled 'related', and completely different labels such as refrigerators or clothing are labeled. A product may be given a label corresponding to 'unrelated'.

As another example, the importance of query-product pair data can be differentially applied by using indicators that indicate the degree of relevance to the query, such as the number of clicks on the product provided as a search result and CTR (click-to-exposure ratio). For example, model parameters can be updated so that the more clicks a query has on a product and the higher the CTR, the higher the similarity between the query and the product.

As another example, negative sampling techniques can be used to construct query-product pair data with low correlation. It is difficult to secure a good quality embedding search model (620) if you simply randomly sample products that were not clicked on in response to the query. Rather than products that are unrelated to the query, products that have an appropriate correlation with the query but have a low correlation in terms of text matching can be negatively sampled and used as query-product pair data with a low correlation. For example, among products of the same brand or seller as the most clicked products in response to the query, or products in a similar price range or category, products that do not match the important terms that make up the query term can be negatively sampled. For query-product pair data composed of negative sampling, you can learn by setting a label according to the importance of terms that do not match between the query and the product, and then adjusting the degree to which the cosine similarity is reduced according to the set label.

Therefore, the model tuning unit 320 considers the characteristics of the shopping domain and generates query-product pair data with a high degree of correlation and query-product pair data with a low degree of correlation as fine tuning data, and then creates the label and cosine of the fine tuning data. An embedding search model 620 in which the correlation coefficient between similarity prediction values converges can be created.

Figures 7 to 9 are example diagrams to explain the process of searching for products based on embedding similarity in one embodiment of the present invention.

Referring to FIG. 7, the product search unit 330 can extract embeddings for all products existing in the product database using the embedding search model 620 through offline processing. You can apply the same method of quickly searching by organizing the product term into a reverse index.

The product search unit 330 can learn an appropriate number of cluster centroids through K-means clustering and then use them to index the embedding of the product. In the process of repeatedly learning centroids, quality indicators can be used to tune the best model parameters. At this time, the silhouette coefficient, which is a clustering quality indicator, and the entropy and impurity of the main fields of the product can be used as quality indicators. The entropy and impurity of each major field of products clustered in a cluster can be considered additional quality indicators that can be confirmed in terms of product clustering, considering the characteristics of the shopping domain.

The product search unit 330 may extract and index the product centroid term using the centroid ID closest to each product in the offline indexing step. For example, as shown in Figure 8, assuming that the embedding of a specific product is closest to the 614th centroid, a term called 'cent614' can be added to the inverse index. The product search unit 330 may search by matching products having the same product centroid term as the query centroid term corresponding to the embedding of the query entered in the real-time search step. Likewise, for embeddings extracted in real time from user queries, products indexed by centroids closest to the embeddings in the query can be quickly partially searched.

Referring to FIG. 9, the product search unit 330 can apply a method of quickly partially searching products with the same PQ code as the PQ code of the query embedding in a manner similar to the cluster centroid. Internally, a separate PQ codebook can be learned and utilized when quickly calculating the cosine similarity approximation between query embedding and product embedding.

In order to maintain the accuracy of search results at a high level, if the quality of the centroid closest to the query embedding is below a certain level, search using the embedding search model may not be used. The quality of a centroid may be determined based on the categories of products assigned to the centroid. For example, the more products in categories that are less related to each other are assigned to the same centroid, the lower the quality of the centroid may be judged to be.

The product search unit 330 can utilize a method of querying a pre-dumped embedding matrix to increase the efficiency of real-time extraction work for query embedding. As a form of caching, it is possible to extract embeddings in advance for frequently incoming queries and then dump the extracted embeddings and the indices of the embeddings. When a pre-dumped query is received, the product search unit 330 searches the embedding index of the incoming query using a data structure such as a trie, and then searches the embedding matrix through the index to perform query embedding in real time. It can be extracted with

Figure 10 is an example diagram for explaining a process of providing search results through ensemble with other search models in one embodiment of the present invention.

The product search unit 330 may provide search results by linking an embedding search model and a term matching search model for the input query. When applying the embedding search model to product matching, products that cannot be searched through term matching alone can be additionally searched.

Referring to FIG. 10, the product search unit 330 can adaptively set the weight between the term matching product 1010 by the term matching search model and the embedding matching product 1020 by the embedding search model according to the given query. . The weight of the embedding matching product 1020 may be determined based on the quality index of the centroid. For example, if the centroid closest to the embedding of the query matches the category of the query to a high degree, the weight of the embedding matching product 1020 can be set to be relatively higher than that of the term matching product 1010. On the other hand, if the products assigned to the centroid closest to the embedding of the query are too diverse and the degree of matching with the category of the query is low, the weight of the embedding matching product (1020) can be set relatively lower than that of the term matching product (1010). You can. Depending on the embodiment, the embedding matching product 1020 may be excluded from the search results depending on the quality (silhouette coefficient, entropy, impurity, etc.) of the centroid closest to the embedding of the query.

Additionally, the product search unit 330 can use the term similarity and embedding similarity between the query and the product as ranking features. In the embedding search model, embedding similarity, which could not be used in the term matching search model, can be provided as a ranking feature. For example, the product search unit 330 may determine the product ranking by adding the embedding similarity with the query to the term similarity. If product ranking is provided by adding embedding similarity and term similarity, too much similarity may be given to products to which embedding matching and term matching are applied simultaneously. As another example to solve this issue, the product search unit 330 may set the weight for term similarity and the weight for embedding similarity differently depending on the query. The product search unit 330 can determine the weight of embedding similarity according to a given query by learning the weight of embedding similarity to be reflected in product ranking through a query level LTR (learning to rank) technique using click logs between queries and products. .

In this way, according to embodiments of the present invention, more accurate product search results can be provided for queries by providing products with embeddings similar to the embedding of the query as search results, and the embedding similarity between the query and the product can be used to rank the search results. The quality of search ranking can be improved by using it as a feature to determine.

The device described above may be implemented with hardware components, software components, and/or a combination of hardware components and software components. For example, the devices and components described in the embodiments include a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), and a programmable logic unit (PLU). It may be implemented using one or more general-purpose or special-purpose computers, such as a logic unit, microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications running on the operating system. Additionally, a processing device may access, store, manipulate, process, and generate data in response to the execution of software. For ease of understanding, a single processing device may be described as being used; however, those skilled in the art will understand that a processing device includes multiple processing elements and/or multiple types of processing elements. It can be seen that it may include. For example, a processing device may include a plurality of processors or one processor and one controller. Additionally, other processing configurations, such as parallel processors, are possible.

Software may include a computer program, code, instructions, or a combination of one or more of these, which may configure a processing unit to operate as desired, or may be processed independently or collectively. You can command the device. The software and/or data may be embodied in any type of machine, component, physical device, computer storage medium or device for the purpose of being interpreted by or providing instructions or data to the processing device. there is. Software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. At this time, the medium may continuously store a computer-executable program, or temporarily store it for execution or download. In addition, the medium may be a variety of recording or storage means in the form of a single or several pieces of hardware combined. It is not limited to a medium directly connected to a computer system and may be distributed over a network. Examples of media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, And there may be something configured to store program instructions, including ROM, RAM, flash memory, etc. Additionally, examples of other media include recording or storage media managed by app stores that distribute applications, sites or servers that supply or distribute various other software, etc.

As described above, although the embodiments have been described with limited examples and drawings, various modifications and variations can be made by those skilled in the art from the above description. For example, the described techniques are performed in a different order than the described method, and/or components of the described system, structure, device, circuit, etc. are combined or combined in a different form than the described method, or other components are used. Alternatively, appropriate results may be achieved even if substituted or substituted by an equivalent.

Therefore, other implementations, other embodiments, and equivalents of the claims also fall within the scope of the claims described below.

Claims (15)

In a product search method executed on a computer device,
The computer device includes at least one processor configured to execute computer-readable instructions contained in a memory,
The product search method is:
providing, by the at least one processor, product search results matching an embedding of the input query using an embedding-based search model for the input query.
Including,
The step of providing the product search results is,
Indexing a product from a cluster centroid learned as a set of product embeddings into a centroid corresponding to the embedding of the product; and
Searching for products of the same centroid as the centroid corresponding to the embedding of the input query among the indexed products.
Product search method including. According to paragraph 1,
The step of providing the product search results is,
Linking the term matching search model, which searches for products based on term matching between queries and products, with the embedding-based search model to provide embedding-matched products and term-matched products as the product search results.
A product search method characterized by . According to paragraph 1,
The embedding-based search model is,
Text extracted from at least one of a product database that maintains product information, a shopping document set (shopping corpus) consisting of the text of product reviews or queries registered by users, and a shopping vocabulary (shopping vocabulary) consisting of text output from a shopping domain. Use it as learning data, but in the case of numeric fields, convert them to text and use them as the learning data.
A product search method characterized by . According to paragraph 3,
The embedding-based search model is,
Using query-product pair data according to the degree of correlation based on click logs between queries and products as fine tuning data
A product search method characterized by . According to paragraph 4,
The fine tuning data is,
A label corresponding to the degree of correlation between the query and the product is assigned to the query-product pair data.
A product search method characterized by . According to paragraph 4,
The fine tuning data is,
Differentially applying the importance of the query-product pair data based on an indicator indicating the degree of correlation between the query and the product.
A product search method characterized by . According to paragraph 4,
The fine tuning data is,
Generating query-product pair data with a low correlation between the query and the product by negative sampling products that are similar to the product clicked on in response to the query but do not match the term of the query.
A product search method characterized by . delete According to paragraph 1,
The step of providing the product search results is,
Searching for products with the same PQ code as the PQ code corresponding to the embedding of the input query using a PQ (product quantization) codebook learned as a set of product embeddings.
A product search method that further includes. According to paragraph 1,
The step of searching for the product is,
Excluding embedded matching products by the embedding-based search model from the product search results according to the quality of the centroid corresponding to the embedding of the input query.
Product search method including. According to paragraph 2,
The step of providing the product search results is,
Adaptively setting a weight for an embedding matching product by the embedding-based search model and a weight for a term matching product by the term matching search model according to the input query.
Product search method including. According to paragraph 2,
The step of providing the product search results is,
Providing a ranking for the product search results using embedding similarity with the input query through the embedding-based search model and term similarity with the input query through the term matching search model.
Including,
The step of providing rankings for the product search results is:
Adaptively setting the weight for the embedding similarity and the weight for the term similarity according to the input query.
Product search method including. A computer program stored in a computer-readable recording medium for executing the product search method of any one of claims 1 to 7 and 9 to 12 on a computer device. In computer devices,
At least one processor configured to execute computer readable instructions contained in memory
Including,
The at least one processor,
A process of providing product search results matching the embedding of the input query using an embedding-based search model for the input query.
Process it,
The at least one processor,
Index products into the centroid corresponding to the product embedding in the cluster centroid learned from the product embedding set,
Searching for products of the same centroid as the centroid corresponding to the embedding of the input query among the indexed products.
A computer device characterized by a. According to clause 14,
The at least one processor,
Linking the term matching search model, which searches for products based on term matching between queries and products, with the embedding-based search model to provide embedding-matched products and term-matched products as the product search results.
A computer device characterized by a.

Images