KR102538519B1 - System and Method for recomending related product using key phrase extraction - Google Patents

Abstract

The present invention relates to a device and method for recommending a related product using a key phrase extraction that extract a phrase or clause configured with one word or two or more consecutive words with connotation, representativeness, and originality from text data as a key phrase, to easily check a meaning and characteristic having the text data. The device comprises: a text data collection part; a key phrase extraction part; and a related product recommendation part.

Description

Apparatus and method for recommending related products using key phrase extraction {System and Method for recomending related product using key phrase extraction}

The present invention relates to an apparatus and method for recommending a related product using key phrase extraction, and more particularly, from text data for a specific product, a word having connotation, representativeness, and originality, or a phrase composed of two or more consecutive words ( An apparatus and method for recommending related products using key phrase extraction that extracts phrases or clauses as key phrases and recommends other products related to the corresponding product using the same.

BACKGROUND As communication technology develops, an E-commerce market for buying and selling goods and services using wireless communication is continuously expanding. In particular, the e-commerce market has been able to grow significantly due to increased accessibility to e-commerce due to the spread of mobile terminals, and recently, it is growing even more significantly due to the global spread of COVID-19.

When purchasing products and services through e-commerce, consumers cannot directly verify the product or service provider, so consumers rely on information provided by the seller (images, descriptions, etc.) and review information written by other users about the product. based on which decision to purchase is made.

However, the information provided by the seller may be exaggerated or hide unfavorable information on the seller's side in order to induce purchase of the product, and thus reliability may be lowered from the consumer's point of view.

Accordingly, user reviews written by other users who have recently purchased products and services greatly contribute to consumers' purchasing decisions. Users' reviews of products are an important factor in supporting decision-making for companies and buyers, and e-commerce sellers make efforts to collect a large number of user reviews by providing benefits such as points or discount coupons to consumers who write user reviews. are leaning

Since such review data is created in thousands of pieces for a single product, and the review data is simply sorted by criteria such as newest, popular, and price, it is difficult for users to check many reviews one by one, and many reviews There is a limit to selectively filtering only reviews with meaningful information among reviews.

Accordingly, studies to quantify and analyze such atypical reviews are continuously being conducted. Conventional review analysis is a method of extracting only the keywords (nouns) of each review and constructing a dictionary of words with close distances for each dimension using the classification dimension and word embedding method, but this is an extension that is difficult to apply to review analysis of other product groups. has gender limitations.

On the other hand, the above-mentioned background art is technical information that the inventor possessed for derivation of the present invention or acquired in the process of derivation of the present invention, and cannot necessarily be said to be known art disclosed to the general public prior to filing the present invention. .

Korean Patent Registration No. 10-2310487

One aspect of the present invention extracts a word or a phrase consisting of two or more consecutive words with connotation, representativeness, and originality as a key phrase from review data on a specific product. and a related review filtering device and method using main phrase extraction for filtering review data related to a corresponding main phrase among all review data for a specific product using this.

The technical problem of the present invention is not limited to the technical problem mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art from the following description.

An apparatus for recommending related products using main phrase extraction according to an embodiment of the present invention extracts at least one sub-phrase from text data on a specific product, clusters the extracted sub-phrases into a plurality of groups, and A phrase or clause composed of one word or two or more consecutive words is extracted as a key phrase from a plurality of sub-phrases included in, and other products related to the corresponding product are extracted using this I recommend you.

The related product recommendation device using the main phrase extraction,

a text data collection unit that collects text data for each product;

Using a pre-learned Named Entity Recognition (NER) model, each word constituting the text data is distinguished whether it is the beginning of a predefined phrase, the middle of a predefined phrase, or not a predefined phrase, and A word corresponding to the beginning and a word corresponding to the middle of a phrase are grouped and extracted as one sub-phrase, a plurality of the sub-phrases are clustered into a plurality of groups, and a plurality of sub-phrases included in each group are grouped into one main phrase. a main phrase extraction unit for extracting phrases; and

A related product recommendation unit recommending related products related to a specific product according to a result of comparing the main phrase extracted for a specific product with sub-phrases extracted for other products is included.

The main phrase extraction part,

Sub-phrases less than the maximum allowable number are extracted from one text data according to the order of importance of attributes within preset passages from one text data.

The main phrase extraction part,

After labeling a first tag on a word corresponding to the beginning of a phrase, labeling a second tag on a word corresponding to the middle of a phrase, and labeling a third tag on a word other than a phrase,

Extract one word labeled with the first tag as a sub-phrase,

A first word labeled with a first tag and at least one second word following the first word and labeled with a second tag are bundled and extracted as one sub-phrase.

The main phrase extraction part,

An embedding vector for the sub-phrase is generated using a BERT (Bidirectional Encoder Representations from Transformers) model, and a plurality of the sub-phrases are divided into a plurality of groups based on the plurality of embedding vectors generated for each of the plurality of sub-phrases. cluster, but

When the number of text data for a specific product is collected above a preset reference value, the number of groups to be clustered is optimized using a silhouette score, and the number of text data for a specific product is collected below a preset reference value. In this case, a plurality of the sub-phrases are clustered into a plurality of groups according to a preset number of clusters.

The main phrase extraction part,

A representative value is extracted from a plurality of embedding vectors included in each group, and a sub-phrase corresponding to the extracted representative value is extracted as a main phrase.

The related product recommendation unit,

By calculating the average embedding similarity between the main phrase extracted from a specific product and sub-phrases extracted from other products, sub-phrases having an embedding similarity higher than a preset threshold are extracted, and the product corresponding to the extracted sub-phrase is selected as a specific product. It is recommended as a related product for

In addition, the related product recommendation method using main phrase extraction according to an embodiment of the present invention is performed by the related product recommendation device using main phrase extraction,

Collecting text data for each product;

Using a pre-learned Named Entity Recognition (NER) model, each word constituting the text data is distinguished whether it is the beginning of a predefined phrase, the middle of a predefined phrase, or not a predefined phrase, and A word corresponding to the beginning and a word corresponding to the middle of a phrase are grouped and extracted as one sub-phrase, a plurality of the sub-phrases are clustered into a plurality of groups, and a plurality of sub-phrases included in each group are grouped into one main phrase. extracting phrases; and

and recommending a related product related to a specific product according to a result of comparing the main phrase extracted for a specific product with sub-phrases extracted for other products.

The step of extracting the main phrase,

It is characterized in that sub-phrases less than the maximum allowable number are extracted from one text data according to the order of importance of attributes within preset passages from one text data.

The step of extracting the main phrase,

After labeling a first tag to a word corresponding to the beginning of a phrase through the entity name recognition model, labeling a second tag to a word corresponding to the middle of a phrase, and labeling a third tag to a word other than a phrase,

Extract one word labeled with the first tag as a sub-phrase,

A first word labeled with a first tag and at least one second word following the first word and labeled with a second tag are bundled and extracted as one sub-phrase.

The step of extracting the main phrase,

An embedding vector for the sub-phrase is generated using a BERT (Bidirectional Encoder Representations from Transformers) model, and a plurality of the sub-phrases are divided into a plurality of groups based on the plurality of embedding vectors generated for each of the plurality of sub-phrases. cluster, but

When the number of text data for a specific product is collected to be less than a preset reference value, a plurality of sub-phrases are clustered into a plurality of groups according to a preset number of clusters.

The step of extracting the main phrase,

When the number of text data for a specific product is collected over a preset reference value, the number of groups to be clustered is optimized using a silhouette score.

The step of extracting the main phrase,

It is characterized in that any one representative value is extracted from a plurality of embedding vectors included in each group, and a sub-phrase corresponding to the extracted representative value is extracted as a main phrase.

The step of recommending the related product is,

By calculating the average embedding similarity between the main phrase extracted from a specific product and sub-phrases extracted from other products, sub-phrases having an average embedding similarity higher than a preset threshold are extracted, and the product corresponding to the extracted sub-phrase is specified. It is characterized in that the product is recommended as a related product.

According to one aspect of the present invention described above, unlike the conventional keyword extraction technique, it is possible to reliably extract a key phrase, which is a phrase well reflecting the characteristics (implication, representativeness, originality) of text data.

In addition, by recommending related products using the extracted main phrases, the user can easily check information on related products that are meaningful to him or her.

1 is a diagram showing a schematic configuration of a related product recommendation device using main phrase extraction according to an embodiment of the present invention.
2 to 4 are diagrams illustrating a specific example of extracting a sub-phrase from one piece of text data about a food product.
5 is a conceptual diagram showing the entire process of the main phrase extraction unit.
6 is a diagram illustrating an example of a main phrase finally extracted for a specific product.
7 is a diagram showing an example of calculating the embedding similarity between a main phrase and a sub phrase in a related product recommendation unit.
8 is a diagram illustrating an example of specific text of a related product recommended by a related product recommendation unit.
9 is a diagram showing a schematic flow of a related product recommendation method using main phrase extraction according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The detailed description of the present invention which follows refers to the accompanying drawings which illustrate, by way of illustration, specific embodiments in which the present invention may be practiced. These embodiments are described in sufficient detail to enable one skilled in the art to practice the present invention. It should be understood that the various embodiments of the present invention are different from each other but are not necessarily mutually exclusive. For example, specific shapes, structures, and characteristics described herein may be implemented in another embodiment without departing from the spirit and scope of the invention in connection with one embodiment. Additionally, it should be understood that the location or arrangement of individual components within each disclosed embodiment may be changed without departing from the spirit and scope of the invention. Accordingly, the detailed description set forth below is not to be taken in a limiting sense, and the scope of the present invention, if properly described, is limited only by the appended claims, along with all equivalents as claimed by those claims. Like reference numbers in the drawings indicate the same or similar function throughout the various aspects.

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

1 is a diagram showing schematic components of an apparatus for recommending related products using main phrase extraction according to an embodiment of the present invention.

The apparatus for recommending related products using the extraction of main phrases according to the present invention may be a user terminal possessed by a user for understanding and analyzing the meaning of text data, and the user terminal is capable of wired and wireless communication with an external device and inputs and outputs information It may be an electronic device capable of processing, such as a smart phone, PC, tablet PC, wearable device, and the like.

The apparatus for recommending related products using main phrase extraction according to the present invention implemented in such a user terminal is one word or a phrase or clause composed of two or more consecutive words having connotation, representativeness, and originality from text data. By modeling a series of processes of extracting as a key phrase, reliable information can be efficiently extracted from a large amount of text data using artificial intelligence.

The main phrase extracted from text data related to a product using the related product recommendation device using main phrase extraction according to the present invention can represent the characteristics of the collection target (eg, product) better than other phrases among the phrases extracted from the text data. It means a key phrase in a sentence, and can have the following characteristics.

1) It is desirable that main passages have representativeness representing common characteristics of text data for each subject of collection.

2) It is desirable that the main passages have originality that shows the differentiation of text data among collection subjects.

3) It is desirable that the main passages have connotation that summarizes the contents of the text data on the subject of collection.

Therefore, among the extracted passages, the passages that have representativeness, originality, and connotation for the collection object compared to other passages can be regarded as effectively representing the characteristics of the collection object, and thus the corresponding passage can be selected as the main passage.

To this end, the related product recommendation device 1 using main phrase extraction according to an embodiment of the present invention includes a text data collection unit 100, a main phrase extraction unit 200, and a related product recommendation unit 300. .

The text data collection unit 100 collects text data related to a specific product.

Here, the text data related to the subject to be collected may be review data written by users who have purchased the product. However, the type of text data is not limited thereto, and may further include various types of text data collected for a specific domain, such as manuals of corresponding products, related articles, promotional materials, and the like.

The text data collecting unit 100 may collect text data in various ways. For example, the text data collecting unit 100 may scrape text data from an external server through wired/wireless communication. Alternatively, the text data collecting unit 100 may directly receive text data from the user terminal.

The text data collecting unit 100 classifies and stores the collected text data according to product (service) type.

Hereinafter, for convenience of description, it will be assumed that the text data collected by the text data collection unit 100 is text written about food products, but the type of text data is not limited thereto, and the main phrase extraction according to the present invention will be described. Of course, the used related product recommendation device can also be applied to texts for various other domains to which the text device can be applied (for example, electronic products, cosmetics, travel products, interior construction services, etc.).

The key phrase extraction unit 200 analyzes the collected text data for each product and extracts a key phrase, which is a phrase that well reflects the characteristics (implication, representativeness, originality) of each text data.

Specifically, the main phrase extractor 200 uses a pre-learned Named Entity Recognition (NER) model to classify entity name types for each word constituting text data.

The main phrase extractor 200 extracts one word or at least two consecutive words corresponding to a sub-phrase previously defined by a user through a named entity recognition (NER) model as a sub-phrase.

To this end, the main phrase extraction unit 200 distinguishes whether each word constituting the text data is the beginning of a predefined phrase, the middle of a predefined phrase, or not a predefined phrase by using an object name recognition model, and distinguishes According to the result, a first tag may be labeled for a word corresponding to the beginning of a phrase, a second tag may be labeled for a word corresponding to the middle of a phrase, and a third tag may be labeled for a word other than a phrase.

That is, the main phrase extractor 200 divides sentences constituting text data into token (word) units, and divides words by applying a tagging system set for each domain to the tokens.

The tagging system includes the BIESO system and the BIO system. In the BIESO system, 'B-(begin)' is placed at the beginning of the entity name, 'I-(inside)' is placed at the end of the entity name when the token is in the middle of the entity name. 'E-(end)' is attached when a token is an entity name, 'S-(singleton)' is attached when a token is an entity name, and 'O(outside)' is appended when a token is not an entity name. The BIO system is similar to the BIESO system, but has the characteristic that E is integrated into I and S is integrated into B.

The tagging system applied to the main phrase extractor 200 is the BIO system. When the entity name starts, the above-mentioned first tag 'B-(begin)' is used, and when the token is in the middle of the entity name, the above-mentioned second tag ' 'I-(inside)' is attached, and 'O(outside)', which is the above-mentioned third tag, is attached when the token is not an entity name.

2 shows an example of text data for a food product. As shown, the main phrase extractor 200 analyzes text data such as 'I bought it yesterday, the meat is delicious' and corresponds to 'I bought it yesterday'. A third tag (O, other) is labeled on the word, a first tag (B, begin) is labeled on the word corresponding to 'meat', and a second tag (I) is labeled on the word corresponding to 'delicious'. , inside).

That is, in the food domain, if a user wants to set a word that represents the type of material, taste/texture, amount of food, object and environment, convenience, etc. As attributes to be extracted among attributes are set, the sub-phrase extractor 200 may extract the entity name having the above expression as a sub-phrase through the entity name recognition model.

In some cases, the sub-phrase extractor 200 may automatically set the attributes to be extracted among the carpenter's attributes pre-learned in the object name recognition model according to the domain set by the user.

To this end, the main phrase extractor 200 may define phrases to be extracted differently according to domains. For example, when the text data is about food products, the main phrase extractor 200 may set properties such as taste, texture, cooking method, amount, object and environment, and convenience as main properties. As another example, if the main phrase extractor 200 sets a word representing price, portability, speed, etc. as a phrase that the user wants to extract in the electronic device domain, the entity name having the above expression through the entity name recognition model. can be extracted into sub-phrases.

In this way, the main phrase extractor 200 extracts sub-phrases consisting only of words that are determined to be important in the food text by setting phrases in the type of product or service among all attributes according to the characteristics of the product or service related to the text data. can be extracted.

As described above, the main phrase extractor 200 distinguishes whether each word constituting the text data is the beginning of a predefined phrase, the middle of a predefined phrase, or not a predefined phrase through the entity name recognition model, and determines the meaning of the phrase. After labeling the first tag (B) for the word corresponding to the beginning, labeling the second tag (I) for the word corresponding to the middle of the phrase, and labeling the third tag (O) for the word other than the phrase, A word corresponding to the beginning of a phrase and a word corresponding to the middle of the phrase are grouped together and extracted as one sub-phrase.

Specifically, when words are listed in the form below, the main phrase extraction unit 200 may extract the corresponding word or the corresponding words as one sub-phrase.

1) B

2) B-I

3) B-I-I

4) B-I-I-I

In the case of 1), the first tag (B) extracts one labeled word into one sub-phrase. At this time, the first tag (B) is labeled and the second tag (O) If there are no labeled words, single words are extracted as one sub-phrase.

In the case of 2), in the case of text data including a first word labeled with the first tag B and a second word following the first word and labeled with the second tag, the first word and the second word are labeled. Group the words and extract them into one sub-phrase.

For example, as shown in FIGS. 2 and 3 , after the main phrase extractor 200 labels each word constituting text data with a tag, the first tag (B) and the second tag (I) are Two consecutive words, 'meat is delicious', are extracted as one sub-phrase.

In the case of No. 2), the first tag (B) is labeled, the labeled first word, the second word subsequent to the first word and labeled with the second tag, and the subsequent word of the second word and labeled with the second tag In the case of text data including the third word and , the first word, the second word, and the third word are grouped and extracted as one sub-phrase. Similarly, in the case of 3), in the case of text data in which the first word labeled with the first tag B and three words labeled with the second tag are consecutively arranged from the first word, four words are grouped together. Extract into one sub-phrase. 4 shows examples of sub-phrases extracted by the main phrase extractor 200 .

In this way, the main phrase extractor 200 distinguishes whether each word constituting the text data is the beginning of a predefined phrase, the middle of a predefined phrase, or not a predefined phrase through the entity name recognition model, and determines the beginning of the phrase. After labeling the first tag to the corresponding word, labeling the second tag to the word corresponding to the middle of the phrase, labeling the third tag to the non-phrase word, and then labeling one word labeled with the first tag as a sub-word. A first word labeled with a first tag and at least one second word that is subsequent to the first word and labeled with a second tag are bundled and extracted as one sub-phrase. The sub-phrase extraction unit 200 repeatedly performs this operation for each text data to extract sub-phrases from all text data.

On the other hand, the main phrase extraction unit 200 selects examples of phrases from one text data in order of importance in order to prevent extraction of sub-phrases equal to or greater than the preset maximum allowable number from one text data in the process of extracting sub-phrases. Only phrases less than the allowable number are labeled and trained to be extracted. For example, even if there are 5 phrases in one text data, the sub-phrase extraction unit 200 configures the data and trains it so that only 3 sub-phrases are extracted according to the order of importance, so that the maximum allowable number of phrases (3) or more. Make sure this is not extracted.

To this end, the main phrase extraction unit 200 may set priorities of main attributes in advance, and may set priorities in the order of taste, texture > cooking method > quantity > object and environment > convenience. In this case, the main phrase extraction unit 200 has the first and second sub-phrases among the 5 sub-phrases for taste/texture, the 3rd sub-phrase for cooking methods, and the 4th sub-phrase for cooking methods. If it is confirmed that it is a sub-phrase for quantity and the 5th sub-phrase is a sub-phrase for convenience, only the 1st, 2nd, and 3rd sub-phrases can be extracted from the text data, excluding the 4th and 5th sub-phrases, which have low priority. there is. It should be understood that such a sub-phrase extraction rule is set in various ways according to the type of product/service.

Below is an example of sub-phrases extracted by the main phrase extractor 200 .

Key attribute 1: taste, texture

"The meat was soft and it tasted better than I expected haha" -> "The meat is soft"

"One bag of 8 dumplings. They are not too spicy and the blood is chewy." -> "It's not too spicy", "The blood is chewy."

"It's so delicious, no need to go to a specialty store ha" -> "No need to go to a specialty store"

Key attribute 2: Cooking method

"It's delicious, so I'll buy it again. Good for air fryer" -> "Good for air fryer"

Key Attribute 3: Quantity

"It's just the right amount to eat alone. I like that it has ham and kimchi in it." -> "Perfect for eating alone"

"It's delicious because it's cooked in an air fryer. It's suitable for two." -> "It's suitable for two."

Key Attribute 4: Target and Environment

"I took it when I went camping. The amount is just right and it's nice" -> "I took it when I went camping"

"It's great as a snack for kids. I'll order it again next time" -> "As a snack for kids

"It's very delicious, so I order it often~ It's also good as a snack." -> "It's good as a snack"

Key Attribute 5: Convenience

"I like it because it's easy to cook. I added more sausage and cooked it, and it's so delicious" -> "I like it because it's easy to cook."

"A really good product that helps busy housewives' daily lives" -> "Helps busy housewives' daily lives"

The main phrase extraction unit 200 clusters the plurality of sub-phrases into a plurality of groups and extracts one main phrase from the plurality of sub-phrases included in each group.

5 is a conceptual diagram showing the entire process of the main phrase extractor 200.

First, the main phrase extraction unit 200 generates an embedding vector for the sub phrase using a BERT (Bidirectional Encoder Representations from Transformers) model.

The BERT model is a natural language processing model that understands context, and is the model that shows the best performance among pre-trained language models learned by various tasks. The BERT model uses a special classification token (CLS), which is a special classification token (CLS) that can contain the information inherent in a sentence or word when used for classification by using the information inherent in a sentence or word unit. It is a learned BERT unique token. The method of using the Special Classification Token (CLS) is as follows. After dividing the natural language data from which you want to extract key sentences into sentence-word units, a special classification token (CLS) is added in front of each sentence (word-word). The special classification token (CLS) can be added to any type of sentence without any additional constraints because the word '[CLS]' can be added to the beginning of the sentence. Then, the vector of special classification tokens (CLS) quantified by the model is used to calculate the degree of similarity between each sentence. That is, the corresponding special classification token (CLS) is a token used as a sentence expression vector when BERT executes a classification task on a sentence-by-sentence basis, and the expression vector of the special classification token (CLS) is used as an embedding vector of sentence tokens. use.

The main phrase extractor 200 may convert each sub-phrase into an embedding vector by using the BERT model, and thus may utilize an embedding vector in which context information other than simple word information is reflected.

Thereafter, the main phrase extractor 200 clusters the plurality of sub-phrases into a plurality of groups based on the plurality of embedding vectors generated for each of the plurality of sub-phrases. At this time, since the number of texts and text content for each product are different due to the nature of text data, the number of clusterings for each product is different.

At this time, the main phrase extraction unit 200 is a group to be clustered using a silhouette score when the number of text data for a specific product is collected at a preset reference value or more, that is, when a sufficiently large amount of text data is collected. optimize the number of

A silhouette score is a value calculated by calculating the distance between each data point and neighboring data points, and is a technique used to determine whether clustered data are appropriately clustered.

For example, assuming that embedding vectors for a total of 10 sub-phrases are divided into 3 clusters, the main phrase extractor 200 extracts the first embedding vectors included in the first cluster and the same cluster as the first embedding vectors (the second cluster). An average distance between other embedding vectors included in cluster 1), an average distance between other embedding vectors included in the second cluster, and an average distance between other embedding vectors included in the third cluster are calculated, respectively. In this way, the main phrase extraction unit 200 calculates average distances for all embedding vectors, and calculates a silhouette score obtained by calculating an average value for the calculated average distances.

The main phrase extraction unit 200 calculates silhouette scores when sub-phrases are clustered by an arbitrary number of groups, and performs clustering with the highest clustering number of calculated silhouette scores.

For example, when the main phrase extractor 200 clusters the sub-phrases into two groups, three groups, four groups, and five groups, the silhouette score for each When it is determined that the silhouette score is the highest when calculated and clustered into five groups, a clustering operation is performed to cluster sub-phrases into five groups.

In some other embodiments, the key phrase extraction unit may assign different weights to silhouette scores calculated for each group according to the number of clustering groups of a specific product and similar products in performing a clustering operation on a specific product.

For example, in performing a clustering task for yukgaejang products, the main phrase extraction unit searches for the number of clusters for other soup products (soybean paste stew, kimchi stew, etc.), calculates the average number of clusters, and calculates the average number of clusters. A higher weight is given to the silhouette score calculated for the cluster most similar to the number of clusters. Accordingly, the main phrase extraction unit can reflect the clustering result for the already verified product to the new product, so that the reliability of data analysis can be improved.

Thereafter, the main phrase extraction unit extracts a sub-phrase closest to the central point in each cluster as a representative phrase of the corresponding cluster, and sets the extracted representative phrase as the main phrase. In the above example, since four groups are clustered, the main phrase extraction unit can extract a total of four main phrases.

On the other hand, when the number of text data for a specific product is less than a preset reference value, the main phrase extractor 200 clusters the plurality of sub-phrases into a plurality of groups according to the preset number of clusters. For example, when the number of text data collected for a specific product is less than 100, the main phrase extraction unit 200 clusters sub-phrases into 4 groups collectively without clustering optimization using the above-described silhouette score.

6 is a diagram showing an example of finally extracted main phrases. In this way, the main phrase extraction unit can represent user text contents in each product and extract a phrase having originality within a product as a main phrase. can

The related product recommendation unit 300 recommends other products related to the main phrase using the main phrase extracted from a specific product.

Conventionally, similar products are recommended based on the user's purchase history or meta information (category). it was only

On the other hand, when the main phrase extraction technique according to the present invention is used, when the user selects the main phrase 'air fryer' extracted from fried dumpling products, the related product recommendation unit does not recommend other dumpling products that are similar products. It is characterized by recommending other products (eg, hot dogs, nuggets, etc.) that can be baked in an air fryer.

To this end, the related product recommendation unit calculates embedding similarity between a main phrase extracted from text data related to a specific product and sub-phrases extracted from text data related to other products.

7 is a diagram showing an example of the average embedding similarity between a specific main phrase extracted from a specific product and sub-phrases extracted from other products, and a related product recommendation unit extracts a product having an average embedding similarity greater than or equal to a preset threshold value; , the extracted product can be recommended as a related product for a specific product.

FIG. 9 is a diagram showing a specific example of related products recommended by the related product recommendation unit. As shown in the figure, among a plurality of key phrases extracted from kimchi cheese rice balls by a user, the main phrase 'when baked in an air fryer' is shown. If you select , you can receive other products related to the selected main phrase as recommended products.

In this way, the related product recommendation unit can recommend more diverse products such as new products and other category products by recommending products related to main phrases rather than products related to the product, and based on the main phrase extracted from the text and sub-phrases of other products You can recommend related products that reflect actual consumer reviews.

9 is a diagram showing a schematic flow of a related product recommendation method using main phrase extraction according to an embodiment of the present invention.

The related product recommendation method using main phrase extraction according to the present invention may be performed by the above-described related product recommendation device using main phrase extraction according to the present invention.

As shown, the related product recommendation method using main phrase extraction according to an embodiment of the present invention includes collecting text data (S10), extracting sub-phrases from text data (S20), and extracting a plurality of extracted phrases. Clustering the sub-phrases into a plurality of groups, extracting one main phrase from the plurality of sub-phrases included in each group (S30), and specifying a specific phrase based on the embedding similarity between the extracted main phrase and sub-phrases of other products. It includes a step (S40) of recommending a product related to the product. Since the main phrase extraction method according to the present invention has been described in the related product recommendation device using the main phrase extraction according to the present invention, repeated description will be omitted.

According to one aspect of the present invention described above, unlike conventional keyword extraction techniques, it is possible to reliably extract a key phrase, which is a phrase in which the characteristics (implication, representativeness, originality) of text data are well reflected, and the extracted By recommending text for related products using key phrases, the user has the advantage of being able to easily check text for related products that are meaningful to him or her.

As described above, the technology according to the present invention may be implemented as an application or implemented in the form of program instructions that can be executed through various computer components and recorded on a computer-readable recording medium. The computer readable recording medium may include program instructions, data files, data structures, etc. alone or in combination.

Program instructions recorded on the computer-readable recording medium may be those specially designed and configured for the present invention, or those known and usable to those skilled in the art of computer software.

Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magneto-optical media such as floptical disks. media), and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like.

Examples of program instructions include high-level language codes that can be executed by a computer using an interpreter or the like as well as machine language codes such as those produced by a compiler. The hardware device may be configured to act as one or more software modules to perform processing according to the present invention and vice versa.

Although the above has been described with reference to embodiments, those skilled in the art will understand that the present invention can be variously modified and changed without departing from the spirit and space of the present invention described in the claims below. You will be able to.

100: text data collection unit
200: main phrase extraction unit
300: Related product recommendation unit

Claims (15)

a text data collection unit that collects text data for each product;
Using a pre-trained Named Entity Recognition (NER) model, a first word corresponding to the beginning of a predefined phrase and at least one second word subsequent to the first word and corresponding to the middle of the phrase a main phrase extraction unit which groups words and extracts them into one sub-phrase, clusters a plurality of the sub-phrases into a plurality of groups, and extracts one main phrase from a plurality of sub-phrases included in each group; and
Includes a related product recommendation unit that recommends related products related to a specific product according to the result of comparing the main phrase extracted for a specific product with sub-phrases extracted for other products,
The main phrase extraction part,
An embedding vector for the sub-phrase is generated using a BERT (Bidirectional Encoder Representations from Transformers) model, and a plurality of the sub-phrases are divided into a plurality of groups based on the plurality of embedding vectors generated for each of the plurality of sub-phrases. cluster, but
When the number of text data for the collection target is collected less than a preset reference value, a plurality of the sub-phrases are clustered into a plurality of groups according to the preset number of clusters;
When the number of text data for the collection target is collected more than a preset reference value, the number of groups to be clustered is dynamically determined using a silhouette score,
The related product recommendation unit,
By calculating the average embedding similarity between the main phrase extracted from a specific product and sub-phrases extracted from other products, sub-phrases having an average embedding similarity higher than a preset threshold are extracted, and the product corresponding to the extracted sub-phrase is specified. A related product recommendation device using key phrase extraction that recommends products as related products.
According to claim 1,
The main phrase is extracted by selecting a sub-phrase that can better represent the characteristics of the collection target from among the extracted sub-phrases compared to other sub-phrases,
Key phrases characterized by being selected by representativeness representing common characteristics of collected data for each collection subject, originality showing differentiation of collected data among collection subjects, and connotation that summarizes the contents of collected data for collection subject A device for recommending related products using extraction.
According to claim 1,
The main phrase extraction part,
An apparatus for recommending related products using main phrase extraction, which extracts sub-phrases of less than the maximum allowable number from text data according to the order of importance of attributes within preset phrases from text data.
According to claim 1,
The main phrase extraction part,
After labeling a first tag to a word corresponding to the beginning of a phrase through the entity name recognition model, labeling a second tag to a word corresponding to the middle of a phrase, and labeling a third tag to a word other than a phrase,
Extract one word labeled with the first tag as a sub-phrase,
Using main phrase extraction, characterized in that a first word labeled with a first tag and at least one second word following the first word and labeled with a second tag are bundled and extracted as one sub-phrase. Related product recommendation device.
delete delete According to claim 4,
The main phrase extraction part,
An apparatus for recommending related products using main phrase extraction, which extracts a representative value from a plurality of embedding vectors included in each group and extracts a sub-phrase corresponding to the extracted representative value as a main phrase.
delete As a related product recommendation method performed by a related product recommendation device using key phrase extraction,
Collecting text data for each product;
Using a pre-trained Named Entity Recognition (NER) model, a first word corresponding to the beginning of a predefined phrase and at least one second word subsequent to the first word and corresponding to the middle of the phrase grouping words and extracting them into one sub-phrase, clustering a plurality of sub-phrases into a plurality of groups, and extracting one main phrase from a plurality of sub-phrases included in each group; and
Comprising a step of recommending a related product related to a specific product according to a result of comparing the extracted main phrase for a specific product with sub-phrases extracted for other products,
Extracting the main phrase,
An embedding vector for the sub-phrase is generated using a BERT (Bidirectional Encoder Representations from Transformers) model, and a plurality of the sub-phrases are divided into a plurality of groups based on the plurality of embedding vectors generated for each of the plurality of sub-phrases. cluster, but
When the number of text data for a specific product to be collected is less than a preset reference value, a plurality of the sub-phrases are clustered into a plurality of groups according to the preset number of clusters;
Characterized in that, when the number of text data for a specific product is collected more than a preset reference value, the number of groups to be clustered is dynamically determined using a silhouette score,
The step of recommending the related product is,
By calculating the average embedding similarity between the main phrase extracted from a specific product and sub-phrases extracted from other products, sub-phrases having an average embedding similarity higher than a preset threshold are extracted, and the product corresponding to the extracted sub-phrase is specified. A related product recommendation method using main phrase extraction, characterized in that the product is recommended as a related product.
According to claim 9,
The step of extracting the main phrase,
A method of recommending related products using main phrase extraction, characterized in that sub-phrases less than the maximum allowable number are extracted from one text data according to the order of importance of attributes within preset passages from one text data.
According to claim 9,
The step of extracting the main phrase,
After labeling a first tag to a word corresponding to the beginning of a phrase through the entity name recognition model, labeling a second tag to a word corresponding to the middle of a phrase, and labeling a third tag to a word other than a phrase,
Extract one word labeled with the first tag as a sub-phrase,
Using main phrase extraction, characterized in that a first word labeled with a first tag and at least one second word following the first word and labeled with a second tag are bundled and extracted as one sub-phrase. How to recommend related products.
delete delete According to claim 9,
The step of extracting the main phrase,
A method for recommending related products using main phrase extraction, characterized in that a representative value is extracted from a plurality of embedding vectors included in each group, and a sub-phrase corresponding to the extracted representative value is extracted as a main phrase.
delete

Images