Recommendation system

Abstract

Providing technology that increases the likelihood of unexpected search results being suggested to users. [Solution] A recommendation system comprising: a classification unit that classifies the content of posts related to a search target into multiple classes; a similarity determination unit that determines that the first search target and the second search target are similar when the similarity of posts related to the first search target and the second search target for each class exceeds a predetermined standard; and a suggestion unit that proposes search targets that match the search conditions specified by the user and search targets that have been determined to be similar to the search targets that match the search conditions.

Description

Technical Field

[0005] , , ,

[0001] The present invention relates to a recommendation system.

Background Art

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In the conventional technology, the user profile indicates the hobbies and orientations of the user, and by proposing products that match the user profile and the product profile, a proposal that conforms to the hobbies and orientations of the user is made. However, products that match the user profile and the product profile are less likely to include products that are unexpected to the user. The present invention has been made in view of the above problems, and an object thereof is to provide a technology that increases the possibility of proposing a search target that is unexpected to the user.

Means for Solving the Problems

[0005] A recommendation system according to one embodiment includes: a classification unit that classifies the content of posts relating to a search target into a plurality of classes; a similarity determination unit that determines that the first search target and the second search target are similar when the similarity of the posts relating to the first search target and the second search target for each class exceeds a predetermined standard; and a suggestion unit that proposes search targets that match the search conditions specified by the user and search targets that have been determined to be similar to the search targets that match the search conditions.

[0006] In other words, search targets are pre-determined to be similar based on the content of the posts. Furthermore, the content of posts in two search targets is classified into multiple classes, and a degree of similarity is specified for each class. Then, two search targets are determined to be similar if the similarity of posts within each class exceeds a predetermined standard. In this configuration, the content of posts can be classified from multiple perspectives, and similarity can be determined from multiple viewpoints. Therefore, compared to configurations that determine similarity based on uniform criteria such as keyword frequency, it is possible to determine that unexpected search targets are similar. Consequently, by suggesting search targets similar to a given search result, the likelihood of users being presented with unexpected search targets can be increased. [Brief explanation of the drawing]

[0007] [Figure 1] A block diagram showing the configuration of the recommendation system. [Figure 2] This figure shows an example of posted information. [Figure 3] Figure 3A shows an example of the results of the classification, and Figure 3B shows an example of the similarity of posts for each class related to two facilities. [Figure 4] This is a flowchart of the similarity detection process. [Figure 5] This is a flowchart of the proposal processing. [Modes for carrying out the invention]

[0008] Here, embodiments of the present invention will be described in the following order. (1) Configuration of the recommendation system: (2) Similarity determination process: (3) Proposal processing: (4) Other embodiments, etc.:

[0009] (1) Configuration of the recommendation system: Figure 1 is a block diagram showing the configuration of the recommendation system 10 according to the present invention. In this embodiment, the recommendation system 10 is composed of a server computer and provides a search service to a user using a user terminal 50 as a client. Users may search for various search targets, but in this embodiment, the search target is a facility.

[0010] In this embodiment, the facility is a building that a user can visit, such as a restaurant or a shopping mall, or other landmark located around a road. The facility accepts visits from anyone and provides various services to visitors. The facility as a search result may be used in any way; for example, it may be set as a destination in the navigation system provided by the user terminal 50, or it may be used to discover a preferred facility by referring to related information about the searched facility, such as reviews, addresses, ratings, etc. This embodiment assumes the latter.

[0011] The recommendation system 10 includes a control unit 20 equipped with a CPU, RAM, ROM, etc., a storage medium 30, and a communication unit 40. The control unit 20 can execute programs stored in the storage medium 30 and ROM. In this embodiment, the recommendation program 21 can be executed as this program. The communication unit 40 is equipped with a circuit for communicating with the user terminal 50, and the control unit 20 can communicate with the user terminal 50 by processing the recommendation program 21.

[0012] The user terminal 50 is a terminal that receives various inputs from the user and outputs various information to the user. The user terminal 50 includes a control unit, a communication unit, a display unit, and an input unit (not shown). The control unit of the user terminal 50 controls the communication unit, the display unit, and the input unit. Through this control, the user terminal 50 can communicate with the recommendation system 10. The user terminal 50 can also display arbitrary information on the display unit. Furthermore, the user terminal 50 can receive various input information based on user input to the input unit.

[0013] In this embodiment, the user can operate the user terminal 50 and input search criteria for facilities. Once search criteria are entered, the user terminal 50 sends the search criteria to the recommendation system 10. The recommendation system 10 retrieves facilities that match the search criteria by referring to the posted information 30a described later. The recommendation system 10 also retrieves similar facilities that are similar to the facilities that match the search criteria. Once facilities that match the search criteria and similar facilities have been retrieved, the recommendation system 10 returns these facilities and similar facilities to the user terminal 50 as search results.

[0014] The user terminal 50 displays the facilities and similar facilities transmitted from the recommendation system 10 on its display unit. As a result, the user can find facilities that match the search criteria. In addition, the similar facilities include facilities that are similar to the facilities that match the search criteria, but are also unexpected. The recommendation system 10 in this embodiment pre-defines the similarity relationships between facilities so that unexpected facilities can be suggested as similar facilities.

[0015] The storage medium 30 of the recommendation system 10 pre-stores post information 30a and LDA model 30b in order to define the similarity relationships between facilities. Furthermore, once the similarity relationships between facilities are defined, similar facility information 30c is stored in the storage medium 30. Post information 30a is text information indicating the content of posts uploaded to social networking services (SNS).

[0016] In this embodiment, text information indicating the content of a post related to a facility is collected in advance. In this embodiment, the post information 30a includes a post ID which is identification information for identifying a post, the name of the facility that the post is about, and text information indicating the content of the post. For a post uploaded to an SNS, other information such as the identification information of the poster, the posting time, the location of the poster, etc. can be associated, but the post information 30a is defined by extracting the post ID, the name of the facility, and the content of the post from these information. FIG. 2 is a diagram showing an example of the post information 30a. As shown in FIG. 2, in the post information 30a, for the post ID of each post, the name of the facility that the post is about and the text indicating the content of the post are associated and defined.

[0017] The LDA model 30b is an LDA (Latent Dirichlet Allocation) model that classifies the post content indicated by the post information 30a into a specified number of classes. The LDA model 30b is machine-learned to classify the topics of each post information 30a into a plurality of classes and identify to which class each post information 30a belongs based on the post information 30a.

[0018] The LDA model 30b is an unlearned model. Learning is performed based on the post information 30a. When the learning is completed, the LDA model 30b can classify the post information 30a into any one of a predetermined number of classes. The predetermined number is not limited and can be any number, but in this embodiment, the predetermined number is 10. That is, the LDA model 30b classifies the input text into a class with a label of any one of 0 to 9.

[0019] In this embodiment, the recommendation system 10 can communicate with an external AI server 60 via the communication of the communication unit 40. A BERT model 60a that has been pre-trained by machine learning is stored in a storage medium (not shown) of the AI server 60. The BERT model 60a is a BERT (Bidirectional Encoder Representations from Transformers) model that identifies the similarity of the post content indicated by the post information 30a. The BERT model 60a is a model that has been trained to output the similarity between two post texts when two post texts are input. An existing model may be reused, or it may be newly generated by machine learning. In this embodiment, when two post texts are input to the BERT model 60a, the similarity between the post texts is output as a numerical value from 0% to 100%. The higher the numerical value, the more similar the texts are. The AI server 60 acquires two post texts transmitted by the control unit 20 via the communication unit 40, inputs them to the BERT model 60a, and returns the obtained similarity.

[0020] The similar facility information 30c is information indicating whether two facilities are similar. Similar facilities may be defined by various methods. In this embodiment, the similar facility information 30c is defined by associating the identification information (e.g., facility name, etc.) of two similar facilities.

[0021] The recommendation system 10 is capable of executing a recommendation program 21 for defining the similarity relationship of facilities and responding to search results for search requests. When the recommendation program 21 is executed, the control unit 20 functions as a classification unit 21a, a similarity determination unit 21b, and a proposal unit 21c. In this embodiment, the control unit 20 performs the processing by the classification unit 21a and the similarity determination unit 21b before the processing by the proposal unit 21c starts, and generates the similar facility information 30c in advance (details will be described later based on FIG. 4).

[0022] The classification unit 21a has the function of classifying the content of posts about facilities into multiple classes. In this embodiment, the control unit 20 uses the function of the classification unit 21a to refer to the post information 30a and obtain the text that represents the content of each post. Then, the control unit 20 uses the function of the classification unit 21a to obtain the LDA model 30b and inputs the text of each post into the LDA model 30b. As a result, the text of each post is classified into one of the classes from 0 to 9. The numerical value indicating the classification result is stored in RAM or the like, associated with the post ID of each post and the name of the facility that was discussed in each post. Figure 3A is an example showing the results of class classification.

[0023] The similarity determination unit 21b has the function of determining that the first facility and the second facility are similar if the similarity of posts for each class related to the first facility and the second facility exceeds a predetermined standard. Specifically, the control unit 20 uses the function of the similarity determination unit 21b to extract any two facilities from the facilities associated with the post information 30a and considers them to be the first facility and the second facility.

[0024] The posts indicated by post information 30a include posts relating to the first facility and the second facility, and each post is classified into one of 10 classes by the classification unit 21a. In this embodiment, the control unit 20 extracts posts relating to each facility that are classified into the same class. The control unit 20 then sends the text of the two extracted posts to the AI ​​server 60 via the communication unit 40. The AI ​​server 60 retrieves the text of the two posts. It inputs the text into the BERT model 60a and obtains the similarity score. The AI ​​server 60 returns the similarity score to the recommendation system 10. The control unit 20 retrieves the similarity score via the communication unit 40. As a result, the class-specific similarity scores of posts relating to the first facility and the second facility are obtained.

[0025] Figure 3B shows an example of the similarity of posts for each class related to two facilities. In Figure 3B, the similarity of each class from 0 to 9 is associated with each combination of two facilities. Note that the similarity of classes 3 to 8 is omitted. For example, for the combination of facility A and facility B, the similarity of class 0 is 70%, the similarity of class 1 is 80%, the similarity of class 2 is 0%, and the similarity of class 9 is 60%. According to this similarity, it can be seen that posts classified as class 0 for facility A and posts classified as class 0 for facility B are similar with a similarity of 70%. In this configuration, similarity is not obtained by comparing posts with different classifications, for example, posts that have no common points in content or topic. Therefore, the possibility of meaningful comparisons can be increased.

[0026] When the similarity of each class is generally high, or when the similarity of some classes is low but the similarity of the remaining classes is high, it means that the content of the posts about the facilities is similar when analyzed from multiple perspectives using 10 classes. For this reason, if the similarity of each class of posts exceeds a predetermined standard, the control unit 20 determines that the first facility and the second facility are similar. Details of this determination will be described later. For two facilities determined to be similar, the facility identification information is associated and stored in the storage medium 30 as similar facility information 30c. The control unit 20 performs the same process for all combinations of two facilities and pre-defines the similar facility information 30c.

[0027] The above processing method differs from configurations that determine similarity from a single perspective, such as keyword frequency, in that it can determine whether facilities are similar from multiple perspectives (10 in this embodiment). Furthermore, in this embodiment, since the similarity of facilities is identified based on posts about the facilities, it is possible to determine whether facilities are similar based on factors other than those that directly indicate the characteristics of the facilities themselves. For example, if a post contains descriptions related to the environment of a facility, such as "quiet place" or "hidden gem," and the facility is classified into a specific class, then facilities with a high similarity of posts in this class may be considered similar facilities. In this case, facilities in completely different genres but with similar environments may be extracted as similar facilities. While similarity determination based on keyword frequency often evaluates whether the main characteristics of the facilities (for example, taste or menu for a restaurant) are similar, this embodiment allows for multifaceted analysis using multiple classes, making it possible for facilities with similar characteristics other than the main characteristics of the facilities to be determined as similar facilities. Therefore, similar facilities that are similar to a given facility may be unexpected for users.

[0028] The suggestion unit 21c has the function of suggesting facilities that match the search criteria specified by the user, and facilities that are determined to be similar to the facilities that match the search criteria. In this embodiment, the control unit 20 obtains the search criteria specified by the user by operating the user terminal 50 via the communication unit 40 using the function of the suggestion unit 21c.

[0029] In this embodiment, the control unit 20 searches for facilities that match the search criteria by referring to the posted information 30a. To this end, the control unit 20 refers to the posted information 30a using the function of the proposal unit 21c and identifies facilities that match the search criteria based on the content of the posts. Various methods may be used to identify facilities that match the search criteria. For example, a configuration can be adopted in which keywords are specified as search criteria, and facilities corresponding to posts containing those keywords are searched. Of course, search criteria may also be specified based on various information such as the genre of the facility, location, nearest station, and business hours. Furthermore, the search method is not limited, and for example, a configuration may be in which posts are identified based on the degree of relevance or similarity to the search criteria or words, and facilities corresponding to those posts are searched. In addition, multiple search results may be output, and in this case, the facilities as search results may be ranked based on the degree of keyword match or frequency, the degree of relevance or similarity to the conditions, etc.

[0030] In addition to the facilities that meet the search criteria obtained as described above, the control unit 20 identifies similar facilities that are similar to the facilities in question. Specifically, the control unit 20 refers to the similar facility information 30c and identifies any facilities that are associated with the identification information of facilities that meet the search criteria as similar facilities.

[0031] Once facilities matching the search criteria and similar facilities are identified, the control unit 20 transmits the facilities matching the search criteria and similar facilities to the user terminal 50 via the communication unit 40 using the function of the suggestion unit 21c. The user terminal 50 receives the transmitted information and displays it as search results on the display unit.

[0032] With the above configuration, the system can suggest facilities that match the search criteria entered by the user, as well as similar facilities that are similar to those facilities. Similar facilities are identified based on a multifaceted analysis of multiple classes of posts, as described above. Therefore, it is possible to increase the likelihood that unexpected search results will be suggested to the user.

[0033] (2) Similarity determination process: Next, the similarity determination process performed by the control unit 20 will be explained with reference to Figure 4. The similarity determination process is performed in advance before the user performs a search. When the similarity determination process is started, the control unit 20 acquires post information using the functions of the classification unit 21a (step S100). The post information acquired here is all the information associated with posts uploaded to the SNS, and includes, for example, the poster's identification information and the posting time. In this embodiment, the post information is stored in the storage medium 30.

[0034] Next, the control unit 20 extracts the facility name and post content using the function of the classification unit 21a (step S105). That is, from each of the post information obtained in step S100, the control unit 20 extracts the post ID, the name of the facility being discussed in the post, and text information indicating the post content. The control unit 20 also associates the facility name and the text information indicating the post content with the post ID and saves it to the storage medium 30 as post information 30a.

[0035] Next, the control unit 20 classifies all posts into multiple classes using the LDA model 30b based on the function of the classification unit 21a (step S110). That is, the control unit 20 refers to the post information 30a generated in step S105, inputs the text information for each post into the LDA model 30b, and classifies each post into one of the multiple classes. As described above, in this embodiment, it is assumed that there are 10 classes, that is, that each post is classified into one of classes 0 to 9.

[0036] Next, the control unit 20 creates a table associating facilities with classes using the functions of the classification unit 21a (step S115). Specifically, the control unit 20 refers to the post information 30a generated in step S105 and creates a table associating the class of the post obtained in step S110 with the facility name that was the topic of the post and the post ID. Figure 3A shows an example of the table generated in step S115. The control unit 20 saves the table to RAM or the like (not shown).

[0037] Next, the control unit 20 sets the unprocessed facility combinations as processing targets using the function of the similarity determination unit 21b (step S120). In this embodiment, two facilities are extracted from a plurality of facilities, and the loop processing of steps S120 to S135 is performed on the combination of these two facilities as the processing target. Therefore, in step S120, the control unit 20 extracts one combination obtained by extracting two facilities from a plurality of facilities that is not the processing target of the loop processing of steps S120 to S135, and sets it as the processing target.

[0038] Next, the control unit 20 uses the function of the similarity determination unit 21b to obtain the similarity of posts from the facilities to be processed that are classified into the same class (step S125). Specifically, in step S115, posts relating to each of the two facilities constituting the combination to be processed are classified into one of 0 to 9, for example, as shown in Figure 3A. As shown in Figure 3A, there may be multiple posts corresponding to facility A. Similarly, there may be multiple posts corresponding to facility B.

[0039] Each post is classified into one of the following categories from 0 to 9. Therefore, the control unit 20 identifies posts that are classified into the same class among the posts corresponding to facility A and facility B. For example, in the example shown in Figure 3A, if the processing targets are facility A and facility B, the post with post ID 0003 and the post with post ID 0105 are classified into the same class. Therefore, the control unit 20 identifies the post with post ID 0003 and the post with post ID 0105 as posts of the same class.

[0040] The control unit 20 then sends the text information of the two identified posts to the AI ​​server 60. The AI ​​server 60 inputs the text information of the two posts into the BERT model 60a to obtain the similarity score and returns it to the control unit 20. The control unit 20 obtains the similarity score of the two posts via the communication unit 40. If there are multiple posts classified into the same class for the same combination of facilities, the control unit 20 obtains the similarity score for each post. For example, in the example shown in Figure 3A, if there are posts from facility A classified into class 9 other than post ID 0003, the similarity score between that post and the posts classified into class 9 from facility B is obtained. If multiple similarity scores for posts classified into the same class are obtained, the control unit 20 obtains statistical values ​​(e.g., the mean) of these similarity scores. The control unit 20 performs the above processing for each class. In this way, the similarity scores of posts corresponding to the combination of facilities being processed are obtained for all classes, and then, for example, the class-specific similarity scores shown in Figure 3B are obtained.

[0041] Next, the control unit 20, using the function of the similarity determination unit 21b, excludes classes with a similarity of 0% and obtains the average value of the remaining similarities (step S130). Specifically, the control unit 20 refers to the similarity obtained in step S125 and does not include classes with a similarity of 0% in the evaluation. That is, the control unit 20 excludes classes with a similarity of 0% from the parameter when obtaining the average. Then, the control unit 20 obtains the sum of the similarities of the remaining classes and divides it by the number of remaining classes to obtain the average value of the similarities. Here, the value obtained by excluding classes with a similarity of 0% and averaging the similarities is called the similarity statistic. In Figure 3B, the similarity statistic obtained by the above process is shown to the right of the similarity of class 9.

[0042] Next, the control unit 20 determines whether processing has been completed for all combinations using the function of the similarity determination unit 21b (step S135). That is, it determines whether the processing in steps S120 to S135 has been performed on all combinations obtained by extracting two facilities from multiple facilities. If it is determined in step S135 that processing has not been completed for all combinations, the control unit 20 repeats the processing from step S120 onward.

[0043] On the other hand, if it is determined in step S135 that processing has been completed for all combinations, the control unit 20 uses the function of the similarity determination unit 21b to save combinations whose similarity statistics are equal to or greater than the threshold as similar facilities (step S140). That is, the control unit 20 compares the similarity statistics of all combinations obtained in the loop processing of steps S120 to S135 with the threshold. The threshold is a value that has been set in advance to determine whether or not the facilities are similar.

[0044] The control unit 20 then determines that combinations of facilities are similar if their similarity statistics are above a threshold. For example, in the example shown in Figure 3B, if the threshold is 70%, the combination of facility A and facility B is determined to be similar. On the other hand, the combination of facility A and facility C, and the combination of facility B and facility C are not determined to be similar.

[0045] The similarity statistics can be calculated using various methods. In this embodiment, the similarity statistics are calculated by excluding classes with a similarity of 0% and using the average of the remaining classes. If these statistics are equal to or greater than a threshold, a predetermined standard is exceeded, and the two facilities are determined to be similar. Through this process, it is possible to determine whether facilities are similar or not based on posts analyzed from multiple perspectives across multiple classes using a simple process. The control unit 20 identifies the identification information of the two facilities determined to be similar and stores the identification information of both facilities in the storage medium 30 as similar facility information 30c, associating them with each other.

[0046] (3) Proposal processing: Next, the suggestion process performed by the control unit 20 will be explained with reference to Figure 5. The suggestion process is performed when the user operates the user terminal 50 and enters search conditions. When the user enters search conditions, the user terminal 50 sends a search request to the recommendation system 10 along with information indicating those search conditions.

[0047] The control unit 20 of the recommendation system 10 receives the search request via the communication unit 40 using the function of the suggestion unit 21c and starts the suggestion process. Once the suggestion process starts, the control unit 20 acquires the search conditions using the function of the suggestion unit 21c (step S200). Next, the control unit 20 acquires facilities that match the search conditions using the function of the suggestion unit 21c (step S205). That is, the control unit 20 refers to the posted information 30a and identifies facilities that match the search conditions.

[0048] Next, the control unit 20 acquires facilities similar to the acquired facility using the function of the proposal unit 21c (step S210). That is, the control unit 20 refers to the similar facility information 30c and determines whether or not there is a facility associated with the identification information of the facility acquired in step S205. If there is a facility associated with the identification information of the facility acquired in step S205, the control unit 20 acquires that facility as a similar facility.

[0049] Next, the control unit 20 proposes the facilities acquired in steps S205 and S210 (step S215). That is, the control unit 20 transmits information indicating these facilities to the user terminal 50 via the communication unit 40. The user terminal 50 acquires this information and displays the facilities indicated by the information as search results on the display unit. With this configuration, the possibility of unexpected search targets being proposed to the user can be increased.

[0050] (4) Other embodiments, etc.: The above embodiments are merely examples for carrying out the present invention, and various other embodiments can be adopted. For example, the recommendation system 10 may be composed of multiple devices. Also, for example, at least a part of the classification unit 21a, similarity determination unit 21b, and suggestion unit 21c in the recommendation system 10 may be implemented in another device, for example, a user terminal 50. Furthermore, some of the configurations of the above embodiments may be omitted, and the order of processing may be changed or omitted. Furthermore, the database for searching the search target may be different from the posted information 30a. For example, facilities may be searched based on a database provided by a facility review site.

[0051] The classification unit only needs to be able to classify the content of posts related to the search target into multiple classes. That is, the classification unit only needs to be able to classify posts with similar characteristics into the same class and posts with different characteristics into different classes. The classification method is not limited to LDA, and various machine learning models can be used. In addition, various methods such as cluster analysis and discriminant analysis may be used. Furthermore, there can be multiple classes, and the number can be any number. It may also be variable depending on the user, etc.

[0052] The search target is not limited to facilities; any object can be included in the search. For example, fruits, fashion items, dishes, or any other items that may be included in a post can be searched. Of course, various types of information, such as genre and coordinates, can be associated with the search target in the database. This information can be used as search criteria to search for the search target.

[0053] Posts can be any text related to the search criteria, and may also be text generated from posted images or audio. Furthermore, the platform on which posts are made is arbitrary and not limited to social media. For example, posts on various websites such as review sites, or posts to publishers or newspapers, are also acceptable.

[0054] The similarity determination unit only needs to be able to determine that the first search target and the second search target are similar if the degree of similarity for each class of posts relating to the first search target and the second search target exceeds a predetermined standard. In other words, the similarity determination unit only needs to be able to determine that the first search target and the second search target are similar if the results of classifying the posts for each of the first and second search targets show similarity, and the degree of similarity exceeds a predetermined standard.

[0055] The first and second search targets may be any two selected from a set of multiple search targets. The similarity of posts by class may be determined by any method. Therefore, the configuration is not limited to using BERT, and other machine learning models may be used, as well as various text analysis methods, such as frequency analysis of similar words.

[0056] The predetermined criteria can be any criteria used to determine whether search targets are similar to each other, based on the similarity of posts in each of multiple classes. Therefore, the predetermined criteria are not limited to a configuration defined by whether the similarity statistics exceed a threshold. For example, the predetermined criteria could be defined by whether the number of classes whose post similarity exceeds a threshold exceeds a default value. Furthermore, the statistics are not limited to the arithmetic mean, but could be a weighted average, etc. In addition, the classes that are not considered for similarity determination are not limited to classes with a similarity of 0%, but could be, for example, classes below a threshold that are not considered for similarity determination.

[0057] The suggestion function should be able to suggest search targets that match the search criteria specified by the user, and search targets that are determined to be similar to those that match the search criteria. In other words, it should be configured to suggest the search results desired by the user by searching for search targets that match the search criteria, and at the same time, to be able to make unexpected suggestions by suggesting similar search targets to the search targets obtained as search results. When suggesting search targets that are determined to be similar to search targets that match the search criteria, the suggested search targets may be limited by various methods. For example, various configurations may be adopted, such as a configuration that suggests the search target with the highest degree of similarity among similar search targets, or a configuration that suggests a predetermined number of search targets in order of highest similarity.

[0058] Furthermore, the methods of the present invention can also be applied as programs and methods. Moreover, such systems, programs, and methods may be implemented as standalone devices, or they may be implemented using parts shared with other devices, encompassing various embodiments. They can also be modified as appropriate, such as being partly software and partly hardware. Furthermore, the invention also functions as a recording medium for a program that controls the system. Of course, the recording medium for the program may be a magnetic recording medium, a semiconductor memory, or any recording medium developed in the future can be considered in exactly the same way. [Explanation of symbols]

[0059] 10...Recommendation system, 20...Control unit, 21...Recommendation program, 21a...Classification unit, 21b...Similarity determination unit, 21c...Proposal unit, 30...Storage medium, 30a...Posting information, 30b...LDA model, 30c...Similar facility information, 40...Communication unit, 50...User terminal, 60...AI server, 60a...BERT model

Claims

1. A classification unit that categorizes the content of posts related to the search target into multiple classes, A similarity determination unit determines that the first search target and the second search target are similar if the similarity of the posts for each class with respect to the first search target and the second search target exceeds a predetermined standard. A proposal unit that proposes search targets that match the search criteria specified by the user, and search targets that are determined to be similar to the search targets that match the search criteria. A recommendation system equipped with [features / features].

2. The similarity determination unit, If the statistical value of the similarity of each class of the aforementioned posts exceeds a threshold, it is deemed to have exceeded the predetermined standard. The recommendation system according to claim 1.

3. The search target is a facility. The recommendation system according to claim 1 or claim 2.

Images