Information processing system, information processing method and program

Abstract

To provide a scheme that generates a search keyword using a generative AI.SOLUTION: An information processing system converts each character string constituting a first data group to be a search target into a distributed representation using a distributed representation model with the character string as input and the distributed representation as output, that has been learned using the first data group or a second data group different from the first data group as learning data, inputs a condition input by a user into a generative AI to obtain a keyword related to the condition as a generative AI keyword, obtains the distributed representation corresponding to the generative AI keyword using the distributed representation model, obtains, as a similar keyword, a character string from among the character strings constituting the first data group whose distributed representation has a similarity degree of a predetermined value or more compared to the distributed representation of the generative AI keyword, and searches the first data group using the generative AI keyword and the similar keyword to extract a corresponding search target.SELECTED DRAWING: Figure 7

Description

[Technical Field]

[0001] The present invention relates to an information processing system, an information processing method, and a program. [Background technology]

[0002] Japanese Patent Laid-Open Publication No. 2022-000822 (Patent Document 1) is a document disclosing background technology in this technical field. This publication states that "A marketing server SV collects purchase information for each user from multiple customer information databases DB1 to DBn, and creates household account book information for each user, income group, and regional area based on this collected purchase information. Then, based on the household account book information for each income group and the household account book information for each regional area, an income group and a regional area are selected in which the consumption ratio of a product or service of interest among a group of products or services with the same expenditure item is equal to or greater than a threshold, and from a set of users belonging to either one or both of these selected income groups and regional areas, users in which the consumption ratio of a product or service of interest among a group of products or services with the same expenditure item is equal to or greater than a threshold are selected as approach target users" (see Abstract). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2022-000822 Summary of the Invention [Problem to be solved by the invention]

[0004] In the invention described in this publication, users whose consumption ratio of a product or service of interest is equal to or exceeds a threshold are selected based on household account book information. In other words, a search for users is performed using the product or service of interest as a search key. However, this publication does not anticipate using a generation AI to generate the search key when performing this search. The present invention has been made in consideration of these circumstances, and provides a mechanism for generating search keywords using a generation AI. [Means for solving the problem]

[0005] In order to solve the above problems, for example, the configurations described in the claims are adopted. The present application includes multiple means for solving the above problem, and one example is an information processing system comprising: a conversion means for converting each string constituting the first data group into a distributed representation using a distributed representation model that takes strings as input and outputs distributed representations and that has been trained using a first data group to be searched or a second data group different from the first data group as training data; a first acquisition means for inputting conditions entered by a user into a generation AI and acquiring keywords related to the conditions as generation AI keywords; a second acquisition means for using the distributed representation model to acquire distributed representations corresponding to the generation AI keywords; a third acquisition means for acquiring, as similar keywords, strings from among the strings constituting the first data group whose distributed representations have a similarity of a predetermined value or more compared to the distributed representation of the generation AI keyword; and an extraction means for searching the first data group using the generation AI keywords and the similar keywords and extracting corresponding search targets. [Effects of the Invention]

[0006] According to the present invention, a mechanism for generating search keywords using a generation AI can be provided. Problems, configurations, and effects other than those described above will become apparent from the following description of the embodiments. [Brief explanation of the drawings]

[0007] [Figure 1] FIG. 1 shows an example of the configuration of an information retrieval system 100 . [Figure 2] FIG. 2 shows an example of the configuration of the information search server 101. [Figure 3] FIG. 3 shows an example of the configuration of the generation AI server 102. [Figure 4] FIG. 4 shows an example of the configuration of the user terminal 103. [Figure 5] FIG. 5 shows an example of the details DB 221. [Figure 6] FIG. 6 shows an example of a distributed representation generation processing flow. [Figure 7] FIG. 7 shows an example of a search processing flow. DETAILED DESCRIPTION OF THE INVENTION

[0008] 1. Example Hereinafter, an embodiment of the present invention will be described with reference to the drawings. Overview In this embodiment, household account book data is used to extract a set of users having a specified arbitrary lifestyle. The lifestyles required for a business change depending on the situation. Therefore, in this embodiment, users with any lifestyle specified by the analyst are extracted using a large-scale language model (LLM). This allows for quick response to changing business conditions. Specifically, in this embodiment, LLM generates detail keywords related to lifestyles and searches the detail database.

[0009] However, there is an issue with this method: the keywords generated by LLM do not always match the description of the detailed data, and may not be found when searching the detailed database. For example, suppose the target lifestyle is "pet lover" and the LLM output keyword is "Pet Shop J." In this case, the keyword does not match the description in the detailed database, "Pet Specialty Store J."

[0010] Furthermore, even if the keywords match the detailed description, the keywords may not be comprehensive enough, and users with the desired lifestyle may not be extracted. For example, suppose the target lifestyle is "child-rearing households" and the LLM output keywords are "baby goods store D" and "baby goods store E." In this case, there are many stores and services related to child-rearing, such as child-rearing stores, and it cannot be said that the keywords have been sufficiently listed.

[0011] In this embodiment, in consideration of these circumstances, a distributed representation of specific information in the search population (for example, store names and product names in household ledger data) is created, and the keywords generated by LLM are expanded using this distributed representation. This complements keywords that are more consistent with the representation of specific information in the search population. In addition, the comprehensiveness of the keywords is improved. The distributed representation here refers to a word expressed as a vector, more specifically, a word expressed as a high-dimensional real vector.

[0012] 1-2.Configuration FIG. 1 shows an example of the configuration of an information retrieval system 100 . The information retrieval system 100 is an information processing system comprising an information retrieval server 101, a generation AI server 102, and multiple user terminals 103, all of which are connected via a network. The network may be wired or wireless, and each device can send and receive information via the network.

[0013] Here, the information search server 101 is a server that uses household account book data to extract a set of users having an arbitrary lifestyle designated by the user. The generation AI server 102 is a server managed by a company that provides services using the generation AI. The user terminal 103 is a terminal used by a user who uses the search service provided by the information search server 101 .

[0014] Each device constituting the information retrieval system 100 may be, for example, a portable terminal (mobile terminal) such as a smartphone, tablet, mobile phone, or personal digital assistant (PDA), or may be a wearable terminal such as glasses, a wristwatch, or clothing. Each device may also be a stationary or portable computer, or a server located on the cloud or a network. Each device may also function as a VR (Virtual Reality) terminal, an AR (Augmented Reality) terminal, or an MR (Mixed Reality) terminal. Alternatively, each device may be a combination of multiple of these terminals. For example, a combination of one smartphone and one wearable terminal can logically function as a single terminal. Other information processing terminals may also be used.

[0015] Each device constituting the information retrieval system 100 includes a processor that executes an operating system, applications, programs, etc., a main storage device such as a RAM (Random Access Memory), an auxiliary storage device such as an IC card, a hard disk drive, an SSD (Solid State Drive), or a flash memory, a communication control unit such as a network card, a wireless communication module, or a mobile communication module, an input device such as a touch panel, a keyboard, a mouse, voice input, or input based on motion detection captured by a camera unit, and an output device such as a monitor or a display. The output device may also be a device or terminal that transmits information to be output to an external monitor, display, printer, or device.

[0016] The main memory stores various programs, applications, etc. (modules), and the processor executes these programs and applications to realize the various functional elements of the overall system. These modules may be implemented in hardware, such as by integration. Each module may be an independent program or application, or may be implemented as a subprogram or function within a single integrated program or application.

[0017] In this specification, each module is described as the entity (subject) that performs the processing, but in reality, the processing is carried out by a processor that processes various programs, applications, etc. (modules). Various databases (DBs) are stored in the auxiliary storage device. A "database" is a functional element (storage unit) that stores a set of data so that it can accommodate any data manipulation (e.g., extraction, addition, deletion, overwriting, etc.) from a processor or an external computer. There are no limitations on how the database is implemented; for example, it can be a database management system, spreadsheet software, or a text file such as XML or JSON.

[0018] 1-1-1. Information Search Server 101 FIG. 2 shows an example of the configuration of the information search server 101. The information search server 101 is configured, for example, by one or more servers arranged on a cloud. The server's main memory device 201 stores programs and applications such as a model learning module 210, a distributed representation generation module 211, a keyword generation module 212, a keyword expansion module 213, a search module 214, and a specific processing module 215. These programs and applications are executed by the processor 203 to realize each functional element of the information search server 101. Each module will be described below.

[0019] The model learning module 210 trains the distributed representation model 220. This distributed representation model 220 is a trained model that is trained using a first data group to be searched as training data, and takes strings as input and outputs distributed representations. This distributed representation model 220 is a model that outputs distributed representations based on the co-occurrence between strings that make up the first data group. In this embodiment, the purchase history of a user group corresponds to the first data group, and the detail strings that make up the purchase history correspond to the strings input to the distributed representation model 220.

[0020] When training the distributed representation model 220, the module may extract, from the character strings constituting the first data group, character strings whose appearance frequency or ranking in appearance frequency is equal to or greater than a predetermined value, and use the extracted character strings as training data to train the distributed representation model 220. This makes it possible to remove words unrelated to the search criteria in advance, which is efficient in terms of performance in the search process.

[0021] The distributed expression generation module 211 converts each character string constituting the first data group to be searched into a distributed expression using the distributed expression model 220. In this embodiment, the purchase history of a user group corresponds to the first data group, and the detailed character strings constituting the purchase history correspond to the character strings to be converted into a distributed expression.

[0022] The keyword generation module 212 inputs the conditions input by the user into the generation AI, and obtains keywords related to the conditions as generation AI keywords.

[0023] The keyword expansion module 213 obtains a distributed representation corresponding to the generated AI keyword using the distributed representation model 220. In doing so, the module identifies, from among the character strings constituting the first data group to be searched, character strings that have a similarity to the generated AI keyword of a predetermined value or more, and converts the identified character strings into distributed representations using the distributed representation model 220. Note that a character string that has a similarity to the generated AI keyword of a predetermined value or more here refers, for example, to a character string whose edit distance from the generated AI keyword is a predetermined value or less. The module also acquires, as similar keywords, strings from among the strings that make up the first data group to be searched, whose distributed representations have a similarity of a predetermined value or more compared to the distributed representation of the generated AI keyword. In this embodiment, the purchase history of a user group corresponds to the first data group, and the detail character strings that make up the purchase history correspond to the character strings that make up the first data group.

[0024] The search module 214 searches the first data group to be searched using the generated AI keyword and similar keywords, and extracts the corresponding search target. In this embodiment, the purchase history of the user group corresponds to the first data group, and the user corresponds to the search target. When searching the first data group, the module can extract search targets corresponding to multiple keywords, thereby preventing errors in the extraction of search targets due to incorrect search keywords.

[0025] The identification processing module 215 performs identification processing on the extracted search target. For example, if the extracted search target is a user of a "child-rearing household," this identification processing involves providing information on child-rearing benefits.

[0026] Next, the auxiliary storage device 202 of the information search server 101 will be described. The auxiliary storage device 202 stores a distributed representation model 220, a detailed DB 221, distributed representation information 222, and a token set 223. Of these, the distributed representation model 220 is a distributed representation model that has been trained by the model learning module 210.

[0027] The statement DB 221 is information equivalent to the purchasing history of a group of users (in other words, household account book data). Figure 5 shows an example of the statement DB 221. The statement DB 221 shown in the figure is made up of records each consisting of fields for user ID, date, statement string, and amount. Of these fields, the user ID is identification information for the user who purchased the product or service. The date is the date of purchase. The statement string is a string representing the name of the store where the product or service was purchased, or the name of the purchased product or service. The amount is the purchase amount.

[0028] The distributed expression information 222 is a collection of distributed expressions generated by the distributed expression generation module 211. Each distributed expression is associated with a corresponding token.

[0029] The token set 223 is a set of tokens generated in the distributed expression generation process described later. Each token is associated with a user ID.

[0030] 1-1-2.Generation AI Server 102 FIG. 3 shows an example of the configuration of the generation AI server 102. The generation AI server 102 is configured, for example, by one or more servers placed on the cloud. The server's main memory device 301 stores programs and applications such as the generation processing module 310. The processor 303 executes these programs and applications to realize the various functional elements of the generation AI server 102.

[0031] The generation processing module 310 of the server inputs the prompt sent from the information retrieval server 101 into the generation AI and obtains an answer. The module then returns the obtained answer to the information retrieval server 101.

[0032] Here we will explain the generation AI. Generative AI (or generative AI) is a deep learning model that has been trained on a large dataset. In other words, the generative AI is a large-scale language model (LLM). Examples of large-scale language models include GPT-3, GPT-4 ("GPT" is a registered trademark), and BERT. Since large-scale language models are created by training deep learning models in advance using large datasets, they can be used without training data or additional training. In particular, this generation AI is a document generation AI that outputs text in response to input text.

[0033] Next, the auxiliary storage device 302 of the generation AI server 102 will be described. The auxiliary storage device 302 temporarily stores a prompt 320 and answer information 321. Of these, the prompt 320 is a prompt sent from the information search server 101. On the other hand, the answer information is answer data generated as a result of the generation process.

[0034] 1-1-3. User terminal 103 FIG. 4 shows an example of the configuration of the user terminal 103. The user terminal 103 is, for example, a smartphone, a tablet terminal, a notebook PC, a desktop PC, or the like. The main memory device 401 of the terminal stores programs and applications such as a browser module 410. The processor 403 executes these programs and applications to realize the various functional elements of the user terminal 103.

[0035] The browser module 410 of the terminal communicates with the information search server 101, acquires a screen corresponding to the user's operation, and displays it on the display. The auxiliary storage device 402 of the terminal stores the screen data 420 sent from the information search server 101 .

[0036] 1-2.Operation 1-2-1. Distributed Representation Generation Processing Next, the distributed expression generation process executed by the information search server 101 will be described with reference to Fig. 6. Fig. 6 shows an example of the distributed expression generation process flow. Flow 600 shown in the figure shows the process flow for generating distributed expressions for each detail string in the detail DB 221.

[0037] First, the model learning module 210 divides the detail strings into tokens for all records in the detail DB 221 (step 601). Here, a token is a text component such as a word, and is generated using a morphological analysis engine such as MeCab. For example, this splitting process splits the statement string "Convenience store A, store X" into [Convenience store A, store X]. As another example, the statement string "Product X, Online store C" would be split into [Product X, Online store C].

[0038] Here, the number of occurrences of each token in the statement DB 221 may be tallied, and only the top N most frequent tokens may be extracted. Since the purpose of this embodiment is to determine the user's lifestyle, removing words unrelated to this in advance will improve performance in subsequent processing. For example, a token such as "XX store" is not particularly important information in determining lifestyle. Since this token such as "XX store" is a unique store name, it is expected to occur less frequently than a chain store name such as convenience store A, and is therefore expected to be excluded.

[0039] Next, the model learning module 210 compiles the token sequences for each user (step 602). Below is an example of the token sequences compiled for each of the user IDs "0001" and "0002." 0001: [Convenience store A, XX store, Electric power company I, January payment] 0002: [[Baby Products], Disposable Diapers B, Online Store C, Baby Goods Store D, Baby Goods Store E] The collected token sequence is stored in the secondary storage device 202 as a token set 223 .

[0040] Next, the module trains the distributed representation model 220 on the token sequence for each user (step 603). This training uses a method called item2vec. Item2vec is a model that embeds items in a latent space based on the co-occurrence relationship between users and items. Item2vec assumes that all items previously purchased by a user are considered a purchase sequence, and that a given item can be predicted from other items in the same purchase sequence. Then, each item vector is trained to increase the dot product of the target item vector and the surrounding item vectors. The trained distributed representation model 220 is stored in the auxiliary storage device 202.

[0041] Next, the distributed representation generation module 211 converts all tokens generated in step 601 into embedding vectors using the distributed representation model 220 (step 604). Here, an embedding vector is a numerical vector that represents the meaning of a word. An example of conversion to an embedding vector is shown below. Convenience store A => [0.8, 0.1, 0.2] Baby Goods Store D => [0.1, 0.2, 0.9] The generated embedding vectors are stored in the auxiliary storage device 202 as distributed representation information 222 .

[0042] This concludes the description of the distributed expression generation processing flow 600. By executing this flow, the distributed expression information 222 is generated.

[0043] 1-2-2. Search process Next, the search processing executed by the information search server 101 will be described with reference to Fig. 7. Fig. 7 shows an example of a search processing flow. Flow 700 shown in the figure shows the flow of processing for extracting users with an arbitrary lifestyle using the distributed representation model 220.

[0044] First, the keyword generation module 212 accepts a lifestyle specification from a user operating the user terminal 103 (step 701). Here, the lifestyle specification to be accepted is assumed to be "child-rearing household."

[0045] Next, the module uses the generation AI server 102 to obtain the names of stores that child-rearing households frequently visit and the names of products that they purchase (step 702).At that time, the module sends an instruction (in other words, a prompt) to the generation AI server 102 instructing it to output the names of stores that child-rearing households frequently visit and the names of products that they purchase.

[0046] The generation processing module 310 of the generation AI server 102 inputs the prompt sent from the information retrieval server 101 into the generation AI and obtains keywords (hereinafter referred to as "LLM output keywords") (not shown in the figure). Then, the module returns the obtained LLM output keywords to the information retrieval server 101 (not shown in the figure). As a result, the keyword generation module 212 of the information retrieval server 101 obtains the LLM output keywords. Here, the following are assumed as the LLM output keywords to be obtained. LLM output keywords: Baby Supplies Store E, Baby Supplies Store D, Diaper Type B A Here, "Baby Supplies Store E" and "Baby Supplies Store D" are examples of store names, and "Diaper Type B A" is an example of a product name.

[0047] Next, the keyword expansion module 213 uses the distributed representation model 220 to obtain the embedding vectors corresponding to each LLM output keyword. Specifically, the module first extracts one token from the token set 223 that best matches the literal representation of each LLM output keyword (step 703). At that time, the module calculates the edit distance between the LLM output keyword and each token and selects the token with the smallest edit distance (hereinafter referred to as the "corresponding token"). The edit distance (Levenshtein edit distance) mentioned here is defined as the minimum number of operations (insertion, deletion, replacement) required to make two strings the same. Note that instead of the edit distance, the corresponding token may be selected based on other similarities such as the Jaro-Winkler distance.

[0048] Examples of pairs of LLM output keywords and corresponding tokens are shown below. <LLM output keyword> <corresponding token> Baby Supplies Store D Baby Supplies Store D Baby Supplies Store E Baby Supplies Store E Diaper Type B A Diaper Type B

[0049] The module then converts each corresponding token into an embedding vector (step 704) using the distributed representation model 220. An example of the conversion is shown below. Baby Goods Store D => [0.1, 0.2, 0.9]

[0050] Next, the module uses the distributed representation model 220 to acquire keywords that are semantically similar to the LLM output keywords (hereinafter referred to as "similar keywords") (step 705). In doing so, the module refers to the distributed representation information 222, calculates the similarity between the embedding vector of the corresponding token generated in step 704 and the embedding vectors of all tokens, and selects tokens with high similarity. Note that the similarity here is an index that indicates how closely two vectors point in the same direction, such as cosine similarity. Token selection methods include, for example, selecting the top n similarities, or selecting all tokens for which similarity > a threshold.

[0051] Below are examples of corresponding tokens and similar tokens. Baby goods store D[0.1, 0.2, 0.9] => Toy store F[0.1, 0.1, 0.9], Baby clothes store G[0.1, 0.2, 0.8] Disposable diaper B [0.1, 0.7, 0.9] => Disposable diaper H [0.2, 0.7, 0.8]

[0052] Next, the search module 214 searches the statement DB 221 using the LLM output keyword and similar keywords as search keywords, and extracts users whose purchase history includes the keywords (step 706). Here, the extracted users may be limited by a condition such as score > threshold, where the number of search keyword hits or the total amount corresponding to the search keywords is used as a score. The score can be flexibly set based on the search keywords that are hit. Below, an example of the number of search keyword hits and the total amount corresponding to the search keywords is explained.

[0053] First, assume the following keywords as search keywords: Baby goods store D, baby goods store E, disposable diaper B type A Baby clothes store G, toy store F, disposable diaper store H

[0054] The following purchase history of the extracted user is assumed. [Table 1]

[0055] In this case, the number of search keyword hits is 3, and the total amount corresponding to the search keywords is 3,500 yen (2,000 + 500 + 1,000).

[0056] However, errors in search keywords may occur due to errors in the LLM output or in the expansion using distributed representations. In this regard, using scores based on multiple search keywords can prevent errors in user extraction due to incorrect search keywords. An example of error prevention is described below.

[0057] First, assume the following keywords as search keywords: Baby goods store D, baby goods store E, disposable diaper B type A Baby clothes store G, toy store F, disposable diaper store H, Pet Shop J Among these search keywords, it is assumed that the incorrect keyword is "Pet Shop J."

[0058] Below is an example of the purchase history of a user who is not a child-rearing household. [Table 2]

[0059] In this case, the number of search keyword hits will be 2, and the number of unique search keyword hits will be 1 (the only search keyword that will hit is "Pet Shop J"). Therefore, for example, by extracting users with unique search keyword hits of ≥ 2, it is possible to prevent incorrect user extraction even if there is one incorrect search keyword. This is based on the idea that the reliability of user extraction is low when there is only 1 hit, so the results can only be trusted when there are 2 or more hits.

[0060] Finally, the identification processing module 215 performs identification processing on the extracted users (step 707). For example, the module provides information suitable for the extracted users, such as information on child-rearing benefits to users in the "child-rearing household" category. The above is the description of the search processing flow 700.

[0061] According to the embodiment described above, a distributed representation of specific information in a search population (for example, store names or product names in household ledger data) is created in advance, and the keywords generated by LLM are expanded using the distributed representation. This complements keywords that are more consistent with the representation of specific information in the search population. In addition, the comprehensiveness of keywords is improved.

[0062] 2. Variations The above embodiment may be modified as follows: The following modifications may be combined with each other.

[0063] (1) Search population In the above embodiment, purchase history is assumed as the search population, but other behavioral history may also be used as the search population. For example, website browsing history or search history may also be used as the search population. In this case, the extraction target is users who have viewed or searched for websites that match the conditions entered by the user.

[0064] (2) Search conditions In the above example, "lifestyle" (for example, "child-rearing household") is assumed as a search condition, but this is only one example of a search condition. Other search conditions such as lifestyle habits, hobbies, and preferences may also be set.

[0065] (3) How to convert LLM output keywords into embedding vectors In the above search processing flow 700, the following processing is performed to obtain an embedding vector corresponding to an LLM output keyword. From the token set 223, extract one corresponding token whose character expression best matches the LLM output keyword (step 703). Convert the corresponding tokens into embedding vectors using the distributed representation model 220 (step 704).

[0066] Instead of these processes, the LLM output keywords may be directly converted into embedding vectors using the distributed representation model 220. Even with this method, it is possible to obtain embedding vectors corresponding to the LLM output keywords.

[0067] (4) Distributed Representation Model In the above embodiment, the distributed representation model 220 is trained using purchase history, which is a search population, as training data. However, this distributed representation model 220 can also be trained using text information (second data group) that is not a search population, such as that published on the web. In this case, training may be performed using a technique called word2vec. In this case, the distributed representation model 220 is a trained model that is trained using text information as training data, takes strings as input, and outputs distributed representations. This distributed representation model 220 is a model that outputs distributed representations based on the co-occurrence between words in each sentence. However, although the distributed representation model 220 trained in this manner can increase the comprehensiveness of the LLM output keywords, since it is not trained on the search population, the consistency of the notation with the search population is reduced compared to the embodiment. Note that word2vec is merely an example, and other methods (such as Sentence BERT) can also be used to train a distributed representation model. Sentence BERT is a method for training a model by providing similarity between text pairs as training data in addition to co-occurrence between words.

[0068] (5)Functional layout In the above-described information retrieval system 100, the generation AI function is provided on a server different from the information retrieval server 101. However, this functional arrangement is merely an example. The function may also be provided on the information retrieval server 101. The functions of each server in this system may be arranged as appropriate according to the environment in which the information retrieval service is provided.

[0069] (6) Other The present invention is not limited to the above-described embodiments and includes various modifications. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to those including all of the described configurations. Furthermore, it is possible to replace part of the configuration of one embodiment with the configuration of another embodiment, or to add the configuration of another embodiment to the configuration of one embodiment. Furthermore, it is possible to add, delete, or replace part of the configuration of each embodiment with other configurations.

[0070] Furthermore, the above-described configurations, functions, processing units, processing means, etc. may be partially or entirely implemented in hardware, for example, by designing them as integrated circuits. The above-described configurations, functions, etc. may also be implemented in software, with a processor interpreting and executing a program that implements each function. Information such as the programs, tables, and files that implement each function can be stored in a memory, a recording device such as a hard disk or SSD (Solid State Drive), or a recording medium such as an IC card, SD card, or DVD.

[0071] In addition, the control lines and information lines shown are those that are considered necessary for the explanation, and do not necessarily show all the control lines and information lines in the product. In reality, it can be assumed that almost all components are interconnected. The above-described embodiments disclose at least the configurations described in the claims. [Explanation of symbols]

[0072] 100...information search system, 101...information search server, 102...generation AI server, 103...user terminal

Claims

1. a conversion means for converting each character string constituting the first data group into a distributed representation using a distributed representation model that receives character strings as input and outputs distributed representations and that has been trained using a first data group to be searched or a second data group different from the first data group as training data; a first acquisition means for inputting conditions input by a user into the generation AI and acquiring keywords related to the conditions as generation AI keywords; a second acquisition means for acquiring a distributed representation corresponding to the generated AI keyword using the distributed representation model; a third acquisition means for acquiring, as a similar keyword, a character string from among the character strings constituting the first data group, the character string whose distributed representation has a similarity of a predetermined value or more compared with the distributed representation of the generated AI keyword; an extraction means for searching the first data group using the generated AI keyword and the similar keyword and extracting corresponding search targets; An information processing system comprising:

2. The second acquisition means acquires a distributed representation corresponding to the generated AI keyword, Identifying character strings that have a similarity of a predetermined value or more compared with the generated AI keyword from among the character strings that make up the first data group; Converting the specified character string into a distributed representation using the distributed representation model.

2. The information processing system according to claim 1, wherein:

3. 2. The information processing system according to claim 1, wherein said extraction means extracts search targets corresponding to a plurality of keywords when searching said first data group.

4. further comprising a learning means for learning the distributed representation model; The learning means extracting character strings from the character strings constituting the first data group whose frequency of appearance is equal to or greater than a predetermined value or whose ranking of frequency of appearance is equal to or greater than a predetermined value; The extracted character strings are used as training data to train the distributed representation model.

2. The information processing system according to claim 1, wherein:

5. 2. The information processing system according to claim 1, wherein the first data group is a purchasing history of a user group.

6. 2. The information processing system according to claim 1, wherein the distributed representation model is a model that outputs distributed representations based on co-occurrence between character strings that make up the first data group.

7. The information processing system according to claim 1 , wherein the distributed representation model is a model trained on the second data group that is not a search target.

8. 1. A computer-implemented information processing method, comprising: converting each character string constituting the first data group into a distributed representation using a distributed representation model that takes character strings as input and outputs distributed representations and that has been trained using a first data group to be searched or a second data group different from the first data group as training data; A step of inputting the conditions input by the user into the generation AI and obtaining keywords related to the conditions as generation AI keywords; Using the distributed representation model, obtain a distributed representation corresponding to the generated AI keyword; acquiring, as similar keywords, character strings from among the character strings constituting the first data group, the character strings whose distributed representations have a similarity of a predetermined value or more compared with the distributed representation of the generated AI keyword; searching the first data group using the generated AI keyword and the similar keyword to extract corresponding search targets; An information processing method including:

9. On the computer, converting each character string constituting the first data group into a distributed representation using a distributed representation model that takes character strings as input and outputs distributed representations and that has been trained using a first data group to be searched or a second data group different from the first data group as training data; A step of inputting the conditions input by the user into the generation AI and obtaining keywords related to the conditions as generation AI keywords; Using the distributed representation model, obtain a distributed representation corresponding to the generated AI keyword; acquiring, as similar keywords, character strings from among the character strings constituting the first data group, the character strings whose distributed representations have a similarity of a predetermined value or more compared with the distributed representation of the generated AI keyword; searching the first data group using the generated AI keyword and the similar keyword to extract corresponding search targets; A program to execute.

Images