JP6802334B1 - Matching system and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a matching system capable of accurately reflecting the characteristics and needs of a matching target person and reducing the burden on a service person in charge. SOLUTION: A matching system 10 creates combined text data by combining self-data and needs data of a large number of matching subjects, and uses these combined text data to estimate topics by soft clustering or a neural language model. Matching is performed by executing processing to obtain the topic distribution (topic value indicating the appearance probability of each topic in the combined text data of each matching target person), and using the obtained topic distribution to perform machine learning using a supervised classification model. A model is constructed, matching is performed between the applicant and each candidate by the matching model, and a score indicating the probability that each candidate is selected is output. [Selection diagram] Fig. 1

Description

The present invention relates to a matching system and program composed of a computer that executes a matching process that combines natural persons, corporations, non-corporate organizations, or other matching target persons, and introduces, for example, a company that is a business partner. It can be used for business matching that combines each other.

Conventionally, in the business partner referral service for corporations (hereinafter referred to as business matching service), when the service staff selects referral candidates, the service staff searches on a keyword basis and matches from the search results. A method based on keyword matching to select a partner was often used. At this time, the search target is a database of company information (information indicating one's own characteristics) and needs information described by each corporation in advance.

As a system for performing business matching, the company information of the company to which the employee belongs and the company information of the company managed by the manager are stored, and the employee or manager in the website displayed on the communication terminal. Based on the company information that is stored in association with the employees and managers who acquired the action log of the company, the information providing system that performs company matching to select the company to be introduced to the company that opened the website It is known (see Patent Document 1).

Further, when the present invention is applied to a business matching service, machine learning can be performed by using a performance list of a company pair selected as a candidate for a business partner in the past by a service person. From this point of view. As a related technique of the above, there is known an information processing device that provides a certain evaluation of a person who wants to be hired based on objective data and reduces the labor of a person in charge of hiring (see Patent Document 2). This information processing device has a first evaluation department that generates first evaluation information for each factor determined in advance based on a document created by the applicant for employment for the target company, and information on the applicant for employment and the first evaluation. It is equipped with a second evaluation department that generates second evaluation information of applicants for employment based on the first evaluation information from one evaluation department, and the second evaluation department has an artificial intelligence function that learns based on past performance data. have. The past performance data includes recruitment applicant information regarding past recruitment applicants, first evaluation information for the past recruitment applicants, and selection results for the past recruitment applicants.

Further, as a system for matching job seekers and job seekers, a human resources matching device that performs ideal and efficient human resource matching using presentation videos of job seekers and job seekers is known (Patent Document 3). reference). This human resources matching device acquires a job seeker video which is a presentation video by a job seeker and a job seeker video which is a presentation video by a job seeker, and provides the job seeker video and the job seeker video to the job seeker and the job seeker. At the same time, obtain follow-up information notifying that the job seeker or job seeker is interested in the job seeker or job seeker, and based on the follow-up information of both the job seeker and the job seeker, the job seeker and the job seeker Matching.

Further, in the present invention, a scoring model that outputs a score indicating the probability or certainty that each candidate is selected for a matching applicant is constructed, but as a related technique from this viewpoint, specific requirements are satisfied. A matching device that matches a human resource that searches for an organization that meets the requirements with an organization that searches for a human resource that meets a specific requirement is known (see Patent Document 4). In this matching device, a matching score indicating the relationship between human resources information and organizational information is calculated.

Publication of Patent No. 6369966 JP-A-2018-116710 Publication of Patent No. 6480077 Japanese Unexamined Patent Publication No. 2015-164022

As described above, in the conventional business matching service, a method of selecting a matching partner based on a keyword match is often used, but this method has the following problems.

Since it is necessary to search after converting company information and matching needs into specific keywords, conversion work to keywords is required, which takes time and effort, and the burden on service personnel is heavy. There was a problem.

In addition, the conversion work to this keyword relies heavily on know-how and is a personal work, so the work content varies among service personnel, that is, if the service personnel changes, the work content There was a problem that the selection results would be different.

Further, when the matching needs are converted into keywords, there is a problem that information is missing and a desired search result may not be obtained.

Then, there is a problem that the output result may be significantly different due to the change of the search keyword, which tends to lead to opportunity loss.

In addition, the results obtained by the search (selected matching partner candidates) do not have a quantitative index, and it is necessary to treat multiple search results (multiple companies selected as matching partner candidates) in the same line. Therefore, there is a problem that the burden of finding a desired partner is heavy. That is, at the stage where the service person narrows down the candidates, or the company that wants matching (the requesting company that receives the business matching service) narrows down the final candidates from the candidates presented by the service person. Since there is no quantitative index at any of the steps to be performed, there is a problem that it takes time and effort.

In addition, the above problems are not limited to business matching services that connect companies, but can also be said when various matching services are provided, such as individuals and individuals, individuals and groups, and groups and groups. is there. A service in which individual-to-individual, individual-to-group, and group-to-group matching is mixed includes, for example, a seller-buyer matching service in real estate sales. In addition, matching between individuals is, for example, matching between persons seeking specialized knowledge and specialists (lawyers, tax accountants, etc.), marriage hunting support services, etc., and matching between individuals and groups is, for example, employment. It is a support service for job hunting and recruiting activities that connects applicants and hiring companies, and matching between groups is, for example, a support service for business succession, a support service for finding opponents for practice games of baseball teams, etc. ..

An object of the present invention is to provide a matching system and a program that can accurately reflect the characteristics and needs of a matching target person and can reduce the burden on a service person.

The present invention is roughly classified into a case where the topic model including the applicant data for the matching applicant is not relearned (for example, in the case of FIG. 8 described later) and a topic model including the applicant data. Re-learning (for example, in the case of FIG. 9 described later) and matching for a specific purpose among existing matching target persons (for example, in the case of FIG. 10 described later). There is a basic configuration.

<Basic configuration of the present invention when the topic model including the applicant data is not relearned (for example, in the case of FIG. 8 described later)>

The present invention is a matching system configured by a computer that executes a matching process that combines natural persons, corporations, non-corporate organizations, or other matching target persons.
Needs data consisting of self-data consisting of text data describing the characteristics of each of a large number of matching target persons, not including those who wish to match as a new matching target person, and text data describing the needs of the matching partner. Appearance probability of each topic in the combined text data of each matching target person obtained by executing topic estimation processing by soft clustering or neural language model using a large number of combined text data obtained by combining and A topic model storage means that stores the topic value indicating the above in association with the identification information of the matching target person and stores the appearance probability of each word in each topic obtained by executing the topic estimation process.
An applicant data storage means that stores self-data and needs data about a matching applicant in association with the identification information of the applicant, and
A combined text data creating means that combines self-data and needs data of a matching applicant stored in the applicant data storage means and executes a process of creating the combined text data of the applicant.
Appearance of each topic in the applicant's combined text data using the applicant's combined text data created by this combined text data creation means and the appearance probability of each word in each topic stored in the topic model storage means. Applicant topic value prediction means that executes the process of predicting the topic value indicating the probability,
The topic value indicating the appearance probability of each topic in the combined text data of the applicant obtained by this applicant topic value prediction means, and the matching target person as each candidate to be the matching partner stored in the topic model storage means. A matching feature data creation means that executes a process of creating matching feature data for a plurality of pairs of applicants and candidates using a topic value indicating the appearance probability of each topic in the combined text data. ,
Using each of the plurality of matching feature data created by this matching feature data creating means as input data, a matching model that has been learned in advance by two-class classification of whether or not to be selected and stored in the matching model storage means is used. The feature is that it is equipped with a matching processing means that executes a processing that outputs a score indicating the probability that each candidate is selected for a matching applicant by performing a classification process using a supervised classification model. Is what you do.

Here, the "matching target person" includes a natural person, a corporation, a non-corporate organization, and other persons. Non-corporate organizations include so-called voluntary organizations, such as pre-establishment companies, neighborhood associations, political organizations, condominium management associations, circles, academic societies, local baseball teams and soccer teams. Therefore, for example, a matching process for searching for an opponent in a practice game such as a baseball team or a soccer team is also an object of the present invention. In addition, non-corporate organizations include organizations within the corporation, such as departments within the corporation (departments, sections, staff, factories, offices, laboratories within the university, faculties, departments, etc.). Furthermore, "matching targets" include national and local governments.

Also, for example, when looking for a business partner, it is usually a match between companies, so it is a match between a group (whether it is a corporation or not) and a group, but one or both are individually managed. In the case of offices, etc., it is also possible to match individuals (natural persons) with groups, and individuals with individuals. Business matching services also include non-profit organizations such as universities and their laboratories, and therefore industry-academia collaboration support services. It also includes non-profit organizations such as universities and their laboratories, and therefore support services for finding partners in collaborative research.

In addition to business matching, for example, matching between sellers and buyers of real estate sales, the successor company (the company that transfers the business) and the successor company (the company that takes over the business) in the business succession support service There are matching with, matching between those who seek specialized knowledge and specialists (lawyers, tax accountants, etc.), matching with marriage hunting support services, etc. These matching services include matching between individuals (natural persons) and individuals. , Matching between individuals and groups (whether or not they are corporations), matching between groups and groups, and matching of these types are the same as in the case of business matching. Mixed matching is, for example, the case where both sellers and buyers of real estate sales include individuals and groups.

In addition, matching between individuals (natural persons) and groups (whether or not they are corporations) includes, for example, matching between job seekers and hiring companies in job hunting and recruiting support services, or personnel affairs. There is matching between employees and departments within the company in the transfer support service. Therefore, the matching target persons to be matched may be not only those who are placed in an equal relationship but also those who are placed in a master-slave relationship.

Furthermore, the "matching target person" of "others" includes a person who can be regarded as a person in matching, for example, a robot or some animals, and in short, a text describing one's own characteristics and needs. Anyone who can prepare the data will do. At this time, the description of the text data does not necessarily have to be done by oneself, and may be represented by another person (human). Therefore, as long as he / she has his / her own characteristics and needs, the "matching" of the present invention Can be the target person. Therefore, the "person" of the "matching target person" is not limited to humans or groups of humans (whether or not they are corporations). For example, matching between humans and housekeeper robots, The present invention can also be applied to matching between humans and pet robots, matching between robots, matching between humans and animals for pets, matching between animals and zoos, matching between animals and trainers, and the like.

Further, the "description of the needs for the matching partner" regarding the "needs data" may be a description of the matters required of the matching partner, or a description indicating the desired future state of the matching partner, and may be a description directly or indirectly to the matching partner. Any description may be directed to the target. Therefore, the description may be centered on the actions of the matching partner, or may be a description centered on one's own actions.

Further, the "score indicating the certainty" in the "matching processing means" is a score indicating the degree or degree of certainty, and may be a value of likelihood or probability (0 to 1, 0 to 100%), or is likely. Not limited to the value of degree and probability, the value or result after threshold processing and the value or result after some filtering may be output. In short, a score indicating the degree or degree may be output. It is a purpose, it may be a continuous numerical value, and it may be a stepwise index (for example, "95% or more, 90% or more and less than 95%, 85% or more and less than 90%, ...", or "Very large, large, slightly large". , Medium, slightly small, small, very small "). Then, the "matching processing means" may be configured to perform secondary processing such as arranging the output scores in descending order.

In such a matching system of the present invention, self-data of a large number of matching subjects and needs data are combined to create combined text data, and these combined text data are used by soft clustering or a neural language model. Execute topic estimation processing to obtain topic distribution (topic value indicating the appearance probability of each topic in the combined text data of each matching target person), and use the obtained topic distribution to perform machine learning using a supervised classification model. Since a matching model is constructed by the matching model and the matching applicant and each candidate are matched by the matching model, manual keyword conversion work becomes unnecessary, and it is possible to eliminate personal operations. ..

For this reason, the time and effort required for the conventional keyword conversion work can be saved, and the load on the service staff can be reduced. In addition, since work that involves individual judgment based on know-how is eliminated, it is possible to eliminate variations in work content among service personnel, and it is possible to provide uniform services.

Furthermore, since machine learning of the matching model is performed using the topic distribution obtained by the topic estimation process, the feature amount that accurately and multifacetedly captures the registration information (self-data, needs data) of the matching target person. (Characteristic data for matching) can be created and appropriate scoring can be performed. Therefore, the lack of information can be suppressed. Therefore, in the case of conventional keyword matching, when the matching needs are converted into keywords, information may be lost and the desired search result may not be obtained. However, such a situation should be avoided. Is possible. Then, in the conventional method, there is a problem that the output result may be significantly different due to the change of the search keyword, which tends to lead to a loss of opportunity, but such a problem is also solved.

In addition, since it is possible to perform scoring of the matching accuracy between the applicant and each candidate by using all the registered matching target persons as candidates for the matching applicant, it is unlikely to fall into an opportunity loss. In addition, it is possible to browse an arbitrary number of information on matching target persons having a high possibility of operation by referring to the high and low scores. A matching target person with high operability is, for example, a matching partner who is likely to actually jointly, collaborate, or cooperate in business in the case of business matching, and may be a matching of real estate sales. For example, it is a matching partner who is likely to actually start sales negotiations or conclude a sales contract, and in the case of marriage-hunting matching, it is a matching partner who is likely to actually associate or get married.

Furthermore, since all the registered matching target persons can be selected as candidates and scoring can be performed mechanically, even if the registration of matching target persons increases, the burden on the service staff will increase accordingly. It becomes possible to avoid it.

In addition, when creating feature data for matching, the topic distribution (topic value indicating the appearance probability of each topic) obtained by executing topic estimation processing by soft clustering or neural language model is used, so applicants and applicants As the attribute data of each candidate (applicant data and matching target data of each candidate), information corresponding to the topic (for example, in the case of business matching, the type of industry (business field) can be considered) is prepared. Since it is not necessary to keep it, the burden on the service person can be reduced in this respect as well, and the above-mentioned purpose is achieved by these points.

<Basic configuration of the present invention when the topic model is relearned including the applicant data (for example, in the case of FIG. 9 described later)>

The present invention is a matching system configured by a computer that executes a matching process that combines natural persons, corporations, non-corporate organizations, or other matching target persons.
Needs consisting of self-data consisting of text data describing the characteristics of each of a large number of matching target persons, not including those who wish to be matched as a new matching target person, and text data describing the needs of the matching partner. A matching target person data storage means that stores data in association with the identification information of the matching target person,
An applicant data storage means that stores self-data and needs data about a matching applicant in association with the identification information of the applicant, and
The self-data and needs data of each matching target person stored in the matching target person data storage means are combined, and the self-data and needs data of the applicant stored in the applicant data storage means are combined. , A combined text data creation means that executes the process of creating combined text data for a large number of matching targets, including applicants.
Using the combined text data of a large number of matching targets including the applicant created by this combined text data creation means, each topic in the combined text data of each matching target including the applicant by soft clustering or a neural language model. A topic estimation means that executes topic estimation processing to obtain a topic value indicating the appearance probability of
The topic value indicating the appearance probability of each topic in the combined text data of each matching target person including the applicant obtained by the topic estimation process by this topic estimation means is associated with the identification information of each matching target person including the applicant. Topic model memory means to memorize,
The applicant and each candidate are stored using the topic value indicating the appearance probability of each topic in the combined text data of the applicant and the matching target as each candidate to be the matching partner stored in this topic model storage means. Matching feature data creation means that executes the process of creating matching feature data for a plurality of pairs consisting of
Using each of the plurality of matching feature data created by this matching feature data creating means as input data, a matching model that has been learned in advance by two-class classification of whether or not to be selected and stored in the matching model storage means is used. The feature is that it is equipped with a matching processing means that executes a processing that outputs a score indicating the probability that each candidate is selected for a matching applicant by performing a classification process using a supervised classification model. Is what you do.

Here, the meanings of "description of needs for matching partner" for "matching target person" and "needs data" and "score indicating accuracy" in "matching processing means" are the topics including the above-mentioned <applicant data. This is the same as the basic configuration of the present invention when the model is not retrained.

In such a matching system of the present invention, the same actions and effects as the above-mentioned <basic configuration of the present invention when the topic model including the applicant data is not relearned> can be obtained.

<Basic configuration of the present invention when matching for a specific purpose is performed among existing matching target persons (for example, in the case of FIG. 10 described later)>

The present invention is a matching system configured by a computer that executes a matching process that combines natural persons, corporations, non-corporate organizations, or other matching target persons.
Self-data consisting of text data describing the characteristics of each of a large number of matching target persons including the specific target person as an existing matching target person who performs matching for a specific purpose, and text data describing the needs for the matching partner. Each of the combined text data of each matching target person obtained by performing topic estimation processing by soft clustering or a neural language model using a large number of combined text data obtained by combining the needs data consisting of A topic model storage means that stores a topic value indicating the appearance probability of a topic in association with the identification information of the matching target person, and
Using the topic value indicating the appearance probability of each topic in the combined text data of each matching target person stored in this topic model storage means, from the specific target person and each candidate to be the matching partner of this specific target person. A plurality of pairs, a plurality of pairs consisting of each of the plurality of specific target persons and each candidate to be a matching partner of the plurality of specific target persons, from each of the specific target person and the other plurality of specific target persons. A matching feature data creation means that creates a plurality of matching feature data for a plurality of pairs, or creates a single matching feature data for a pair of specific target persons.
Each of the plurality of matching feature data created by this matching feature data creating means is used as input data, or one matching feature data is used as input data, and the two-class classification of whether or not to be selected is learned in advance. A score indicating the probability that each candidate or another specific target person will be selected for a specific target person by performing classification processing by a supervised classification model using the matching model stored in the matching model storage means. It is characterized by being provided with a matching processing means for executing a process of outputting.

Here, the meanings of "description of needs for matching partner" for "matching target person" and "needs data" and "score indicating accuracy" in "matching processing means" are the topics including the above-mentioned <applicant data. This is the same as the basic configuration of the present invention when the model is not retrained.

In addition, "specific target person as an existing matching target person who performs matching for a specific purpose" is different from a person who desires matching as a new matching target person, and already has matching target person data (attribute data of the matching target person). Is the matching target when the topic estimation process including the data is performed and the topic model is constructed, and for a specific purpose, with other existing matching targets including that person. It is a person who needs to perform matching.

Furthermore, the "specific purpose" is the same as when a matching request is made by a person who wishes to match as a new matching target person, and an existing matching target person (specific target person) causes another existing matching target. It may be a standard specific purpose of responding to a request for matching with a person, but it also includes the following special specific purpose.

For example, for a plurality of existing matching target persons (for example, 10 companies, 10 people, etc.) requested, the same number of candidates suitable for holding a group meeting, etc. from among the existing matching target persons. There is a specific purpose of choosing (for example, 10 companies, 10 people, etc.). In this case, a plurality of existing matching target persons (for example, 10 companies, 10 persons, etc.) who are clients are a plurality of specific target persons, and each candidate is an existing matching target other than the plurality of specific target persons. Is a person. A group meeting, etc. is not a one-to-one combination, but a many-to-many combination, and outputs scores for a large number of one-to-one combinations (pairs) in order to obtain an appropriate many-to-many combination. Will be done. More specifically, for example, if there is a noun exchange meeting setting corresponding to a group match between companies, or if there is a master-slave relationship, for example, a plurality of employers (for example, music production) and a plurality of applicants. For example, when setting up a joint audition with a musician or the like. It should be noted that selecting the same number of candidates is an example, and the number is not limited to the same number. The ratio between the number of existing matching targets who are the clients and the number of each candidate to be selected. Is arbitrary, and either number may be large.

It also includes a specific purpose of finding the matching probability (the probability of matching each other) for a pair of specific target persons and investigating the compatibility. For example, two companies, company A and company B, are determined as targets for obtaining the accuracy of matching, and data is obtained as to whether or not those companies A and company B are in a relationship capable of conducting transactions, collaborative work, etc. When there is a request to make an objective judgment based on the above, the score between company A and company B may be output. In this case, the request may be brought in by either company A or company B, or may be a joint request by both company A and company B. In both cases, company A and company B are specific targets. is there.

Furthermore, it also includes a specific purpose of obtaining the accuracy of matching for a plurality of pairs consisting of a specific target person and each of a plurality of other specific target persons, and investigating their compatibility. For example, company A and its matching partners B, C, and D are determined as targets for determining the accuracy of matching, and the score between company A and company B and the score between company A and company C are determined. The score, the score between the company A and the company D may be output. In this case, it is usually the company A that brings in the request, but the companies A, B, C, and D are all specific target persons. In addition, company C may request the calculation of the score for the pair (pair not including the client) between company A and company B, and investigate their compatibility.

In such a matching system of the present invention, the same actions and effects as the above-mentioned <basic configuration of the present invention when the topic model including the applicant data is not relearned> can be obtained, and the existing matching can be obtained. It is possible to respond to various requests by the target person for a specific purpose.

<Structure that uses the product of topic values and the absolute value of the difference between topic values as synthetic variables>

In addition, in the matching system described above,
The means for creating feature data for matching is
Using each topic value of both pairs, find a composite variable consisting of the product of both topic values for the same topic and a composite variable consisting of the absolute value of the difference between both topic values for the same topic. It is desirable that the configuration is such that processing is performed using each of the obtained composite variables as matching feature data.

When the product of topic values and the absolute value of the difference between topic values are used as synthetic variables in this way, both of the matching target persons (both applicants and candidates, specific target persons) to be paired. Appropriate scoring can be performed compared to the case where the topic distribution (each topic value) of both the candidate and the candidate or both the specific target person and the other specific target person is used as it is as the matching feature data. It will be possible.

That is, if the topic distributions of both pairs are used as they are as the matching feature data, the learning data may be biased at the time of learning, and biased learning may be performed. More specifically, for example, there is a possibility that a matching model (scoring model) that gives a relatively high score to a pattern that appears frequently (here, a topic distribution that appears frequently) is constructed in the training data. On the other hand, by generating a new synthetic variable that shows the characteristics of the pair of matching subjects, it is possible to avoid biased learning.

It is considered that the product of the topic values of both pairs represents the degree of overlap of the topic distributions of both pairs, and the absolute value of the difference between the topic values of both pairs represents the difference / spread of the topic distributions of both pairs.

<Structure that includes the inner product or cosine similarity of the TFIDF vector or other word importance vector in the matching feature data>

Furthermore, in the matching system described above,
A word importance vector calculation means that executes a process of calculating a word importance vector consisting of a TFIDF value or other word importance index values of each word for the combined text data of each matching target person.
The word importance vector storage means for storing the word importance vector of the combined text data calculated by the word importance vector calculation means in association with the identification information of the matching target person is provided.
The means for creating feature data for matching is
Find and find the inner product or cosine similarity between the word importance vector for one combined text data of the pair stored in the word importance vector storage means and the word importance vector for the other combined text data of the pair. The inner product of the word importance vectors or the cosine similarity may be included in the matching feature data to execute the process.

Here, the calculation target of the word importance vector by the "word importance vector calculation means" is, if there is a person who wants to match as a new matching target person, the combination of each matching target person including the person who wants to match. It becomes text data.

When the inner product or cosine similarity of the TFIDF vector or other word importance vector is included in the matching feature data in this way, the feature included in the combined text data is further emphasized by the word importance. It becomes possible to create matching feature data, and it becomes possible to perform more appropriate scoring.

Further, the word importance vectors of both pairs can be used as they are as feature data for matching, but here, the inner product of the word importance vectors of both pairs or the cosine similarity is obtained, and the inner product of the obtained word importance vectors is obtained. Alternatively, since the cosine similarity is used as the matching feature data, the number of dimensions of the matching feature data can be reduced, and the composite variable is used instead of being used as it is as in the case where the composite variable is generated by the topic distribution described above. By generating it, it becomes possible to perform more appropriate scoring.

<Structure that includes the needs flag in the matching feature data>

In addition, in the matching system described above,
Self-data and needs data are accompanied by at least one need flag indicating the type of need for the matching partner.
The means for creating feature data for matching is
The needs flags of both pairs may be included in the matching feature data to execute the process.

Here, the "at least one need flag" is stored in the matching target person data storage means in association with the identification information of the matching target person together with the self-data and the needs data of each matching target person for the existing matching target person. You should keep it. In addition, if there is a person who wishes to match as a new matching target person, store it in the applicant data storage means in association with the identification information of the applicant together with the self-data and needs data of the applicant for matching. Just do it.

When the needs flag is included in the matching feature data in this way, both pairs (both applicants and candidates, both specific target persons and candidates, both specific target persons and other specific target persons). Since the type of needs of) is reflected in the matching feature data, more appropriate scoring can be performed.

<Structure that includes synthetic variables based on needs flags in matching feature data>

Then, in the case where the above-mentioned needs flag is included in the matching feature data,
The means for creating feature data for matching is
Using both pair needs flags, the logical sum, logical product, exclusive OR, negative logical sum, negative logical product, negative exclusive OR, arithmetic sum, or other composite variable of both needs flags. It may be configured to execute a process of obtaining and including the obtained composite variable in the matching feature data.

When the composite variable based on the needs flag is included in the matching feature data in this way, the composite variable is generated instead of using the needs flag as it is, as in the case of generating the composite variable with the topic distribution described above. It becomes possible to perform more appropriate scoring.

<Structure for business matching>

In addition, in the matching system described above,
Matching targets are natural persons, corporations, non-corporate organizations, or other matching targets who do business.
The matching process is a business matching process that combines matching target persons who conduct business.
The topic estimation means can be configured to execute the topic estimation process by the late tent Dirichlet allocation.

In the case of the configuration for performing business matching in this way, the type of industry (business field) can be assumed as each topic obtained by the topic estimation process, so the matching characteristics that accurately reflect the characteristics of the combined text data. Data can be created and appropriate scoring can be performed. In other words, when performing business matching, since there are many business-related descriptions in the combined text data that combines self-data and needs data, it is possible to assume the type of business (business field) as a topic. -The number of topics specified in Dirichlet Allocation (LDA) can be set to the number of industries (business fields) to be classified. Therefore, it is easy to match the design image of the system builder with the result of the topic estimation process by LDA, so that it is possible to perform an appropriate system design. In addition, it does not matter what kind of industry (business field) each topic obtained by LDA corresponds to, there is no clear correspondence, there is no need to know the correspondence, and the topic It suffices if a distribution can be obtained.

<Invention of the program>

The program of the present invention is for operating the computer as the matching system described above.

The above program or a part thereof may be, for example, an optical magnetic disk (MO), a compact disk (CD), a digital versatile disk (DVD), a flexible disk (FD), a magnetic tape, or a read-only memory (ROM). , Electrically erased and rewritable read-only memory (EEPROM), flash memory, random access memory (RAM), hard disk drive (HDD), solid state drive (SSD), flash disk, etc. It can be stored and distributed, and for example, local area network (LAN), metropolitan area network (MAN), wide area network (WAN), Internet, intranet, extranet, etc. It is possible to transmit using a transmission medium such as a wired network, a wireless communication network, or a combination thereof, and it is also possible to carry it on a carrier. Further, the above program may be a part of another program, or may be recorded on a recording medium together with a separate program.

As described above, according to the present invention, a matching model is constructed by performing machine learning by a supervised classification model using a topic distribution obtained by executing topic estimation processing by soft clustering or a neural language model. However, the matching model is used to match applicants for matching with each candidate, or to match specific target persons with each candidate or between specific target persons, so that the characteristics and needs of the matching target person can be determined. There is an effect that it is possible to create matching feature data that accurately reflects it and perform appropriate scoring, and it is possible to reduce the burden on the service staff.

The whole block diagram of the matching system of one Embodiment of this invention. The figure which shows the specific example of the matching target person data stored in the matching target person data storage means of the said embodiment. The figure which shows the specific example of the data from the creation of the combined text data of the said embodiment to the estimation of a topic distribution. The explanatory view of the method of creating the synthetic variable to be included in the matching feature data of the said embodiment. The explanatory view of the method of creating the characteristic data for matching at the time of operation of the said embodiment. The explanatory view of the preparation method of the feature data for matching at the time of learning of the said embodiment. The figure of the flowchart which shows the flow of the learning process performed in advance of the said embodiment. The figure of the flowchart which shows the flow of processing at the time of operation when the topic model including the applicant data of the said embodiment is not relearned. The figure of the flowchart which shows the flow of the process at the time of operation at the time of re-learning a topic model including the applicant data of the said embodiment. The figure of the flowchart which shows the flow of processing at the time of operation in the case of performing matching for a specific purpose in the existing matching target person of the modified form of this invention.

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 shows the overall configuration of the matching system 10 of the present embodiment. This matching system 10 is a business matching service that introduces a business partner (including industry-academia collaboration) to a business operator (mainly a company, but may be an individual or a non-profit organization such as a university). It is a system for implementing.

FIG. 2 shows a specific example of the matching target person data stored in the matching target person data storage means 42, and FIG. 3 shows a specific example of the data from the creation of the combined text data to the estimation of the topic distribution. It is shown. Further, FIG. 4 shows a method of creating synthetic variables to be included in the matching feature data, FIG. 5 shows a method of creating matching feature data during operation, and FIG. 6 shows a method of preparing matching feature data during learning. The explanation of each is shown. Further, FIG. 7 shows a flow of learning processing performed in advance, FIG. 8 shows a flow of processing during operation when the topic model including the applicant data is not relearned, and FIG. 9 shows the applicant. The flow charts show the flow of processing during operation when re-learning the topic model including data.

<Overall configuration of matching system 10>

In FIG. 1, the matching system 10 is composed of a computer and includes a main body 20, a display means 70 such as a liquid crystal display, and an input means 80 such as a mouse, a keyboard, and a touch panel.

The main body 20 includes a processing means 30 that executes various processes related to matching, a desired person data storage means 41 that stores various data required for processing by the processing means 30, a matching target person data storage means 42, and unnecessary word dictionary storage. The means 43, the topic model storage means 44, the word importance vector storage means 45, the matching model storage means 46, the learning data storage means 47, and the matching candidate selection result list storage means 48 are included.

The processing means 30 includes an input receiving means 31, a combined text data creating means 32, a topic estimating means 33, a desired person topic value predicting means 34, a word importance vector calculating means 35, and a matching feature data creating means 36. , The matching processing means 37, the learning means 38, and the learning data preparation means 39.

Here, each of the means 31 to 39 included in the processing means 30 is a central processing unit (CPU) provided inside the main body 20 of the matching system 10, and one or a plurality of means defining the operation procedure of the CPU. Realized by the program. Further, as the storage means 41 to 48, for example, a hard disk drive (HDD), a solid state drive (SSD), or the like can be adopted. The applicant data storage means 41 may only be formed on the main memory (although it may be a volatile memory), but after scoring, the applicant data is used as newly registered matching target data. , Together with the matching target person data registered up to that point, it is stored and registered in the matching target person data storage means 42 which is a non-volatile memory. Therefore, the applicant data storage means 41 and the matching target person data storage means 42 are described as separate storage means for convenience of explanation, but since the data formats are the same, they can be combined into the same storage means. Good.

Further, although the matching system 10 is described in a stand-alone configuration in FIG. 1, it may be a system configuration in which communication is performed via a network. In the case of the server / client type matching system 10, the main body 20 may be composed of one or a plurality of servers, and the display means 70 and the input means 80 may be provided on the terminal side. At this time, the network may be an external network such as the Internet, an internal network such as an intranet or a LAN, or a combination thereof, and whether it is wired, wireless, or a mixture of wired and wireless. Does not matter. The terminal is mainly the terminal of the person in charge of service, but the applicant for matching directly inputs the applicant data (attribute data of the applicant), or the matching result (scoring result) is displayed on the screen or the like. In the case of direct reference, the network may be an external network such as the Internet, and the terminal of the desired person (matching target person) may be installed. In addition, a terminal for a system administrator who manages system construction, maintenance, and updates will also be installed.

<Detailed configuration of matching system 10>

<Structure of input receiving means 31>

The input receiving means 31 receives the input of the desired person data (attribute data of the desired person, see FIG. 2) for the matching requester (requester who receives the matching service), and stores the desired person data storage means 41. It executes the process. The input of the applicant data is mainly performed by the service person operating the input means 80 based on the documents described by the applicant himself or the information heard from the applicant. When the matching system 10 has a network configuration, it may be input from the terminal of the person in charge of service, or may be input from the terminal of the desired person operated by the desired person himself / herself. Further, the input receiving means 31 may execute a process of taking in the desired person data input by another computer and recorded in a recording medium such as a USB memory or a DVD and storing the desired person data in the desired person data storage means 41. ..

As shown in FIG. 2, the applicant data has the same data format as the matching target person data registered so far (before the request for matching by the applicant), and is newly added by the input receiving means 31. Item number (identification information of matching target person), own industry (handled products), own location, annual sales, at least one (here, seven as an example) need flag, desired industry (here, seven) It is data that corresponds to the industry of the matching partner), the desired area (location of the matching partner), self-promotion (text data in free description format), and needs details (text data in free description format).

Here, the needs flags are data indicating the types of needs (data of 1,0 whether or not they are applicable), and in FIG. 2, seven are provided as an example, but the number of needs flags installed is arbitrary. It may be one or a plurality. In addition, the content and the degree of roughness thereof can be set arbitrarily. For example, a need flag of "sell" and a need flag of "buy" may be provided, and more finely, for example, "sell parts". Needs flags such as "Masu", "Sell materials", "Buy parts", "Buy materials" may be set. For example, "sell" and "seek a supplier" have the same content, and "buy" and "seek a supplier" have the same content, so the difference in expression is a problem. do not become.

Similarly, a need flag of "manufactured by us" and a need flag of "manufacture by your company" may be set, and more finely, for example, "manufacture parts" and "materials". Needs flags such as "manufacture", "manufacture parts", and "manufacture materials" may be provided. In addition, for example, "Build a system", "Please build a system", "Transport products / products", "Transport products / products", "Undertake advertising" , "Please advertise", "Dispatch human resources", "Please dispatch human resources", "Retail", "I am looking for a retail store", "I will undertake nationwide expansion" , "I want to expand nationwide", "I will provide real estate", "I want to provide real estate", "I will provide services", "I want to provide services", "Delivery will be on the same day or Needs flags such as "It's the next day" and "I want the delivery date to be the same day or the next day" can be set. In addition, a needs flag of "other" may be provided.

<Structure of Combined Text Data Creating Means 32>

The combined text data creating means 32 is a process of creating [α] combined text data, [β] removing unnecessary symbols and tags, and [γ] decomposing into words by morphological analysis and extracting only nouns. And [δ] the process of removing unnecessary words is executed. The data obtained by executing the process of [α] and the data obtained by executing the processes of [α] to [δ] are both referred to as combined text data. At this time, after the processing of [γ], it is decomposed into words, but when estimating the topic, it has the same information as the data obtained by executing the processing of [α]. Therefore, for convenience of explanation, it is called combined text data. Further, when the combined text data that has undergone the processing up to [δ] is particularly distinguished and called, it is referred to as the combined text data after processing.

In the processing of [α] to [δ] by the combined text data creating means 32, in the processing during operation when the topic model including the applicant data is not relearned (see FIG. 8), the applicant data is stored. It is executed using the applicant data (attribute data of the applicant, see FIG. 2) for the matching applicant stored in the means 41. The obtained combined text data after processing may be stored in the combined text data storage means or the desired person data storage means 41 (not shown) in association with the identification information (case number) of the applicant. In this case, since the topic model is not relearned, the combined text data for the matching target person other than the applicant (the matching target person registered before the matching request by the applicant) is not required, so the matching target person The processing of [α] to [δ] above using each matching target person data stored in the data storage means 42 is not performed.

On the other hand, in the operation-time processing when the topic model is relearned including the applicant data (see FIG. 9), the above processes [α] to [δ] are stored in the applicant data storage means 41. Not only is it executed using the desired person data (attribute data of the desired person, see FIG. 2), but also each matching target person data stored in the matching target person data storage means 42 (before the request for matching by the desired person). The above-mentioned processes [α] to [δ] using the attribute data of the matching target person registered from (see FIG. 2) are also executed. In this case, since the topic model is relearned, both combined text data are required. In addition, the above-mentioned processes [α] to [δ] are performed using each matching target person data stored in the matching target person data storage means 42 in the prior learning process (learning process before the request for matching by the desired person). When the combined text data after processing obtained by execution is saved, the saved combined text data after processing may be used. Therefore, in the processing at the time of operation, the above [α] to [ It is not necessary to execute the process of [δ] again.

Further, also in the prior learning process (learning process before the request for matching by the applicant), the above-mentioned [α] to [δ] processes by the combined text data creating means 32 are received by the learning data preparing means 39. Is executed. In this case, since it is the time before the request for matching by the applicant, the applicant data (attribute data of the applicant, see FIG. 2) stored in the applicant data storage means 41 does not exist. Therefore, only the above-mentioned processes [α] to [δ] using each matching target person data stored in the matching target person data storage means 42 are executed. As described above, the combined text data after the processing obtained by executing the above-mentioned processes [α] to [δ] in this prior learning process may be saved. For example, the obtained combined text data after processing may be associated with the identification information (case number) of the matching target person and stored in the combined text data storage means or the matching target person data storage means 42 (not shown). .. In the present application, the learning includes re-learning performed after the request for matching by the applicant (re-learning including the applicant data) and normal learning performed before the request for matching by the applicant (there is the applicant data). There is learning at the stage where it is not done), and especially when distinguishing between them, the former learning is called operational learning (re-learning), and the latter learning is called prior learning.

[Α] In the combined text data creation process, as shown in FIG. 2, the combined text data creating means 32 comprises a self-industry (handling product) and a self-PR (free description) that constitute the applicant data and each matching target person data. By combining self-data consisting of text data (text data in format) and needs data consisting of text data including desired industry (industry of matching partner) and needs details (text data in free description format). , Create combined text data. The obtained combined text data is treated as one document data.

At this time, as shown in FIG. 3, the combined text data creating means 32 combines the self-industry, the self-PR (free description format), the desired industry, and the needs details (free description format) with the punctuation mark (.) In between. Here, the self-data includes the self-industry and the needs data includes the desired industry. However, only the self-PR (free description format) may be used as the self-data without including the self-industry, and the desired industry is included. Instead, only the needs details (free description format) may be used as the needs data.

Further, since the combined text data created by the combined text data creating means 32 is text data, if the own industry or the desired industry is a selectable symbol or number (1, 2, ..., etc.), Convert those industries into text data (electrical, construction / civil engineering, IT, etc.) and then combine them. On the other hand, if your own industry or desired industry is in a free description format, or if the industry selected from the industries prepared in advance by the matching service provider is described in text, the text data is already available. Since it is, it can be combined as it is.

Furthermore, the industry selected from the industries prepared in advance by the matching service provider (regardless of whether it is selected and specified by a symbol or number or described in text) and the matching target person are in a free description format. If the described industry is used together (when there is no applicable industry among the prepared industries, the corresponding industry is described in the free description format), free description Only the type of industry described in the format may be included in the self-data and needs data. In this case, it is considered that only the self-industry and the desired industry described in the free description format are the same information as the self PR (free description format) and the needs details (free description format).

Although the illustrations in FIGS. 2 and 3 are omitted, if there is a free description format "other" column, the text data described in the "other" column is self-promoted (free description). The information may be considered to be equivalent to the format) and needs details (free description format), and may be combined and included in the combined text data.

[Β] In the process of removing unnecessary symbols, tags, etc., the combined text data creating means 32 deletes, for example, ☆, <br>, etc. from the combined text data obtained in the process of [α].

In the process of decomposing into words by [γ] morphological analysis and extracting only nouns, the combined text data creating means 32 first uses the combined text data obtained in the process of [β] as shown in FIG. Morphological analysis is performed, and the combined text data is decomposed (divided) into words and divided into words. This morphological analysis can be performed using existing analysis tools.

At this time, a dictionary of words and synonyms may be created for morphological analysis and stored in a word dictionary storage means or a synonym dictionary storage means (not shown). In the word dictionary, words that are not included in the dictionary of existing analysis tools such as "earphone", "Instagram", and "organic" are registered. Further, in the synonym dictionary, for example, swaying words such as "battery" and "battery" are registered.

[Δ] In the unnecessary word removal process, the combined text data creating means 32 narrows down the words. First, the words are filtered based on the part of speech and the number of occurrences of the words. Specifically, only a part of the noun (for example, "general", "sahen connection", etc.) is left, and other words are discarded. In addition, in a set of all combined text data (combined text data of all matching targets), words whose number of occurrences is less than 3, for example, are discarded. The relationship between each word and the number of occurrences is stored in a word appearance number storage means (not shown). Therefore, when the combined text data of the applicant as a new matching target person increases, the number of occurrences of the words contained therein is added, so that the threshold value of, for example, 3 times may be exceeded, thereby discarding. In some cases, the words that have been used will be put to good use. In addition, when the combined text data of the applicant as a new matching target person increases, a completely new word appears in the combined text data, and if the number of appearances exceeds the threshold value of, for example, 3 times, that is the case. New words are not discarded words, but utilized words.

Next, the combined text data creating means 32 eliminates unnecessary words (noise words) stored in the unnecessary word dictionary storage means 43. Specifically, for example, words that are considered to appear regardless of the type of business, such as "company", "needs", and "support", are excluded as unnecessary words. In addition, for example, domestic place names such as "Kyoto", "Kanto", and "Eastern Japan" are excluded as unnecessary words. However, overseas place names such as "Europe", "China", and "Dalian" are left. The word to be registered as an unnecessary word is, for example, 1,000 to 2,000 words.

In the example of FIG. 3, the words remaining after the above processing of [α] to [δ] by the combined text data creating means 32 are, in order from the beginning, “food and drink”, “store”, “liquor”, and “brand”. , "Food", "Food", "Alcoholic beverages", "Food", "Brand", "Awamori", and so on. Therefore, as shown in FIG. 3, a relationship between each remaining word and the number of times they appear is obtained, and this relationship becomes information necessary for performing topic estimation. That is, it is the number of occurrences of each word in the combined text data (i = 00001234) treated as one document data. i = 00001234 is a matter number, identification information of a matching target person, and a number of combined text data (document data). The matter numbers do not have to be consecutive numbers (although they may be discrete numbers), but for the sake of explanation, the numbers will be packed together here, including the explanation of the next topic estimation.

<Structure of topic estimation means 33>

When the topic estimation means 33 retrains the topic model including the applicant data (see FIG. 9), the combined text data of a large number of matching target persons including the applicant created by the combined text data creating means 32. A topic value (vertical vector π (vertical vector π (vertical vector π)) indicating the appearance probability of each topic in the combined text data (i) of each matching target person including the applicant by performing Gibbs sampling or the like by soft clustering or a neural language model. i)) and the topic estimation process for obtaining the appearance probability (matrix β) of each word in each topic are executed, and the topic model storage means using π (i) and β obtained by this topic estimation process as a topic model. The process of storing in 44 is executed.

However, the appearance probability (matrix β) of each word in each topic is not used in the operation processing when the topic model is relearned including the applicant data (see FIG. 9).

Further, also in the preliminary learning process shown in FIG. 7 (learning process before the request for matching by the applicant), the topic estimation process by the topic estimation means 33 is executed in response to the instruction by the learning data preparation means 39, and each of them is executed. The topic value (vertical vector π (i)) indicating the appearance probability of each topic in the combined text data (i) and the appearance probability (matrix β) of each word in each topic are obtained and obtained by this topic estimation process. Π (i) and β are stored in the topic model storage means 44 as a topic model. In this pre-learning process, the applicant data does not exist, and therefore the combined text data of the applicant is not created. Therefore, the combined text data of a large number of registered matching target persons created by the combined text data creating means 32. The topic estimation process is executed using.

Regarding the soft clustering or neural language model when executing the topic estimation process, in this embodiment, as an example, Latent Dirichlet Allocation (LDA: Latent Dirichlet Allocation) is adopted. The "model" of the neural language model referred to here is a concept including an algorithm and parameters, while the "model" of the topic model stored in the topic model storage means 44 is a parameter (a parameter obtained as a learning result). It is a concept that refers to learning result data). Therefore, as shown in FIG. 1, the estimator 50 is composed of the topic estimation means 33 and the desired topic value prediction means 34 realized by the algorithm, and the topic model storage means 44 for storing the topic model.

In addition to Latent Dirichlet Allocation (LDA), for example, Fuzzy c-means, mixture distribution model, non-negative Matrix Factorization (NMF), pLSI (probabilistic). Latent Semantic Indexing), Doc2Vec, SCDV (Sparse Compose Document Vecotors) and the like can be adopted. For example, when implementing Doc2Vec, an existing library called Gensim can be used.

Here, the topic value indicating the appearance probability of each topic in each combined text data (i) is a K-dimensional vertical vector π (i) as shown in FIGS. 7 and 9, and this is used as a topic in the present application. We call it the distribution. i is a document number, and when i = 1 to n, n is the number of documents. In the present invention, the number of documents n corresponds to the number of combined text data (that is, the number of matching targets). In the business matching of the present embodiment, it corresponds to the number of registered companies, for example, n = about 100,000 companies. K is the number of topics and is specified by the system builder. Therefore, the topic distribution is composed of each topic value π (i, 1), π (i, 2), π (i, 3), ..., Π (i, K) of the topic number = 1 to K. For example, in the business matching of the present embodiment, 20 industries can be assumed as topics, and K = 20 or the like can be set. However, the number of topics K is not limited to 20, and it is not always necessary to assume an industry as a topic. Since this embodiment is business matching, it is easiest to apply and think about it. The system is only designed assuming the industries that can be used.

Further, the appearance probability of each word in each topic can be represented by a matrix β of K rows and p columns, as shown in FIGS. 7 and 9. K is the number of topics. p is the number of words, for example, p = about 3,000 words.

As shown in FIG. 3, the information directly obtained from the combined text data (i = 00001234) treated as one document data is data showing the relationship between each word included in the combined text data and the number of times they appear. Is. Then, if the number of occurrences of each word is known, the probability of occurrence of each word in a certain combined text data can be found by dividing the number of occurrences of each word by the sum of the number of occurrences of all words. It becomes known information. The topic estimation means 33 uses data indicating the relationship between each word and the number of occurrences of each word in each of a large number of combined text data as known information, and Gibbs sampling or the like is performed using the known information. The topic distribution π (i) in each combined text data (i) (topic value indicating the appearance probability of each topic of topic number = 1 to K) and the appearance probability of each word in each topic (matrix of K rows and p columns). β) is calculated. When the matrix β of K rows and p columns and the topic distribution π (i) represented by the K-dimensional vertical vector are obtained, each word (1st to pth words) in the combined text data (i) is obtained. ) Appearance probability is calculated by multiplying the transposed matrix of the matrix β (p rows and K columns) and the topic distribution π (i) (K-dimensional vertical vector) in this order.

<Structure of Applicant Topic Value Predicting Means 34>

When the desired person topic value predicting means 34 does not relearn the topic model including the desired person data (see FIG. 8), the combined text data after processing of the desired person created by the combined text data creating means 32, And, using the appearance probability of each word in each topic stored in the topic model storage means 44 (matrix β of K rows and p columns), a topic value (topic value) indicating the appearance probability of each topic in the combined text data of the applicant. The process of predicting the topic distribution represented by the K-dimensional vertical vector π (i)) is executed.

Here, as shown in FIG. 3, the relationship between each word in the combined text data of the applicant and the number of occurrences thereof (that is, the appearance probability of each word in the combined text data of the applicant treated as one document data). And the appearance probability (matrix β of K rows and p columns) of each word (1st to pth words) in each topic (topic number = 1 to K) is known information. Therefore, the applicant topic value prediction means 34 uses the known information as a topic value indicating the appearance probability of each topic in the combined text data of the applicant (topic distribution represented by the K-dimensional vertical vector π (i)). ) Predict.

In addition, in the combined text data of the applicant, a new word that is not included in each word (1st to pth words) prepared as the appearance probability of each word in each topic (matrix β of K rows and p columns) If the word is present, the prediction including the new word cannot be performed. Therefore, assuming that the new word does not exist, the topic distribution in the combined text data of the applicant is predicted. Therefore, the information of this newly-developed word is reflected in the topic model stored in the topic model storage means 44 in the topic estimation process performed by the topic estimation means 33 after that.

<Structure of word importance vector calculation means 35>

The word importance vector calculation means 35 calculates a word importance vector consisting of a TFIDF value of each word or other word importance index values for the combined text data of each matching target including the applicant, and obtains the word. The process of associating the importance vector with the identification information of the matching target person (including the applicant) and storing it in the word importance vector storage means 45 is executed. The combined text data used for calculating the word importance vector is a set of words remaining after the processing up to [δ] by the combined text data creating means 32.

Here, the word importance index value adopts the TFIDF value as an example in the present embodiment, but is not limited to this, and may be, for example, Okapi-BM25 or the like. The TF (Term Frequency) is the frequency of appearance of words in a document (here, combined text data), and the number of occurrences of each word in a certain combined text data is the number of appearances of all words in the combined text data. It is the value divided by the sum of. IDF (Inverse Document Frequency) is the inverse document frequency, and the logarithm is taken for the value obtained by dividing the total number of documents (here, the total number of combined text data) by the number of combined text data including each word. The value.

The word importance vector is a p-dimensional vector when the number of words is p. For example, when p = about 3,000 words, it is a vector of about 3,000 dimensions. It is preferable that this word importance vector is subjected to normalization (for example, L2 norm normalization) processing.

<Structure of feature data creating means 36 for matching>

As shown in FIG. 2, the matching feature data creating means 36 describes the topic distribution π (i) (topic value indicating the appearance probability of each topic of topic number = 1 to K) in the combined text data and the combined text data. Matching processing means 37 and matching model storage means 46 using the word importance vector (in the present embodiment, a TFIDF vector is used as an example) and the needs flags and annual sales for the applicant and each candidate. It executes a process of creating matching feature data which is input data to the classifier 60. However, what is described in the lower part of FIG. 2 is the matching feature data on one side of the pair consisting of the matching applicant and the candidate, and is for matching in the state before the composite variable is created. It is feature data.

More specifically, the matching feature data creating means 36 is obtained by the desired person topic value predicting means 34 in the operation operation when the topic model including the desired person data is not retrained (see FIG. 8). Topic distribution in the combined text data of the applicant (topic value indicating the appearance probability of each topic) and topic distribution in the combined text data of each candidate stored in the topic model storage means 44 (appearance probability of each topic) The word importance vector (TFIDF vector, etc.) for the combined text data of the applicant and each candidate calculated by the word importance vector calculation means 35 and stored in the word importance vector storage means 45. The needs flags and annual sales included in the applicant data stored in the applicant data storage means 41, and the needs included in the matching target person data of each candidate stored in the matching target person data storage means 42. Using flags and annual sales, as shown in FIG. 5, matching feature data for a plurality of pairs consisting of applicants and each candidate is created.

Further, the matching feature data creating means 36 is obtained by the topic estimation process by the topic estimation means 33 in the operation operation when the topic model is relearned including the applicant data (see FIG. 9). The topic distribution (topic value indicating the appearance probability of each topic) in the combined text data of the applicant and each candidate stored in the model storage means 44, and the word importance vector calculated by the word importance vector calculation means 35. A word importance vector (TFIDF vector, etc.) for the combined text data of the applicant and each candidate stored in the storage means 45, and a needs flag included in the applicant data stored in the applicant data storage means 41. As shown in FIG. 5, the applicant and each candidate are used by using the annual sales and the needs flag and the annual sales included in the matching target data of each candidate stored in the matching target data storage means 42. Create matching feature data for multiple pairs of people.

Further, the matching feature data creating means 36 also receives an instruction from the learning data preparation means 39 in the prior learning process (learning process before the request for matching by the applicant) shown in FIG. 7, and the matching feature data To create. At this time, in the pre-learning process shown in FIG. 7, since there is no applicant because it is a stage before the request for matching by the applicant, it is for matching for a plurality of pairs consisting of the applicant and each candidate. As shown in FIG. 6, instead of creating feature data, matching features for a large number of matching target pair (a pair corresponding to a candidate and a candidate because it is learning) as learning data. Data will be created, the details of which will be described later with reference to FIG. Specifically, in the prior learning process shown in FIG. 7, the matching feature data creating means 36 is obtained by the topic estimation process by the topic estimation means 33 and stored in the topic model storage means 44. The topic distribution (topic value indicating the appearance probability of each topic) in the combined text data of the above, and the combined text of each matching target person calculated by the word importance vector calculation means 35 and stored in the word importance vector storage means 45. A large number of words are used for the word importance vector (TFIDF vector, etc.) for the data, and the needs flags and annual sales included in the matching target person data of each matching target person stored in the matching target person data storage means 42. Create matching feature data for a pair of matching targets.

Further, in FIG. 2, the matching feature data (however, FIG. 2 shows the state before creating the composite variable) for each matter number (identification information of the matching target person) is the topic distribution and the word importance vector (for example,). It is composed of a TFIDF vector, etc.), a needs flag, and an annual sales. As shown by the two-point chain line in FIG. 2, the word importance vector, the needs flag, and the annual sales are used to create matching feature data. It is not essential for you, and you can choose not to use it. Further, although the illustration in FIG. 2 is omitted, the document length may be used for creating the matching feature data. The document length is, for example, the combined text data in a state in which [α] combined text data is created by the combined text data creating means 32, or [β] is processed in which unnecessary symbols, tags, etc. are removed. (That is, the number of characters in the combined text data before being decomposed into words by the morphological analysis of [γ]), or the total number of words in the combined text data after processing up to [δ]. Etc. can be used.

Further, as shown in FIG. 4, in the creation of matching feature data, it is effective to include synthetic variables in order to perform appropriate scoring.

As shown in FIG. 4A, when a composite variable is created based on each topic value, a composite consisting of the product of the topic distribution π (X) of the applicant X and the topic distribution π (Y) of the candidate Y. The variable distribution λ (XY) can be obtained and included in the matching feature data. In FIG. 4A, the notation λ (XY) = π (X) * π (Y) is used, but the inner product of the vectors is not calculated, but the product of the corresponding elements of the vector is calculated. It means that. That is, since it is the product of topic values for the same topic, the product of topic values with topic number = 1, the product of topic values with topic number = 2, ..., Topic number = K (K is the number of topics). Find the product of topic values. To be more accurate, λ (XY, 1) = π (X, 1) * π (Y, 1), λ (XY, 2) = π (X, 2) * π (Y, 2), ... , Λ (XY, K) = π (X, K) * π (Y, K). Therefore, the distribution λ (XY) of the composite variable, which is a K-dimensional vector, is obtained from the topic distribution π (X), which is a K-dimensional vector, and the topic distribution π (Y), which is a K-dimensional vector.

Further, as shown in FIG. 4 (A), the distribution μ (XY) of the composite variable consisting of the absolute value of the difference between the topic distribution π (X) of the applicant X and the topic distribution π (Y) of the candidate Y. Can be obtained and included in the matching feature data. The absolute value of the difference, not the difference, is used to eliminate the directionality, and to make μ (XY) and μ (YX) the same (so that they are the same even if the order is changed). Is. In this case, find the absolute value of the difference between the corresponding elements of the vector. That is, since it is the absolute value of the difference between topic values for the same topic, the absolute value of the difference between topic values with topic number = 1, the absolute value of the difference between topic values with topic number = 2, ..., Topic number = Find the absolute value of the difference between the topic values of K (K is the number of topics). To be more accurate, μ (XY, 1) = | π (X, 1) -π (Y, 1) |, μ (XY, 2) = | π (X, 2) -π (Y, 2) ) |, ..., Μ (XY, K) = | π (X, K) -π (Y, K) |. Therefore, the distribution μ (XY) of the composite variable, which is a K-dimensional vector, is obtained from the topic distribution π (X), which is a K-dimensional vector, and the topic distribution π (Y), which is a K-dimensional vector.

In the example of FIG. 4A, both the distribution λ (XY) of the composite variable consisting of the product of the topic distribution and the distribution μ (XY) of the synthetic variable consisting of the absolute value of the difference in the topic distribution are matched. Although it is preferable to include it in the feature data, either one of the product distribution λ (XY) and the absolute value distribution μ (XY) of the difference may be used.

Further, as shown in FIG. 4B, when a composite variable is created with each need flag, the logical sum (OR), logical product (AND), of the corresponding needs flags (of the same type of needs flags), Exclusive OR (XOR), NOR, Negative OR (NAND), Negative OR (NXOR), Arithmetic and the like can be adopted. In the example of FIG. 4 (B), the composite variables by the logical sum (OR), the logical product (AND), and the arithmetic sum are shown. This example does not mean that all of the multiple types (3 types) of synthetic variables based on logical sum (OR), logical product (AND), arithmetic sum, etc. are used, but one of the synthetic variables is selected and used. It is described to the effect that it should be done. Note that a plurality of types of composite variables may be included in the matching feature data.

Then, when creating a composite variable with each need flag, the composite variable may be created across different types of needs flags instead of the corresponding needs flags. For example, when there are needs flags related to the content such as (1) sell, (2) buy, among the needs flags, the needs flags of (1) and (2) are used to describe the above. You may create a composite variable by the logical sum (OR) or the like.

Further, as shown in FIG. 4C, when the composite variable is created by the annual sales, the value of the ratio of the annual sales of both can be adopted. When the composite variable M (XY) is based on the ratio value, if the annual sales of the applicant X is S (X) and the annual sales of the applicant Y is S (Y), then M (XY) is given a direction. ) = S (X) / S (Y), or M (XY) = S (Y) / S (X). Further, in order to eliminate the directionality, the larger value of M (XY) = S (X) / S (Y) or S (Y) / S (X), or M (XY) = S (X). ) / S (Y) or S (Y) / S (X), whichever is smaller. When the directionality is given, M (XY) and M (YX) have different values.

Further, although the illustration in FIG. 4 is omitted, when the composite variable is created by the document length, the value of the ratio of the document lengths of both can be adopted. When the composite variable is based on the ratio value, the directionality may be kept waiting or the directionality may be lost, as in the case of the annual sales described above.

<Structure of Matching Processing Means 37>

The matching processing means 37 uses each of the plurality of matching feature data created by the matching feature data creating means 36 (each matching feature data of a plurality of pairs consisting of a desired person and each candidate) as input data. By using the matching model that has been learned in advance by the two-class classification of whether or not to be selected and stored in the matching model storage means 46, the classification process is performed by the supervised classification model, and each of the applicants for matching is subjected to the classification process. It executes a process of outputting a score indicating the probability that a candidate is selected.

At this time, the score is output for each candidate (for each pair consisting of the applicant and the candidate). For example, when the score is output with the likelihood, it is output in the range of 0 to 1, and the closer the score is to 1, the more suitable the candidate is as a matching partner.

Here, the "model" of the "supervised classification model" is a concept including an algorithm and parameters, and the "model" of the matching model stored in the matching model storage means 46 is a parameter (learning result) obtained by learning. It is a concept only for data). Therefore, the discriminator 60 is composed of the matching processing means 37 realized by the algorithm and the matching model storage means 46 for storing the matching model.

Further, the "supervised classification model" may be any machine learning method as long as it can perform machine learning by two-class classification as to whether or not it is selected. For example, SGD Classifier (SGD Classifier: Stochastic Gradient Descent Classifier), K-Neighbors Classifier (K-Neighbors Classifier), Naive Bayes Classifier (Naive Bayes Classifier), Support Vector Classifier (SVC) ), Random Forest Classifier, Multi Layer Perceptron Classifier (MLP Classifier), neural network, etc. can be adopted.

The classifier 60 may input the same input data (matching feature data) into a plurality of types of classifiers as described above, and integrate the scores output from each (for example, each classifier). The average value of the scores output from the above may be used as the integrated score), or may be weighted at the time of integration (the weighted average value of the scores output from each classifier may be used as the integrated score). ).

Further, the matching processing means 37 displays the output score on the screen on the display means 70, and when the display means 70 is provided on the terminal connected via the network, the score data is displayed on the network. It also executes the process of sending to the terminal via the above, or attaching it to an e-mail and sending it. Further, score printing, voice output, and the like may be performed as appropriate.

Further, the matching processing means 37 may be configured to perform secondary processing such as arranging the output scores in descending order of numerical values. This secondary processing also includes an optimization calculation using a score for performing matching for a specific purpose in a form of deformation (see FIG. 10) described later.

<Structure of learning means 38>

The learning means 38 is provided with learning data (selection / non-selection labels (tags) for each of a large number of matching feature data) created by the learning data preparation means 39 and stored in the learning data storage means 47. Using the attached annotated data (see FIG. 6), learning processing is performed by a supervised classification model, and the obtained matching model (here, it means a parameter as learning result data) is stored in the matching model. The process of storing in the means 46 is executed.

The learning process by the learning means 38 is a learning process (process performed before the requester requests mating) shown in FIG. 7, and the matching model obtained by this learning process includes a new model after updating. Model is also included.

If the number of combined text data is increased and the topic estimation process by the topic estimation means 33 is executed again, strictly speaking, the topic distributions in each combined text data will have slightly different values. Each matching feature data created using the data is also in a slightly different state. Then, if the learning process by the learning means 38 is performed using each of the slightly different matching feature data, the matching model stored in the matching model storage means 46 will be in a slightly different state. However, if the number of combined text data to be added is one or a relatively small number, the topic distribution hardly changes. Therefore, the matching model stored in the matching model storage means 46 is used as it is without re-learning. be able to.

Therefore, in the operation-time processing when the topic model is re-learned including the applicant data as shown in FIG. 9, the matching model stored in the matching model storage means 46 is not re-learned as it is. Can be used.

On the other hand, in the case of FIG. 9, in addition to re-learning the topic model, the matching model stored in the matching model storage means 46 is also re-learned, and the updated matching model is used to relearn the matching processing means 37. The score may be output by. In the re-learning of the matching model in this case, it is not possible to tag the applicant as a new matching target with the registered matching target, so that the applicant data cannot be selected or not. It is sufficient to use the matching target person data of the registered matching target person without using. That is, the matching model may be relearned by updating only the topic distribution part of the matching feature data for the pair of registered matching target persons that have been tagged as selected / unselected.

<Structure of learning data preparation means 39>

The training data preparation means 39 creates, as training data, post-annotation data (see FIG. 6) with selected / non-selected labels (tags) for each of a large number of matching feature data, and tags the data. A process of associating a large number of the matched feature data with the pair of identification information of the matching target person and storing them in the learning data storage means 47 is executed.

At this time, the learning data preparation means 39 gives instructions to the combined text data creation means 32, the topic estimation means 33, the word importance vector calculation means 35, and the matching feature data creation means 36 to execute their respective processes. By doing so, a large number of matching feature data are prepared.

Further, the selection / non-selection tagging process for each matching feature data is executed using the matching candidate selection result list stored in the matching candidate selection result list storage means 48.

<Structure of each storage means 41 to 48>

As shown in FIG. 2, the applicant data storage means 41 has its own industry (handling products) and self as attribute data (applicant data) about the applicant (matching requester) as a new matching target. Location, annual sales, at least one (here, seven as an example) needs flag, desired industry (matching partner's industry), desired area (matching partner's location), and self-promotion (freedom) The description format text data) and the needs details (free description format text data) are stored in association with the case number (identification information of the matching target person) given to the applicant.

As shown in FIG. 2, the matching target person data storage means 42 has self-industry (handling products), self-location, and annual sales as attribute data (matching target person data) for the registered matching target person. , At least one (here, seven as an example) needs flag, desired industry (matching partner's industry), desired area (matching partner's location), and self-promotion (free description format text data) And the needs details (text data in free description format) are stored in association with the matter number (identification information of the matching target person). Therefore, the matching target person data stored in the matching target person data storage means 42 has the same data format as the desired person data stored in the desired person data storage means 41.

The unnecessary word dictionary storage means 43 stores unnecessary words (noise words) used in the [δ] unnecessary word removal process by the combined text data creation means 32. Unnecessary words include words that are considered to appear regardless of industry, such as "company," "needs," and "support," and domestic words such as "Kyoto," "Kanto," and "Eastern Japan." There is a place name.

The topic model storage means 44 is a topic distribution (topic number = 1 to K) in the combined text data of each matching target person (including a desired person in the case of FIG. 9) obtained by the topic estimation process by the topic estimation means 33. Topic value indicating the appearance probability of each topic of: π (i, 1), π (i, 2), ..., π (i, K)), which are elements of the vertical vector π (i), are matched. Each word (1 to 1) in each topic (topic number = 1 to K) obtained by this topic estimation process is stored in association with the identification information (i) (including the applicant in the case of FIG. 9). It stores the appearance probability (each element of the matrix β in the K row and p column) of the p-th word).

The word importance vector storage means 45 is a word importance vector (TFIDF value of each word or other word importance) for the combined text data of each matching target person (including the applicant) calculated by the word importance vector calculation means 35. The degree index value) is stored in association with the identification information of the matching target person (including the applicant).

The matching model storage means 46 stores a matching model (parameter as learning result data) that constitutes the classifier 60 based on the supervised classification model. This matching model is obtained by the learning process by the learning means 38.

As shown in FIG. 6, the learning data storage means 47 has a large number of matchings tagged with selection / non-selection as learning data (learning data of a matching model) created by the learning data preparation means 39. Characteristic data is stored in association with a pair of identification information of a matching target person.

The matching candidate selection result list storage means 48 identifies the matching target person (case number) actually selected as the matching candidate by the service person based on the request of the matching applicant in the past, and the identification of the applicant. It stores a matching candidate selection record list consisting of a combination with information (case number).

<Flow of learning process performed in advance: Fig. 7>

In such an embodiment, the matching system 10 performs a preliminary learning process (learning process before a request from a matching applicant) as follows.

In FIG. 7, first, the learning data preparation means 39 issues an instruction to the combined text data creation means 32, and the attribute data (attribute data) for all the registered matching target persons stored in the matching target person data storage means 42 ( As shown in FIGS. 2 and 3 from the matching target person data), the self-industry and self-PR (text data in free description format) are acquired as self-data, and the desired industry and needs details (free) are obtained as needs data. (Text data in description format) is acquired, the process of [α] is executed by the combined text data creating means 32, the acquired self-data and needs data are combined, and the combined text for all the registered matching target persons is obtained. Create data (step S1).

Next, as shown in FIGS. 2 and 3, the combined text data creating means 32 removes [β] unnecessary symbols, tags, etc. from the combined text data obtained by the processing of [α]. [Γ] The process of decomposing into words by morphological analysis and extracting only nouns and [δ] the process of removing unnecessary words are executed, and the combined text data after processing for all registered matching targets is executed. Create (step S2). In the state of undergoing the processing up to [δ], the combined text data after the processing is data showing the relationship between each word and the number of times they appear, as shown in the lower part of FIG. 3, thereby in the combined text data. It means that the appearance probability of each word has been obtained. Since the processes of [α] to [δ] have already been described in detail in the description of the combined text data creating means 32, detailed description thereof will be omitted here.

Subsequently, the learning data preparation means 39 issues an instruction to the topic estimation means 33, and soft clustering or using the processed combined text data (see the lower part of FIG. 3) for all the registered matching subjects. A topic value indicating the appearance probability of each topic (topic number = 1 to K) in each combined text data (i) by executing topic estimation processing by a neural language model (in this embodiment, LDA is used as an example). (K-dimensional vertical vector π (i)) and the appearance probability (matrix β of K rows and p columns) of each word (1st to pth words) in each topic (topic number = 1 to K) are obtained. The obtained π (i) and β are stored in the topic model storage means 44 as a topic model (step S3).

The above is the flow of processing for constructing a new model and updating the model for the topic model (π (i), β) stored in the topic model storage means 44. Then, the flow of the process of constructing a new model and updating the model for the matching model stored in the matching model storage means 46 will be described below.

In FIG. 7, first, the learning data preparation means 39 gives an instruction to the word importance vector calculation means 35, and the combined text data of all the registered matching targets (up to [δ] by the combined text data creating means 32). The word importance vector (TFIDF value of each word or other word importance index value) for the word importance vector (set of words remaining after the process of) is calculated, and the calculated word importance vector is used as the identification information of the matching target person. It is stored in the word importance vector storage means 45 in association with (case number) (step S4). Since FIG. 7 is a prior learning process, there is no applicant for matching. Therefore, the word importance vector for the combined text data of the applicant is not calculated here.

Next, the learning data preparation means 39 issues an instruction to the matching feature data creating means 36, and as shown in FIG. 6, a large number of matching targets are used as the learning data of the matching model to be stored in the matching model storage means 46. Matching feature data for a pair of persons is created, and the matching feature data of a large number of created pairs is stored in the learning data storage means 47 in association with the pair of identification information (case number) of the matching target person. Since there is no applicant in the prior learning process shown in FIG. 7, matching feature data for a plurality of pairs consisting of the applicant and each candidate is created as in the operation process shown in FIG. I don't do it.

Specifically, a large number of matching candidate selection result lists for creating matching feature data as learning data by using the matching candidate selection result list stored in the matching candidate selection result list storage means 48 by the learning data preparation means 39. Prepare a pair of identification information (case number) of the matching target person (step S5). The matching candidate selection record list includes the identification information (case number) of the matching target person who was actually selected as the matching candidate by the service person based on the request of the matching applicant in the past, and the identification information of the applicant (case number). Since the pair with the case number) is recorded, the pair of the identification information (case number) of the matching target person is used as a correct example (correct answer data), and "1 (selection)" is used as the selection / non-selection. Tagging. On the other hand, a pair other than the pair recorded in the matching candidate selection record list is regarded as a negative example (incorrect answer data), and "0 (non-selected)" is tagged as a distinction between selection and non-selection. As a result, for example, 10,000 positive examples (correct answer data) and 20,000 negative examples (incorrect answer data) are prepared.

For negative examples (incorrect answer data), a pair that is a combination of randomly sampled identification information (case number) of the matching target person and is not a pair recorded in the matching candidate selection record list. Can be. In addition, instead of sampling randomly, all combinations (round-robin combinations) of the identification information (case number) of the registered matching target person are excluded from the pairs recorded in the matching candidate selection record list. It may be a negative example (incorrect answer data).

Then, the pair of identification information (case number) of all the matching target persons prepared in step S5 above for creating the matching feature data as the learning data by the matching feature data creating means 36 (all correct answers). For each pair and all incorrect answers), the topic distribution π (i) in the combined text data of each matching target person obtained by the topic estimation process by the topic estimation means 33 and stored in the topic model storage means 44 ( Topic number = topic value indicating the appearance probability of each topic from 1 to K), and the needs flag and annual sales included in the matching target person data of each matching target person stored in the matching target person data storage means 42. Obtain and set the composite variables as shown in FIG. 4 (here, as an example, the product of the topic distribution, the absolute value of the difference of the topic distribution, the logical sum of the needs flags, the value of the ratio of the annual sales, etc.). Calculate (step S6).

Subsequently, a pair of identification information (case number) of all matching target persons prepared in step S5 described above for creating matching feature data as learning data by the matching feature data creating means 36 (all). For correct answer pairs and all incorrect answer pairs), the word importance vector for the combined text data of each matching target person calculated by the word importance vector calculation means 35 and stored in the word importance vector storage means 45. (TFIDF vector, etc.) is acquired, and the inner product of the word importance vector is calculated (step S7).

Then, as shown in FIG. 6, the pair (all correct answer pairs and all incorrect answer pairs) of the identification information (case number) of all the matching target persons prepared in step S5 described above is added to step S6 described above. The composite variables calculated in step 2 (product of topic distribution, absolute value of difference in topic distribution, logical sum of needs flags, value of annual sales ratio, etc.), and word importance vector (TFIDF vector) calculated in step S7 above. Etc.) are associated with each other to obtain matching feature data as learning data, which is stored in the learning data storage means 47. As shown in FIG. 6, it is not a composite variable (product of topic distribution, absolute value of difference in topic distribution, logical sum of needs flags, value of ratio of annual sales, etc.), but topic distribution of both pairs and both pairs. The annual sales of both the needs flag and the pair may be included in the matching feature data, or although the illustration is omitted, the word importance vector is not the inner product of the word importance vectors, but both. The word importance vector may be included in the matching feature data.

After that, the learning data 38 is created by the learning data preparation means 39 and stored in the learning data storage means 47 (each of a large number of matching feature data is tagged with or without selection). The learning process is performed by the supervised classification model using the obtained data (see FIG. 6), and the obtained matching model (parameter as learning result data) is stored in the matching model storage means 46 (step S8).

<Process flow during operation when the topic model including applicant data is not relearned: Fig. 8>

In FIG. 8, first, when the service person who received the request from the matching applicant or the applicant himself / herself operates the input means 80 and inputs the attribute data (applicant data) of the applicant, the input receiving means 31 This input is accepted, and the input applicant data is stored in the applicant data storage means 41 (step S21). In FIG. 1, the processing and data flow in the case of FIG. 8 are mainly shown by solid lines, while the processing and data flow in the case of FIG. 9 described later are mainly shown by dotted lines. Further, the processing and data flow in the case of the pre-learning in FIG. 7 described above are mainly shown by the alternate long and short dash line.

Next, from the applicant data (attribute data of the applicant) of the matching applicant stored in the applicant data storage means 41 by the combined text data creating means 32, as shown in FIGS. 2 and 3, as shown in FIGS. As self-data, self-industry and self-PR (free description format text data) are acquired, and as needs data, desired industry and needs details (free description format text data) are acquired, and the combined text data creation means 32 is used. The process of [α] is executed, and the acquired self-data and needs data are combined to create combined text data for the applicant (step S22).

Further, as shown in FIGS. 2 and 3, the combined text data creating means 32 removes [β] unnecessary symbols, tags, etc. from the combined text data obtained in the process of [α], and []. γ] The process of decomposing into words by morphological analysis and extracting only nouns and [δ] the process of removing unnecessary words are executed to create processed combined text data for the desired person (step S23). In the state of undergoing the processing up to [δ], the combined text data after the processing is data showing the relationship between each word and the number of times they appear, as shown in the lower part of FIG. 3, thereby in the combined text data. It means that the appearance probability of each word has been obtained. Since the processes of [α] to [δ] have already been described in detail in the description of the combined text data creating means 32, detailed description thereof will be omitted here.

Subsequently, the combined text data after processing of the applicant created by the applicant topic value predicting means 34 and the combined text data creating means 32, and the appearance of each word in each topic stored in the topic model storage means 44. Using the probability (matrix β of K rows and p columns), predict the topic value (topic distribution represented by the K-dimensional vertical vector π (i)) indicating the appearance probability of each topic in the combined text data of the applicant. (Step S24). The matrix β is a topic model obtained by the prior learning of FIG. In the prior learning of FIG. 7, a topic using n combined text data of i = 1 to n (n is the number of matching targets, which is the number of companies in this embodiment and corresponds to the number of documents). Assuming that a model is constructed, the combined text data of the applicant is the i = (n + 1) th combined text data, so it is not used for learning the topic model, but if the matrix β is used, i = ( The topic distribution π (i) (i = n + 1) in the n + 1) th combined text data can be predicted. However, if the combined text data after processing of the i = (n + 1) th applicant contains a new word (word other than the 1st to pth words) that does not exist in the matrix β, , The new word is treated as nonexistent because it cannot be predicted.

Then, the word importance vector calculation means 35 is used to obtain a word importance vector (a set of words remaining after the processing up to [δ] by the combined text data creation means 32) of the applicant (for each word). The TFIDF value or other word importance index value) is calculated, and the calculated word importance vector is stored in the word importance vector storage means 45 in association with the identification information (case number) of the applicant (step S25). The word importance vector for the combined text data of the matching target person other than the desired person is calculated by the prior learning process of FIG. 7 and stored in the word importance vector storage means 45.

Subsequently, by the matching feature data creating means 36, as shown in FIG. 5, the identification information (case number) of the applicant and the identification information (case) of all the registered matching target persons who are candidates for the matching partner. Prepare multiple pairs consisting of numbers). In the example of FIG. 5, since the case number of the applicant X is i = 00100001 and the case number of the candidate Y is i = 00000001 to 00100000, one person (1 individual, 1 corporation, or 1 group). 100,000 candidates, who are all registered registrants, are combined and 100,000 pairs are prepared for those who wish. Therefore, the matching feature data creating means 36 creates matching feature data for these plurality of pairs (100,000 pairs).

Specifically, the topic distribution π (i) (i) in the combined text data of the applicants obtained by the applicant topic value prediction means 34 for the plurality of pairs prepared as described above by the matching feature data creation means 36. = N + 1) (topic value indicating the appearance probability of each topic of topic number = 1 to K) and topic distribution π (i) (i =) in the combined text data of each candidate stored in the topic model storage means 44. Matching target with 1 to n) (topic value indicating the appearance probability of each topic of topic number = 1 to K) and the needs flag and annual sales included in the applicant data stored in the applicant data storage means 41. Using the needs flag and annual sales included in the matching target person data of each candidate stored in the person data storage means 42, a synthetic variable (here, as an example, of a topic distribution) is used as shown in FIG. The product, the absolute value of the difference in the topic distribution, the logical sum of the needs flags, the value of the ratio of the annual sales, etc.) are calculated (step S26).

Subsequently, the applicant and the applicant who are calculated by the matching feature data creating means 36 by the word importance vector calculating means 35 and stored in the word importance vector storing means 45 for the plurality of pairs prepared in step S26 above. A word importance vector (TFIDF vector or the like) for the combined text data of each candidate is acquired, and the inner product of the word importance vectors is calculated (step S27).

Then, as shown in FIG. 5, the matching feature data creating means 36 adds the composite variable (topic distribution product) calculated in step S26 to the plurality of pairs (case number pairs) prepared in step S26 described above. , The absolute value of the difference in topic distribution, the logical sum of the needs flags, the value of the annual sales ratio, etc.), and the inner product of the word importance vector (TFIDF vector, etc.) calculated in step S27 above. Feature data for matching is created for a plurality of pairs consisting of a person and each candidate (step S28). As shown in FIG. 5, it is not a composite variable (product of topic distribution, absolute value of difference in topic distribution, logical sum of needs flags, value of ratio of annual sales, etc.), but topic distribution of both pairs and both pairs. The annual sales of both the needs flag and the pair may be included in the matching feature data, or although the illustration is omitted, the word importance vector is not the inner product of the word importance vectors, but both. The word importance vector may be included in the matching feature data.

After that, each of the matching feature data for the plurality of pairs created by the matching feature data creating means 36 by the matching processing means 37 (each matching feature data of the plurality of pairs consisting of the applicant and each candidate). Is selected as input data, and matching is desired by performing classification processing by a supervised classification model using a matching model that has been learned in advance by two-class classification of whether or not to be selected and stored in the matching model storage means 46. A score indicating the probability that each candidate is selected is output to the person (step S29).

<Process flow during operation when re-learning the topic model including applicant data: Fig. 9>

In FIG. 9, first, when the service person who received the request from the matching applicant or the applicant himself / herself operates the input means 80 and inputs the attribute data (applicant data) of the applicant, the input receiving means 31 This input is accepted, and the input applicant data is stored in the applicant data storage means 41 (step S41). In FIG. 1, the processing and data flow in the case of FIG. 9 are mainly shown by dotted lines, while the processing and data flow in the case of FIG. 8 described above are mainly shown by solid lines. Further, the processing and data flow in the case of the pre-learning in FIG. 7 described above are mainly shown by the alternate long and short dash line.

Next, the combined text data creating means 32 stores the applicant data (attribute data of the matching applicant) stored in the applicant data storage means 41, and each stored in the matching target person data storage means 42. From the matching target person data (attribute data of the matching target person registered before the request for matching by the applicant), as shown in FIGS. 2 and 3, self-industry and self-PR (free description format) are used as self-data. (Text data) is acquired, desired industry and needs details (text data in free description format) are acquired as needs data, the processing of [α] by the combined text data creation means 32 is executed, and the acquired self-data and the acquired self-data and The needs data is combined to create combined text data for all matching target persons (including applicants) (step S42).

Further, as shown in FIGS. 2 and 3, the combined text data creating means 32 removes [β] unnecessary symbols, tags, etc. from the combined text data obtained in the process of [α], and []. γ] The process of decomposing into words by morphological analysis and extracting only nouns and [δ] the process of removing unnecessary words are executed, and the combined text after processing for all matching target persons (including applicants) Data is created (step S43). In the state of undergoing the processing up to [δ], the combined text data after the processing is data showing the relationship between each word and the number of times they appear, as shown in the lower part of FIG. 3, thereby all matching targets. It means that the appearance probability of each word in the combined text data for the person (including the applicant) is obtained.

It should be noted that the above [α] to [δ] are used by using each matching target person data stored in the matching target person data storage means 42 in the prior learning process (learning process before the request for matching by the desired person) shown in FIG. ] Is executed, and the combined text data after the processing is saved, the saved combined text data after the processing may be used. Therefore, in steps S42 and S43, the applicant It is not necessary to re-execute the above-mentioned processes [α] to [δ] for the matching target persons other than the above.

Subsequently, the topic estimation means 33 uses the processed combined text data (see the lower part of FIG. 3) of a large number of matching target persons including the applicant created by the combined text data creating means 32, and soft clustering or neural. By performing Gibbs sampling etc. in the language model, the topic distribution π (i) (i = 1 to (n + 1)) (topic) in the combined text data (i) of each matching target including the applicant (i = n + 1). It is a topic value indicating the appearance probability of each topic with numbers = 1 to K, and is a K-dimensional vertical vector π (i)) and each word (1st to pth) in each topic (topic number = 1 to K). A topic estimation process for obtaining the appearance probability of a word) (matrix β in rows K and p columns) is executed, and π (i) (i = 1 to (n + 1)) and β obtained in this topic estimation process are used as a topic model. Is stored in the topic model storage means 44 (step S44).

Then, the word importance vector calculation means 35 is used to obtain a word importance vector (a set of words remaining after the processing up to [δ] by the combined text data creation means 32) of the applicant (for each word). The TFIDF value or other word importance index value) is calculated, and the calculated word importance vector is stored in the word importance vector storage means 45 in association with the identification information (case number) of the applicant (step S45). The word importance vector for the combined text data of the matching target person other than the desired person is calculated by the prior learning process of FIG. 7 and stored in the word importance vector storage means 45.

Subsequently, by the matching feature data creating means 36, as shown in FIG. 5, the identification information (case number) of the applicant and the identification information (case) of all the registered matching target persons who are candidates for the matching partner. Prepare multiple pairs consisting of numbers). This is the same as the case of FIG. 8 described above.

Specifically, the topic distribution π (in the combined text data of the applicant and each candidate stored in the topic model storage means 44 for the plurality of pairs prepared as described above by the matching feature data creation means 36) i) (i = 1 to (n + 1)) (topic value indicating the appearance probability of each topic of topic number = 1 to K) and a need flag included in the applicant data stored in the applicant data storage means 41. And the annual sales, and the needs flag and the annual sales included in the matching target person data of each candidate stored in the matching target person data storage means 42, as shown in FIG. As an example, the product of the topic distributions, the absolute value of the difference in the topic distributions, the logical sum of the needs flags, the value of the ratio of the annual sales, etc.) are calculated (step S46).

The subsequent processes of steps S47, S48, and S49 are the same as those of steps S27, S28, and S29 of FIG. 8 described above.

<Effect of this embodiment>

According to the present embodiment as described above, there are the following effects. That is, in the matching system 10, combined text data is created by combining the self-data and needs data of a large number of matching subjects, and the combined text data is used to perform topic estimation processing by soft clustering or a neural language model. Execute to obtain the topic distribution (topic value indicating the appearance probability of each topic in the combined text data of each matching target person), and use the obtained topic distribution to perform machine learning with a supervised classification model to create a matching model. Since it is constructed and the matching model is used to match the applicant for matching with each candidate, manual keyword conversion work becomes unnecessary, and personal operations can be eliminated.

Therefore, it is possible to save the time and effort required for the conventional keyword conversion work, and it is possible to reduce the load on the service staff. In addition, since work that involves individual judgment based on know-how is eliminated, it is possible to eliminate variations in work content among service personnel, and it is possible to realize uniform service provision.

Furthermore, since machine learning of the matching model is performed using the topic distribution obtained by the topic estimation process, the feature amount that accurately and multifacetedly captures the registration information (self-data, needs data) of the matching target person. (Characteristic data for matching) can be created and appropriate scoring can be performed. Therefore, the lack of information can be suppressed. Therefore, in the case of conventional keyword matching, when the matching needs are converted into keywords, information may be lost and the desired search result may not be obtained. However, such a situation should be avoided. Can be done. Then, in the conventional method, there is a problem that the output result may be significantly different depending on the change of the search keyword, which tends to lead to opportunity loss, but such a problem can be solved.

In addition, since it is possible to perform scoring of the matching accuracy between the applicant and each candidate by using all the registered matching target persons as candidates for the matching applicant, it is unlikely to fall into an opportunity loss. In addition, referring to the high and low scores, information on matching targets with high operability (in this embodiment, companies that are likely to actually jointly, collaborate, or cooperate in business) can be arbitrarily provided. You can browse the number of cases.

Furthermore, since all the registered matching target persons can be selected as candidates and scoring can be performed mechanically, even if the registration of matching target persons increases, the burden on the service staff will increase accordingly. It can be avoided.

In addition, when creating feature data for matching, the topic distribution (topic value indicating the appearance probability of each topic) obtained by executing topic estimation processing by soft clustering or neural language model is used, so applicants and applicants It is necessary to prepare information corresponding to the topic (in this embodiment, an industry (business field) can be considered) as attribute data of each candidate (applicant data and matching target data of each candidate). Since there is no such thing, the burden on the service staff can be reduced in this respect as well.

Further, as shown in FIG. 4, the matching feature data creating means 36 can use the product of the topic values and the absolute value of the difference between the topic values as synthetic variables. Therefore, appropriate scoring can be performed as compared with the case where the topic distribution (each topic value) of both the applicant and the candidate (both of the paired matching target persons) is used as it is as the matching feature data. ..

That is, if the topic distributions of both pairs are used as they are as the matching feature data, the learning data may be biased at the time of learning, and biased learning may be performed. More specifically, for example, there is a possibility that a matching model (scoring model) that gives a relatively high score to a pattern that appears frequently (here, a topic distribution that appears frequently) is constructed in the training data. On the other hand, by generating a new synthetic variable that shows the characteristics of the pair of matching target persons, it is possible to avoid biased learning.

It is considered that the product of the topic values of both pairs represents the degree of overlap of the topic distributions of both pairs, and the absolute value of the difference between the topic values of both pairs represents the difference / spread of the topic distributions of both pairs. ..

The effect of using synthetic variables could be confirmed by conducting the following experiments [1] and [2]. That is, using the attribute data (matching target person data) of the already registered matching target person (company in this case), a large number of matching feature data for the pair of matter numbers (matching target person identification information) are prepared, and they are used. Was divided into training data (70%) and evaluation data (30%). Each of the training data (70%) and the evaluation data (30%) contains a positive example (correct answer data) and a negative example (incorrect answer data).

LDA was adopted for the "soft clustering or neural language model" for topic estimation processing, and random forest was adopted for the "supervised classification model" for matching processing (for scoring).

In the experiment [1], the seven needs flags of both pairs were not used as synthetic variables, but were included as 7 × 2 = 14-dimensional data in the matching feature data, and the topic distribution of both pairs (number of topics K =). As for 20), 54-dimensional matching feature data was created in total by including it as 20 × 2 = 40-dimensional data as it is without using it as a composite variable.

On the other hand, in the experiment [2], the composite variable (logical sum) based on the seven needs flags of both pairs is calculated and included in the matching feature data as 7-dimensional data, and the topic distribution of both pairs (number of topics K =). For 20), two types of composite variables (product and absolute value of difference) are calculated and included in the matching feature data as 20 × 2 = 40-dimensional data, and the composite variable of the document length of both pairs. (Ratio value) was calculated and included in the matching feature data as one-dimensional data to create a total of 48-dimensional matching feature data.

Then, the matching model (model for scoring) was trained using the training data (70%). At this time, the correct example (correct answer data) is tagged with selection (= 1), and the negative example (incorrect answer data) is tagged with non-selection (= 0).

Subsequently, evaluation data (30%) was input to the matching model (model for scoring) constructed by performing the above learning, and a score was obtained. Projects with a score of 0.5 or more were evaluated as selected, and projects with a score of less than 0.5 were evaluated as not selected.

In the experiment [1], the precision rate (Precision) = 0.767, the recall rate (Recall) = 0.750, and the F1 value (harmonic mean of the precision rate and the recall rate) = 0.756. Moreover, the accuracy (Accuracy) was 0.833.

On the other hand, in the experiment [2], the precision rate (Precision) = 0.781, the recall rate (Recall) = 0.803, and the F1 value (harmonic mean of the precision rate and the recall rate) = 0.792. Moreover, the accuracy (Accuracy) was 0.854.

Therefore, in both the experiment [1] and the experiment [2], high values were obtained for all of the precision, recall, F1 value, and accuracy, so that the effect of the present invention was remarkably shown and the experiment was performed. Since the value of [2] was higher than that of the experiment [1], it was found that more appropriate scoring can be performed by using the synthetic variable.

Further, since the matching system 10 includes the word importance vector calculating means 35, the inner product of the TFIDF vector or other word importance vectors can be included in the matching feature data. Therefore, depending on the word importance, it is possible to create matching feature data in which the features included in the combined text data are emphasized, and more appropriate scoring can be performed.

Further, the word importance vectors of both pairs can be used as they are as the matching feature data, but in the present embodiment, the inner product of the word importance vectors of both pairs is obtained, and the obtained inner product is used as the matching feature data. Therefore, the number of dimensions of the matching feature data can be reduced, and a more appropriate score can be obtained by generating the composite variable instead of using it as it is, as in the case of generating the composite variable with the topic distribution described above. You can make a ring.

Further, since the matching feature data creating means 36 can include the needs flag in the matching feature data, the types of needs of the applicant and each candidate are reflected in the matching feature data, which is more appropriate. Can perform scoring.

Then, the matching feature data creating means 36 uses a composite variable based on the needs flags (for example, logical sum, logical product, exclusive logical sum, negative logical sum, negative logical product, negative exclusive logical sum, arithmetic) of both needs flags. Since sum, etc.) can be included in the matching feature data, it is possible to generate and use the composite variable instead of using the needs flag as it is, as in the case of generating the composite variable with the topic distribution described above. Appropriate scoring can be performed.

Further, since the matching system 10 is a system for performing business matching, it is possible to assume an industry (business field) as each topic obtained by the topic estimation process. Therefore, matching feature data that accurately reflects the features of the combined text data can be created, and appropriate scoring can be performed. In other words, when performing business matching, since there are many business-related descriptions in the combined text data that combines self-data and needs data, it is possible to assume the type of business (business field) as a topic. -The number of topics K specified in Dirichlet Allocation (LDA) can be set to the number of industries (business fields) to be classified. Therefore, it is easy to match the design image of the system builder with the result of the topic estimation process by LDA, so that an appropriate system design can be performed.

<Form of deformation>

The present invention is not limited to the above-described embodiment, and modifications and the like within a range in which the object of the present invention can be achieved are included in the present invention.

For example, in the above embodiment, when one applicant (one individual, one corporation, or one organization) (matching requester) appears as a new matching target person, a topic model including applicant data is included. The explanation will be divided into the operation-time processing when the re-learning is not performed (see Fig. 8) and the operation-time processing when the topic model is re-learned including the applicant data (see Fig. 9). It was done, but the following operations can be performed. That is, instead of re-learning the topic model as shown in FIG. 9 each time one applicant appears, it is possible to deal with a plurality of applicants without re-learning the topic model as shown in FIG. When a certain number of applicants appear, the topic model may be relearned as shown in FIG. Specifically, for example, when the 1st to 9th applicants appear, the topic model is not relearned as shown in FIG. 8, and when the 10th applicant appears, the request is made. Including the applicant data for the 10th applicant (naturally, the applicant data for the 1st to 9th applicants is also included, but at that time, these 9 applicant data have already been registered. The matching target person data has already been stored and is stored in the matching target person data storage means 42), and the topic model may be relearned as shown in FIG.

In addition, the topic model may be updated (re-learned) at any time, and it is not necessary to update the topic model at the timing when the applicant appears. That is, it is not always necessary to perform the process during operation as shown in FIG. Specifically, for example, when the first to tenth applicants appear, the topic model is not re-learned as shown in FIG. Then, before the 11th applicant appears, the applicant data for the 1st to 10th applicants (at this stage, the 10 applicant data are already registered matching target data. The topic model is updated by re-learning including the matching target person data storage means 42). Therefore, updating the topic model in this case corresponds to the prior learning shown in FIG. 7. Then, when the 11th applicant appears, the updated topic model will be used, but as shown in FIG. 8, the topic model including the 11th applicant data will not be re-learned. To do.

Further, in the above-described embodiment, as shown in FIGS. 8 and 9, when a new matching target person (requester) appears, the applicant is matched with each candidate. Although it was configured, as shown in FIG. 10, it may be configured to perform matching for a specific purpose among the registered (existing) matching target persons. This specific purpose includes a standard specific purpose and a special specific purpose.

As for the standard specific purpose, if the number of registered (existing) matching targets is, for example, 100,000 companies, one of them is designated as a specific target person, and the remaining 99,999 companies are each. Matching feature data for 99,999 pairs will be created as candidates, and 99,999 scores will be output as scores indicating the probability that each candidate will be selected for a specific target person.

As a special specific purpose, if the number of registered (existing) matching target persons is, for example, 100,000 companies, a plurality of companies (for example, 10 companies) among them are input-designated as specific target persons, and the rest. 99,990 companies are candidates, and from the remaining 99,990 companies, the same number (for example, 10 companies) suitable for a plurality of specific target persons (for example, 10 companies) to hold a group meeting or the like. Select each candidate. In this case, feature data for matching between each of the plurality of specific target persons (for example, 10 companies) and each candidate of the remaining 99,990 companies is created, and the plurality of specific target persons (for example, 10 companies) are created. As a score indicating the probability that each candidate will be selected for each of the above, 10 × 99,990 scores are output. Then, the optimized calculation process is performed using the output score, and the same number (for example, 10 companies) of each candidate is selected so that all of the plurality of specific target persons (for example, 10 companies) are satisfied. However, the same number is an example and is not limited to the same number. It should be noted that each candidate with a score within the top 10 for one of a plurality of specific target persons (for example, 10 companies) is for another one of a plurality of specific target persons (for example, 10 companies). It usually does not match each candidate with a score within the top ten. Therefore, the optimization calculation process is performed so that each candidate of the same number (for example, 10 companies) selected is an appropriate matching partner from the viewpoint of any of a plurality of specific target persons (for example, 10 companies). To do.

The rule of this optimization calculation process is arbitrary. For example, when a plurality of specific target persons A, B, C, D, ... (For example, 10 companies) are specified, the same number for the specific target person A (for example, 10 companies) is specified. The total score of each candidate (10 companies), the total score of each candidate with the same number (for example, 10 companies) for the specific target B, and the same number (for example, 10 companies) for each candidate C. The total value of the scores of the persons can be obtained, and the same number of candidates (for example, 10 companies) can be selected so that the total value of these total values is maximized.

In addition, a rule that selects each candidate with the highest score ranking for each of a plurality of specific target persons (for example, 10 companies), and selects a candidate with a score ranking of 2nd or lower when duplication occurs. It may be. For example, a candidate having the highest score ranking for the specific target person A, a candidate having the highest score ranking for the specific target person B, a candidate having the highest score ranking for the specific target person C, and so on are selected. If there is no duplication in each of the selected candidates, the same number of candidates (for example, 10 companies) are determined. On the other hand, for example, if the candidate with the highest score ranking for the specific target person A and the candidate with the highest score ranking for the specific target person B match, one candidate is sufficient. Since there is no situation, the candidate with the higher score is selected from the candidate with the second highest score ranking for the specific target person A and the candidate with the second highest score ranking for the specific target person B. Further, for example, a candidate having the highest score ranking for the specific target person A, a candidate having the highest score ranking for the specific target person B, and a candidate having the first highest score ranking for the specific target person C. If they match, the number of candidates is two companies short, so the candidate with the second highest score ranking for the specific target person A and the second highest score ranking for the specific target person B are ranked second. Two companies with the highest scores are selected from the candidates and the candidates with the second highest score ranking for the specific target person C. At this time, the candidate with the second highest score ranking for the specific target person A, the second highest score ranking for the specific target person B, and the second highest score ranking for the specific target person C. If they match, the candidate is selected and one candidate is missing. Therefore, the candidate with the third highest score ranking for the specific target person A and the score for the specific target person B The candidate with the highest score is selected from the candidate with the third highest ranking and the candidate with the third highest score with respect to the specific target person C. When selecting candidates that are not the same number but twice the number, for example, select each candidate whose score rank is 1st and 2nd for each of a plurality of specific target persons (for example, 10 companies). , If duplication occurs, the same adjustment processing may be performed.

In FIG. 10, first, the service person in charge of receiving a request from the specific target person (one or more companies among the existing matching target persons) or the specific target person himself operates the input means 80 to operate the specific target person. When the identification information of the above is input and the specific target person is designated, this input is accepted by the input receiving means 31 (step S61). At the time of this reception, the topic model and the matching model have already been constructed by the prior learning process of FIG. 7.

At this time, in the case of the standard specific purpose of calculating the score between the specific target person (1 company) and each candidate, the identification information of the specific target person (1 company) may be specified. In the case of a special specific purpose of calculating the score between multiple specific target persons (for example, 10 companies) and each candidate for group meetings, group interviews, joint auditions, etc., those multiple identifications The identification information of the target person (for example, 10 companies) may be specified. In addition, the number of each candidate to be selected (for example, the same number, double number, triple number, etc., or numerical values of 5, 10, 15, 20, etc.) is also specified. In addition, in the case of a special specific purpose of investigating the compatibility between specific target persons (companies), the identification information of the specific target persons of both pairs is specified. Furthermore, in the case of a special specific purpose of investigating the compatibility between a specific target person (1 company) and a plurality of other specific target persons (for example, 10 companies), the identification information of the specific target person of both pairs. For example, when investigating the compatibility between the specific target person A and a plurality of other specific target persons B, C, D (the compatibility of A and B, A and C, and A and D), , The identification information of the specific target person A and the identification information of each of the specific target persons B, C, and D are specified.

The subsequent processes of steps S62, S63, S64, and S65 are substantially the same as those of steps S26, S27, S28, and S29 of FIG. 8 described above, except that the pairs are different. That is, in FIG. 8, a person who wishes to match as a new matching target person is included in the pair, but in FIG. 10, a specific target person (existing matching target person) is included in the pair. Only.

Further, in the above-described embodiment, the matching system 10 is a system for performing business matching that introduces a business partner, but the matching system of the present invention is not limited to this, for example, an intermediary service for buying and selling real estate. Matching between sellers and buyers, matching between the successor company (the company that transfers the business) and the successor company (the company that takes over the business) in the business succession support service, men and women in the marriage hunting support service A system that performs matching, matching between applicants for employment and hiring companies, matching between high school students who want to go on to school and universities who want to gather students, matching between employees who want to change personnel and departments within the company that recruit human resources, etc. However, in short, any system may be used as long as it is a system for matching between matching target persons who can prepare self-data and needs data. At this time, it is not necessary for the self-data and the needs data to be prepared by oneself, and a substitute person may perform the preparation. Therefore, even if it is an animal, a human can represent the description of its intention. In some cases, the animal can be a matching target, and therefore, for example, a system for matching between an animal and a zoo, matching between an animal and a trainer, matching between animals, and the like is also included in the matching system of the present invention. ..

Specifically, for example, in the case of a matching system between a seller and a buyer in a real estate sales brokerage service, a large number of seller attribute data (matching target data) and a large number of buyer attribute data (matching target data). Data) is used to prepare feature data for matching a large number of pairs of sellers and buyers. Then, for the matching feature data of these large numbers of pairs, for example, matching feature data for pairs that have been introduced as candidates for matching partners in the past is selected as a correct example (correct answer data) (= 1). ), And the matching feature data for the other pairs is tagged as a negative example (incorrect answer data) with non-selection (= 0), and the matching model is trained. It can be said that the self-data included in the attribute data (matching target person data) of the seller and the buyer shows the characteristics of the real estate, but since the real estate cannot have the needs data, the self-data is also the characteristics of the real estate. Rather, it is considered to indicate the characteristics of sellers and buyers. Further, since the seller and the buyer may be individuals or groups, the real estate sales brokerage service is in a state where matching of individuals and individuals, individuals and groups, and groups and groups is mixed.

Similarly, in the case of a matching system between the successor company (the company that transfers the business) and the successor company (the company that takes over the business) in the business succession support service, the attribute data of a large number of successor companies. Using (matching target person data) and attribute data of a large number of successor companies (matching target person data), feature data for matching a large number of pairs of the successor company and the successor company is prepared. , Learn the matching model. The needs data of the successor company (the company that transfers the business) includes a description of what kind of company wants the business to be taken over.

In addition, in the case of a matching system that matches men and women in the marriage hunting support service, a large number of male attribute data (matching target data) and a large number of female attribute data (matching target data) are used. , Prepare feature data for matching a large number of male and female pairs. Then, for the matching feature data of these large numbers of male and female pairs, for example, matching feature data for male and female pairs that have been introduced as candidates for matching partners in the past is selected as a correct example (correct answer data) (correct answer data). = 1) is tagged, and the matching feature data for other male and female pairs is tagged as a negative example (incorrect answer data) with no selection (= 0), and the matching model is trained.

It should be noted that, without distinguishing between men and women, matching feature data for a pair of matching subjects is prepared, that is, not only a pair of men and women but also a pair of men and a pair of women are prepared for matching feature data. , The matching model may be trained. In tagging between the same sex, it is possible to select (= 1) tagging of matching feature data for a pair that has been introduced as a friend as a correct example (correct answer data). In the matching process at the time of operation, if the person who wants to match is a man and wants to match with a woman, a plurality of men who want to match and all the registered women. Characteristic data for matching for each pair of females may be created, input to the matching model, and a score indicating the probability of each female candidate being selected may be output. Also, if the person who wants to match is a man and wants to match with a woman and also wants to find a friend, it will consist of the man who wants to match and all the registered men and women. Matching feature data for multiple pairs may be created and input into the matching model. A woman with a high score is a woman who is likely to develop into a relationship, and a man with a high score is a man who is suitable for referral as a friend.

In addition, matching employees who wish to transfer with departments within the company who want to secure human resources, matching sick people who want medical examinations with hospitals who want to secure patients, students who want to enroll, and schools who want to recruit students Even in the case of matching when there is a master-slave relationship between the matching targets, such as matching with, the matching model is learned in the same way as when the matching feature data for male and female pairs is prepared in the marriage hunting support service. Scoring can be performed with the constructed matching model.

As described above, the matching system and program of the present invention are suitable for use, for example, in the case of introducing a company as a business partner and performing business matching in which the companies are combined.

10 Matching system 32 Combined text data creation means 33 Topic estimation means 34 Applicant topic value prediction means 35 Word importance vector calculation means 36 Matching feature data creation means 37 Matching processing means 41 Applicant data storage means 42 Matching target person data storage Means 44 Topic model storage means 45 Word importance vector storage means 46 Matching model storage means

Claims (9)

A matching system consisting of a computer that executes a matching process that combines natural persons, corporations, non-corporate organizations, or other matching targets.
Needs data consisting of self-data consisting of text data describing the characteristics of each of a large number of matching target persons, not including those who wish to match as a new matching target person, and text data describing the needs of the matching partner. Appearance probability of each topic in the combined text data of each matching target person obtained by executing topic estimation processing by soft clustering or neural language model using a large number of combined text data obtained by combining and A topic model storage means that stores the topic value indicating the above in association with the identification information of the matching target person and stores the appearance probability of each word in each topic obtained by executing the topic estimation process.
An applicant data storage means for storing the self-data and the needs data of the matching applicant in association with the identification information of the applicant.
A combined text data creating means that combines the self-data and the needs data of the matching applicant stored in the applicant data storage means and executes a process of creating the combined text data of the applicant.
Using the combined text data of the applicant created by this combined text data creation means and the appearance probability of each word in each topic stored in the topic model storage means, each topic in the combined text data of the applicant is used. Applicant topic value prediction means that executes the process of predicting the topic value indicating the appearance probability,
The topic value indicating the appearance probability of each topic in the combined text data of the applicant obtained by this applicant topic value prediction means, and the matching target person as each candidate to be the matching partner stored in the topic model storage means. Matching feature data creation means that executes a process to create matching feature data for a plurality of pairs of applicants and candidates using a topic value indicating the appearance probability of each topic in the combined text data of When,
Using each of the plurality of matching feature data created by this matching feature data creating means as input data, a matching model that has been learned in advance by two-class classification of whether or not to be selected and stored in the matching model storage means is used. The feature is that it is equipped with a matching processing means that executes a processing that outputs a score indicating the probability that each candidate is selected for a matching applicant by performing a classification process using a supervised classification model. Matching system to do. A matching system consisting of a computer that executes a matching process that combines natural persons, corporations, non-corporate organizations, or other matching targets.
Needs consisting of self-data consisting of text data describing the characteristics of each of a large number of matching target persons, not including those who wish to be matched as a new matching target person, and text data describing the needs of the matching partner. A matching target person data storage means that stores data in association with the identification information of the matching target person,
An applicant data storage means for storing the self-data and the needs data of the matching applicant in association with the identification information of the applicant.
The self-data of each matching target person stored in the matching target person data storage means and the needs data are combined, and the self-data of the desired person stored in the desired person data storage means and the above-mentioned self-data. A combined text data creation means that combines with needs data and executes a process of creating combined text data of a large number of matching targets including applicants.
Using the combined text data of a large number of matching targets including the applicant created by this combined text data creation means, each topic in the combined text data of each matching target including the applicant by soft clustering or a neural language model. A topic estimation means that executes topic estimation processing to obtain a topic value indicating the appearance probability of
The topic value indicating the appearance probability of each topic in the combined text data of each matching target person including the applicant obtained by the topic estimation process by this topic estimation means is associated with the identification information of each matching target person including the applicant. Topic model memory means to memorize,
The applicant and each candidate are stored using the topic value indicating the appearance probability of each topic in the combined text data of the applicant and the matching target as each candidate to be the matching partner stored in this topic model storage means. Matching feature data creation means that executes the process of creating matching feature data for a plurality of pairs consisting of
Using each of the plurality of matching feature data created by this matching feature data creating means as input data, a matching model that has been learned in advance by two-class classification of whether or not to be selected and stored in the matching model storage means is used. The feature is that it is equipped with a matching processing means that executes a processing that outputs a score indicating the probability that each candidate is selected for a matching applicant by performing a classification process using a supervised classification model. Matching system to do. A matching system consisting of a computer that executes a matching process that combines natural persons, corporations, non-corporate organizations, or other matching targets.
Self-data consisting of text data describing the characteristics of each of a large number of matching target persons including the specific target person as an existing matching target person who performs matching for a specific purpose, and text data describing the needs for the matching partner. Each of the combined text data of each matching target person obtained by performing topic estimation processing by soft clustering or a neural language model using a large number of combined text data obtained by combining the needs data consisting of A topic model storage means that stores a topic value indicating the appearance probability of a topic in association with the identification information of the matching target person, and
Using the topic value indicating the appearance probability of each topic in the combined text data of each matching target person stored in this topic model storage means, from the specific target person and each candidate to be the matching partner of this specific target person. A plurality of pairs, a plurality of pairs consisting of each of the plurality of specific target persons and each candidate to be a matching partner of the plurality of specific target persons, from each of the specific target person and the other plurality of specific target persons. A matching feature data creation means that creates a plurality of matching feature data for a plurality of pairs, or creates a single matching feature data for a pair of specific target persons.
Each of the plurality of matching feature data created by this matching feature data creating means is used as input data, or one matching feature data is used as input data, and is learned in advance by two-class classification as to whether or not it is selected. A score indicating the probability that each candidate or other specific target person will be selected for a specific target person by performing classification processing by a supervised classification model using the matching model stored in the matching model storage means. A matching system characterized by being equipped with a matching processing means for executing a process of outputting. The matching feature data creation means
Using both topic values of the pair, a composite variable consisting of the product of both topic values for the same topic and a composite variable consisting of the absolute value of the difference between both topic values for the same topic are obtained. The matching system according to any one of claims 1 to 3, wherein a process of using each of the obtained synthetic variables as the matching feature data is executed. A word importance vector calculation means that executes a process of calculating a word importance vector consisting of a TFIDF value or other word importance index values of each word for the combined text data of each matching target person.
The word importance vector storage means for storing the word importance vector of the combined text data calculated by the word importance vector calculation means in association with the identification information of the matching target person is provided.
The matching feature data creation means
The inner product or cosine similarity between the word importance vector for one of the combined text data of the pair and the word importance vector for the other combined text data of the pair stored in the word importance vector storage means. The matching system according to any one of claims 1 to 4, wherein the inner product or cosine similarity of the obtained word importance vector is included in the matching feature data. .. The self-data and the needs data are accompanied by at least one needs flag indicating the type of needs for the matching partner.
The matching feature data creation means
The matching system according to any one of claims 1 to 5, wherein a process of including the needs flags of both of the pairs in the matching feature data is executed. The matching feature data creation means
Using both of the needs flags of the pair, the logical sum, logical product, exclusive OR, negative logical sum, negative logical product, negative exclusive OR, arithmetic sum, or other composition of both needs flags. The matching system according to claim 6, wherein the matching system is configured to obtain a variable and execute a process of including the obtained composite variable in the matching feature data. The matching target person is a natural person, a corporation, a non-corporate organization, or other matching target person who conducts business.
The matching process is a business matching process that combines matching target persons who conduct business.
Matching system according to claim 1 wherein the topic estimation process, characterized in that Ru processing der that by the Reitento-Dirichlet allocation. A program for operating a computer as the matching system according to any one of claims 1 to 8.

Images