JP7192039B1 - Matching system and program - Google Patents

Abstract

[Problems] To prevent biased referrals, prevent re-learning of matching models based on biased referral results, realize registration of appropriate case data for effective matching, and reduce the labor of service personnel. To provide a matching system that can achieve SOLUTION: In a matching system 10, when a user inputs a keyword into an input unit 110, associative search processing is executed, related items and related words are displayed on the same screen as the input unit 110, and the input unit 110 When the item identification information is entered in , matching processing and similar item search processing are executed, and on the same screen as the input unit 110, recommended items as matching partners and similar items to the target item are displayed. At this time, the associative search process, the matching process, and the similar case search process are all executed using the unique text data of each case. [Selection drawing] Fig. 11

Description

The present invention relates to a computer-configured matching system and program that executes a matching process that combines natural persons, corporations, non-corporation groups, or other projects, and can be used, for example, in a business matching system that combines companies.

In conventional business matching services, for companies that wish to find partner companies for business tie-ups or cooperation, service personnel select companies that can be introduced mainly by the following two methods.

The first method is a method based on a keyword search function in which a service person performs a keyword search and selects a matching partner company from the search results. In this first method, each company prepares its own data consisting of text data describing its own characteristics and needs data consisting of text data describing needs for a matching partner, and then prepares these text data. Search target.

The second method is based on an AI recommendation function that calculates a score that indicates the degree of matching accuracy between a company that is looking for a matching partner and any other company, and selects the company with the highest score. In other words, past data that summarizes the selection results of candidate companies to be matched (a set of data that shows what characteristics companies were selected as matching partners for what characteristics) In this method, a matching model is constructed by performing machine learning as learning data, and information on a company that is looking for a matching partner is input to this matching model, and a score is output. As a system for realizing this second method, a matching system developed by the applicant of the present application can be used (see Patent Document 1). In this matching system, the self-data and needs data of each company described above are combined, and using a large number of combined text data obtained, topic estimation processing by soft clustering or neural language model (for example, LDA: latent Dirichlet allocation, etc.) is executed, and matching feature data is created using topic values that indicate the appearance probability of each topic in the combined text data.

In order to realize the above-described first method, it is sufficient to have a system for inputting a keyword and extracting companies related to the keyword. An extraction system developed by man is known (see Patent Document 2). This extraction system was developed mainly as a system for extracting candidate stocks (i.e. companies issuing such stocks) to be constituent stocks of theme-type funds by inputting keywords. Investor Relations information (IR information) sent by each company to investors and home page information (HP information) of each company are used as unique text data of each company.

Publication of Japanese Patent No. 6802334 Publication of Japanese Patent No. 6596565

However, the first and second techniques described above have the following problems.

In both the keyword search of the first method and the AI recommendation by the matching model of the second method, the content of the text data describing each company greatly affects the search results and recommendation results. Therefore, from the viewpoint of business matching, text data that does not accurately represent company information or text data that has a small amount of information in the first place cannot provide desirable search results or recommendation results.

Further, the aforementioned self-data and needs data used for the keyword search of the first method and the construction of the matching model of the second method are created by the company itself that desires matching. At this time, it is unclear how to describe self-data and needs data effectively from the viewpoint of business matching, and it is difficult to distinguish between self-data and needs data. or omit one or the other. On the other hand, it is conceivable that the person in charge of the matching service will instruct each company on the appropriate description method, but it will take time and effort, and the content of the guidance will differ depending on the difference in experience and knowledge of the person in charge of the service. , the characteristics of the guidance by the person in charge of service appear directly in the contents of the description.

Furthermore, when performing a keyword search of the first method, it is necessary to input an appropriate keyword that matches the matching condition, which greatly affects the know-how of the person in charge of service, and thus limits the number of referrals. In addition, when using both the keyword search of the first method and the AI recommendation by the matching model of the second method, the number of introductions by keyword search is limited as described above, so from companies selected by keyword search However, as a result, the number of introductions of companies recommended by the matching model increases, resulting in a bias in the introduction results. As a result, the matching model re-learns the biased referral results, and the recommended companies are biased toward similar ones.

The above problems are not limited to business matching systems that connect companies. It is what happens.

In the case of business matching, it is matching between companies and companies, companies and sole proprietors, and sole proprietors and sole proprietors, and in the case of matchmaking support matching, it is matching between unmarried men and women. However, the essence of data processing is that it is the matching of information registered by those persons for the purpose of matching. In addition, matching can be performed even for animals such as pets and machines such as robots if humans can represent (describe) their possessed information. (natural persons, legal persons, unincorporated bodies, animals, robots, etc.) shall be referred to as "Items".

The purpose of the present invention is to prevent biased referrals, prevent re-learning of the matching model based on biased referral results, realize registration of appropriate case data for effective matching, and troublesome service personnel. It is to provide a matching system and program that can reduce the

The present invention performs three processes using each company's unique text data (self data and needs data, or combined text data of these): associative search process using keywords, and matching process for AI recommendation using a matching model. , and similar case search processing can be repeatedly executed to achieve the above object.

That is, the present invention is a matching system configured by a computer that executes a matching process that combines natural persons, legal persons, non-juristic organizations, or other cases,
Self-data consisting of text data describing the characteristics of each of a large number of projects, and needs data consisting of text data describing needs for matching partners, or obtained by combining these self-data and needs data Using the combined text data obtained as a large number of unique text data that describe information specific to each project, and performing topic estimation processing by soft clustering or neural language models, in the unique text data of each project topic model storage means for storing a topic value indicating the probability of occurrence of each topic in association with item identification information, and storing the probability of occurrence of each word in each topic obtained by executing the topic estimation process;
input receiving means for executing processing for receiving input of keywords or project identification information;
When the input of item identification information is received by the input reception means, the item of the received item identification information is treated as a subject item for matching processing, and each topic in the unique text data of each item stored in the topic model storage means. Using the topic value that indicates the probability of appearance, matching feature data is created for each of a plurality of pairs consisting of a target case and each candidate case that is a matching partner of this target case, and a plurality of created matching feature data are created. Using each of these as input data, using a matching model that has been pre-learned by two-class classification of whether or not it is selected and stored in the matching model storage means, classification processing by a supervised classification model is performed. a matching means for executing a process of calculating a score indicating the probability that each candidate project is selected for
When input of project identification information is received by the input reception means, in addition to processing by the matching means, a high degree of project similarity between the target project calculated using the unique text data of each project and other projects a similar case search means for executing a process of extracting similar cases of the target case in order;
When an input of a keyword is accepted by the input accepting means, the general similarity between the keyword calculated using the distributed representation of the words stored in the distributed representation model storage means and other words is taken as the final similarity. Alternatively, the final similarity obtained by calculating the final similarity using this general similarity and the unique similarity between the keyword calculated using the unique text data of each project and other words In addition to determining the related words of the keyword in descending order of degree, the number of occurrences for each word in which each of the keywords and related words appear in the unique text data of each case is obtained, and the obtained number of occurrences for each word and the number of related words are calculated. an associative search means for executing a process of determining related cases in descending order of the degree of relevance of each case to the keyword calculated using the final similarity;
When the input of a keyword is received by the input reception means, the related matter determined by the associative search means, or the related word in addition to the related matter, is displayed on the same screen as the input unit of the keyword or matter identification information by the input reception means. When the input reception means receives the input of project identification information, candidate projects with high scores calculated by the matching means are displayed as recommended projects on the same screen as the input unit, and similar projects are displayed. and output means for executing processing for displaying similar cases extracted by the search means on the same screen as the input part.

Here, "on the same screen as the input section" in the "output means" does not necessarily mean that the input section and the processing result display section are provided in the same window. When the main window provided with the input section and the small window provided with the input section are simultaneously displayed on the screen (including superimposed display), the input section and the processing result display section can be displayed on the screen of the same display means. exists at the same time, it corresponds to "on the same screen as the input section" here. In addition, when the screen size of the display means is small, or when a large amount of information is displayed on the display section as a result of processing, the input section and the display section of the processing result cannot be seen at the same time unless the screen is scrolled. Even if there is no input section, it corresponds to "on the same screen as the input section".

In such a matching system of the present invention, when a user (a person receiving a matching service or a person in charge of providing a matching service) enters a keyword into the input unit, the associative search means executes processing, Related cases are displayed on the same screen as the input section, or related words are displayed in addition to the display of the related cases. On the other hand, when project identification information is input to the input section, processing by the matching means and processing by the similar project search means are executed, and on the same screen as the input section, a recommended project as a matching partner and the input project identification information are displayed. Similar projects to the target project are displayed.

At this time, since the three processes by the associative search means, the matching means, and the similar case search means are all executed using the unique text data of each case, that is, the same data, all of the three processes are suitable for matching. In addition to being able to implement the same processing, it is possible to make the processing contents related to each other. Therefore, the matching system of the present invention is not simply a combination of the three existing systems, but enables each process to share data and to standardize the contents of the process.

In addition, the results of each process by the associative search means, matching means, and similar case search means can be used for the next process by these means, and the processes by these means can be repeated. , a variety of searches can be performed.

More specifically, when a keyword is entered and a job related to the keyword is obtained through processing by the associative search means, if the job identification information of the related job is entered, the related job can be recommended as a matching partner. A project is obtained, and similar projects to the related project are obtained. Then, by inputting the project identification information of the obtained recommended project, a recommended project as a matching partner for the recommended project can be obtained, and similar projects to the recommended project can be obtained. Further, by inputting the project identification information of the obtained similar project, a recommended project as a matching partner for the similar project can be obtained, and a similar project for the similar project can be obtained. In this way, it is possible to repeatedly search across related projects, recommended projects, and similar projects, and it is also possible to input various keywords at any timing during the search.

For this reason, if the user is a person who registers a new project or a service representative who assists, guides, and guides the user, he or she may be in the same position as the customer he/she is in charge of (for example, a certain product or merchandise). It is possible to easily grasp the contents of the project data (self-data and needs data included in it) regarding companies that wish to match for the purpose of selling the , it is possible to easily grasp what kind of project data is recommended as a matching partner (that is, what kind of project data has a high degree of matching accuracy). Therefore, the information referred to in those searches can be utilized when registering the item data of the new item. Therefore, the user can register suitable job data from the viewpoint of matching and receive effective matching service.

In addition, since the person who registers the new item can search and refer to the item data by himself/herself, it is possible to reduce the burden on the person in charge of the service. Furthermore, for example, even when a service representative conducts a search at a service counter or at a customer's site, it is possible to directly show the contents of project data of several reference projects obtained by the search to the person who registers a new project. Therefore, it is possible to reduce the burden on the person in charge of service (time and effort in providing assistance, guidance, and guidance).

In addition, the matching means produces the same output for the same input unless the topic model obtained by the topic estimation process or the matching model obtained by learning using the matching feature data is updated. , the same item is always output as a recommended item with the same rank (same score) for a certain item. And if this becomes a selection result and is reflected in the updating of the matching model, it will lead to promoting biased introductions. However, according to the present invention, by performing processing by the similar case search means, it is possible to grasp cases similar to the target case for which matching is desired. Since the recommended cases are output, not only the recommended cases for the target case, but also the recommended cases for similar cases to the target case can be output, and the service staff can introduce them. Similarly, by repeating the search, it is possible to output similar cases of the recommended cases to the target case and similar cases of the recommended cases to the similar cases of the target case, and the service personnel can introduce them. Therefore, since the service staff can find a project to introduce from various angles, even a staff member with little know-how can introduce a suitable project as a matching partner to the matching requester, and the introduction project Therefore, it is possible to prevent the matching model from being updated by biased learning data, thereby achieving the above object.

It should be noted that when a keyword is entered, in addition to the output of the related items obtained by the processing by the associative search means, the related words of the keyword may also be output. can be entered as a new keyword, which enhances the convenience of searching. In this case, the related words may be determined in descending order of the final similarity calculated using the general similarity and the specific similarity. Instead of general related words (words with high general similarity to the input keyword), it is possible to use related words that reflect the content of the unique text data of each case, making it even more effective from the viewpoint of matching. related words can be input as new keywords. In other words, in the processing by the associative search means, the related item is determined using the unique text data of each item, and the related word used in determining the related item is also determined using the unique text data of each item. As a result, more effective processing can be realized from the viewpoint of matching.

<Configuration that reflects the unique text data of the new matter in the search without updating the topic model>

Also, in the matching system mentioned above,
Occurrence of each topic in the unique text data of the new issue using the unique text data of the new issue not used in the topic estimation process and the appearance probability of each word in each topic stored in the topic model storage means. A topic value prediction means for executing a process of predicting a topic value indicating a probability;
The matching means are
The topic value indicating the appearance probability of each topic in the unique text data of the new project obtained by the topic value prediction means, and each of the unique text data of each candidate project to be matched with the new project stored in the topic model storage means. matching feature data for a plurality of pairs consisting of a new case and each candidate case using a topic value that indicates the appearance probability of a topic, and using each of the created plurality of matching feature data as input data, Classification processing by a supervised classification model is performed using a matching model pre-learned by two-class classification of whether or not to be selected and stored in a matching model storage means, so that each candidate for a new case is classified. It is configured to execute a process of outputting a score indicating the probability of being selected,
The similar project search means is
It is configured to execute a process of extracting similar projects, including the project similarity between the new project and other projects calculated using the unique text data of the new project and each other project,
The associative search means is
Keywords and related words are calculated using the number of occurrences of each word in the specific text data of each project, including new projects, and the final similarity of related words. It may be configured to execute the process of determining the related cases in the order of the degree of relevance of each case to the high order.

In this way, if the topic value prediction means is provided and the unique text data of the new item is reflected in the search without updating the topic model, then the topic estimation process including the unique text data of the new item will be performed. Even before the topic model is updated, it is possible to perform matching processing with a new matter as the target matter. Therefore, before updating the topic model, it is possible to perform a search by repeating each process by the associative search means, the matching means, and the similar case search means, including new cases.

<Details of vector processing using unique text data by similar project search means>

Furthermore, in the matching system described above,
The similar project search means is
A topic distribution vector consisting of topic values indicating the occurrence probability of each topic in the unique text data of each item stored in the topic model storage means, and the mean value of the distributed expression of each word constituting the unique text data of each item and a word importance vector consisting of the TFIDF value of each word or other word importance index values for the unique text data of each project. It may be configured to execute the process of extracting similar cases of the target case in descending order of the case similarity between the case.

In this way, when the similar project search means performs vector processing using unique text data, the content of description of the unique text data (own data, needs data) is accurately reflected, and similar projects useful for matching are found. An extraction process can be implemented.

<Details of search history display processing>

And in the matching system mentioned above,
Search history storage means for storing the keyword or project identification information received by the input reception means and log information including the input order thereof,
The input reception means is
Using the log information stored in the search history storage means, in addition to indicating the input type and input order of whether the input of the keyword or the matter identification information was accepted,
Recommendation/similarity distinguishing information for distinguishing whether input of item identification information of a recommended item for the target item has been received or whether input of item identification information of a similar item to the target item has been received;
While repeatedly accepting input of project identification information for recommended projects and project identification information for similar projects, from the perspective of whether the project for which the input of project identification information is received is a project on the matching requesting side or a matching partner side. Therefore, for recommended projects, the projects on the opposite side of the projects for which input of project identification information was accepted, for similar projects, projects on the same side as the projects for which the input of project identification information was accepted, and for recommended projects for recommended projects is side information that distinguishes whether each matter for which the input of matter identification information is on the same side or on the opposite side by making the matter on the same side as the matter for which the matter identification information has been input. Execute search history display processing including information on at least one of
Alternatively, if the project data of each project contains a needs flag indicating whether it is for sale or purchase, whether it is a worker or an employer, whether it is a man or a woman, or other paired information, recommend or similar A configuration may be adopted in which search history display processing including at least one of the distinguishing information and the side information and including a needs flag is executed.

As described above, in addition to the usual search history display processing indicating the input type and input order of which input of keyword or item identification information has been received, at least one of recommendation/similarity distinction information and side information is performed. If the search history display processing including the information of and the search history display processing including the needs flag are performed, during the search that repeats each process by the associative search means, the matching means, and the similar case search means , users can easily understand what they have been doing, what they are doing now, and what they should do from now on, so they can perform efficient searches. .

<Configuration for displaying a reference project when registering a new project>

Also, in the matching system mentioned above,
a new matter registration means for accepting input of matter data of a new matter, assigning matter identification information, and storing the received matter data in the matter data storage means in association with the given matter identification information;
The output means are
a process of storing the item identification information of a related item, a recommended item, or a similar item selected as a reference item when registering a new item, or the item data thereof in a reference item storage means at the time of registration;
The new project registration means is
Acquiring the case data of the reference case from the case data storage means using the case identification information of the reference case stored in the reference case storage means at the time of registration, or obtaining the case data of the reference case stored in the reference case storage means at the time of registration A configuration may be adopted in which case data is acquired, and the process of displaying the acquired case data of the reference case on the same screen as the input unit for the case data of the new case may be executed.

In this way, when a reference case is displayed at the time of registering a new case, the user who registers the new case has to input appropriate case data (including self data and needs data) as data for matching. can be easily done.

<Details of Specific Similarity Calculation Processing in the Case of Using General Similarity and Specific Similarity by Associative Search Means to Calculate Final Similarity>

Furthermore, in the matching system described above,
The associative search means is
configured to execute a process of calculating the final similarity using the general similarity and the unique similarity, and
Based on the number of appearances of general similar words in set P of all projects that include both keywords and general similar words selected by high general similarity in unique text data, or set P An evaluation value Ps of the set P based on the number of occurrences of keywords and the number of occurrences of general similar words in
Using the evaluation value Qs of the set Q based on the number of occurrences of the general similar words in the set Q of each case in which the specific text data does not contain the keyword and contains the general similar words,
A configuration may be adopted in which the process of calculating the unique similarity is performed by Ps/(Ps+Qs).

In this way, when the associative search means is configured to calculate the unique similarity by using the number of appearances of words, the unique similarity that accurately reflects the description content of the unique text data of each case is calculated, and the unique similarity is calculated. It is possible to calculate a final similarity using the similarity and determine related words using the final similarity.

<Structure for accepting input of keywords and input of project identification information in the same input unit>

And in the matching system mentioned above,
The input reception means is
After the input of the keyword and the input of the item identification information are accepted by the same input unit, a process of determining whether the information input to the input unit is the keyword or the item identification information may be executed.

When the same input unit accepts the input of a keyword and the input of project identification information in this way, the configuration of the screen for disposing the input unit and displaying the search results is simplified, and the operation by the user is simplified. It is possible to improve the quality.

<Configuration that can receive simultaneous input of multiple item identification information>

Also, in the matching system mentioned above,
The input reception means is
It is configured to be able to receive simultaneous input of multiple project identification information,
The matching means are
When simultaneous input of a plurality of item identification information is received by the input reception means, each item of the received plurality of item identification information is grouped together as a target item group for matching processing and stored in the topic model storage means. Using the topic value that indicates the appearance probability of each topic in the unique text data of each project, each of the multiple target projects that make up the target project group and each candidate project that is a matching partner for each of these multiple target projects and each of the plurality of created matching feature data is used as input data, and is pre-learned by two-class classification as to whether or not it is selected, and the matching model is stored. By performing classification processing using a supervised classification model using the matching model stored in the means, a score indicating the probability that each candidate project is selected for each of the multiple target projects that make up the target project group is calculated. calculate,
Select multiple candidate projects that have a large total score between each of the multiple target projects that make up the target project group and each of the multiple candidate projects that are matching partners, and select multiple candidate projects are configured to execute a process of determining as a plurality of recommended cases constituting a recommended case group,
The similar project search means is
When simultaneous input of multiple project identification information is received by the input reception means, calculation is performed using the unique text data of each project for each of the multiple target projects that make up the target project group in conjunction with processing by the matching means It is configured to execute the process of extracting similar cases of the target case in the order of high degree of case similarity between the target case and other cases,
The output means are
When simultaneous input of a plurality of project identification information is received by the input reception means, a plurality of recommended projects constituting a recommended project group determined by the matching means are displayed on the same screen as the input unit, and a similar project search is performed. It may be configured to execute a process of displaying similar cases for each of a plurality of target cases constituting the target case group extracted by the means on the same screen as the input unit.

In the case where the configuration is such that it is possible to receive simultaneous input of multiple project identification information, the similar project search means extracts similar projects for each of the multiple target projects that make up the input target project group. Since it is displayed and displayed, it is possible to decide the project group to introduce from various viewpoints, and it is possible for even service personnel with little know-how to perform appropriate introduction work, including biased introduction results It is possible to prevent the matching model from being updated by the learning data. That is, the matching means produces the same output for the same input unless the topic model obtained by the topic estimation process or the matching model obtained by learning using the matching feature data is updated. , the same recommended case group is always output for a certain target case group. However, according to the present invention, at least one of the plurality of target cases that make up the target case group is changed to a similar case of those target cases to perform a search, or a plurality of cases that make up the recommended case group are searched. It is possible to change at least one item out of the recommended items to an item similar to those recommended items and introduce it.

<Program Invention>

A program of the present invention is for causing a computer to function as the matching system described above.

Note that the above program or part thereof can be, for example, a magneto-optical disk (MO), a compact disk (CD), a digital versatile disk (DVD), a flexible disk (FD), a magnetic tape, a read-only memory (ROM) , electrically erasable and rewritable read-only memory (EEPROM), flash memory, random access memory (RAM), hard disk drive (HDD), solid state drive (SSD), flash disk, etc. It is possible to store, distribute, etc. by using, for example, a local area network (LAN), a metropolitan area network (MAN), a wide area network (WAN), the Internet, an intranet, an extranet, etc. It can be transmitted over a transmission medium such as a wired network, a wireless communication network, or a combination thereof, or it can be carried on a carrier wave. Furthermore, the above program may be 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, it is possible to perform a search by repeating the three processes of the associative search means, the matching means, and the similar case search means. It is possible to prevent the re-learning of the matching model based on the system, to realize the registration of appropriate project data for effective matching, and to reduce the labor of the person in charge of the service.

1 is an overall configuration diagram of a matching system according to an embodiment of the present invention; FIG. FIG. 4 is an explanatory diagram of processing using distributed representation of words according to the embodiment; FIG. 10 is an explanatory diagram of a process of calculating a unique similarity B according to the embodiment; FIG. 10 is an explanatory diagram of the calculation processing of the degree of association CS according to the embodiment; FIG. 10 is an explanatory diagram of the unique text data preparation process according to the embodiment; FIG. 4 is a diagram showing a specific example of data from preparation of unique text data to estimation of topic distribution in the embodiment; FIG. 4 is an explanatory diagram of a method for creating synthetic variables to be included in matching feature data according to the embodiment; FIG. 4 is an explanatory diagram of a method for creating matching feature data when the embodiment is operated; FIG. 4 is an explanatory diagram of a method of preparing matching feature data during learning according to the embodiment; FIG. 4 is an explanatory diagram of calculation processing of the case similarity degree according to the embodiment; Explanatory drawing which shows the flow of the search in the matching system of the said embodiment. The figure which shows the example of a screen which displayed the result of the process by the associative search means of the said embodiment. FIG. 10 is a diagram showing an example of a screen displaying results of each process performed by the matching means and the similar case search means of the embodiment; The figure of the flowchart which shows the whole flow of the pre-processing of the matching system of the said embodiment. The figure of the flowchart which shows the whole flow of the search processing by the matching system of the said embodiment. FIG. 10 is a flowchart showing the overall flow of processing when the time comes to update a topic model including new item data according to the embodiment; FIG. 4 is a flowchart showing a detailed flow of pre-processing for associative search in the embodiment; The figure of the flowchart which shows the detailed flow of the pre-processing for matching of the said embodiment. The figure of the flowchart which shows the detailed flow of the pre-processing for similar matter search of the said embodiment. FIG. 4 is a flowchart showing a detailed flow of associative search processing of the embodiment; The figure of the flowchart which shows the detailed flow of the matching process of the said embodiment. The figure of the flowchart which shows the detailed flow of the similar matter search process of the said embodiment. FIG. 4 is a flowchart showing the flow of new matter registration processing according to the embodiment; The figure which shows the form of 1st modification of this invention. The figure which shows the form of 2nd modification of this invention. The figure which shows the form of a 3rd modification 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 a matching system 10 of this embodiment. This matching system 10 is a business that introduces business partners (including industry-academia collaboration) to businesses (mainly companies, but may also be sole proprietors or non-profit organizations such as universities). It is a system for implementing a matching service. In the specification of the present application, the term "enterprise" may be used in the sense of a business operator including sole proprietors and non-profit organizations such as universities.

As a form of service, upon receiving a request for matching from a business operator, the person in charge of the service exclusively searches for a matching partner that meets the conditions and introduces it to the requester. etc., the service representative and the customer (requester) conduct a search while talking to each other to search for a matching partner that meets the conditions, and the service representative discloses the name and contact information of the selected partner to the requester. The matching requester may conduct a search (including online search from a remote location), search for a matching partner that meets the conditions, and provide the name and contact information of the selected partner to the person in charge of the service. (Service company) may be introduced. Therefore, the subject of the search is the business operator who wishes to be matched, or the person in charge of the service who assists, guides, and guides the business operator.

2 to 10 show details of each process performed by the matching system 10. FIG. 11 to 13 are explanatory diagrams of the search function of the matching system 10 executed in accordance with user's operation. 14 and 15 show the overall flow of pre-processing and search processing during operation of the matching system 10, FIG. 16 shows the flow of processing such as topic model updating, and FIGS. shows the flow of preprocessing for associative search, matching, and similar case search, and the flow of associative search process, matching process, similar case search process, and new case registration process is shown in FIGS. Each is shown in a flow chart.

<Overall Configuration of Matching System 10>

In FIG. 1, the matching system 10 is configured by a computer and includes a main body 20, display means 80 such as a liquid crystal display, and input means 90 such as a mouse, keyboard, touch panel, or the like.

The main body 20 comprises a processing means 30 for executing various processes related to matching and accompanying associative search and similar case search, and a storage means 50 for storing various data necessary for the processing by the processing means 30. ing.

The processing means 30 includes a setting means 31, a distributed representation model learning means 32, an item data preparation means 33, a unique text data preparation means 34, a topic estimation means 35, a topic value prediction means 36, and a word importance vector. Calculation means 37, matching learning data preparation means 38, matching model learning means 39, input reception means 40, associative search means 41 (including word similarity calculation means 41A and relevance calculation means 41B), and matching means 42 (including matching feature data creation means 42A and matching score calculation means 42B); similar case search means 43 (including case similarity calculation means 43A and similar case extraction means 43B); output means 44; It is configured including an item registration means 45 .

Here, each means 31 to 45 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 Realized by a program. Details of each of these means 31 to 45 will be described later.

The storage means 50 includes setting data storage means 51, distributed representation model learning data storage means 52, distributed representation model storage means 53, case data storage means 54, unnecessary word dictionary storage means 55, and topic model storage means. 56, word importance vector storage means 57, matching candidate selection record list storage means 58, matching learning data storage means 59, matching model storage means 60, search history storage means 61, model similarity storage. means 62, general similarity storage means 63, unique similarity storage means 64, final similarity storage means 65, relevance storage means 66, text vector storage means 67, case similarity storage means 68, It is configured including a reference case storage means 69 at the time of registration.

Here, as each of the storage means 51 to 69 included in the storage means 50, for example, a hard disk drive (HDD), a solid state drive (SSD), etc. can be adopted. (storage area temporarily secured in work memory, etc.). This point will be described later. Details of each of the storage means 51 to 69 will also be described later.

In addition, in FIG. 1, the matching system 10 is described as having a stand-alone configuration, but it may have a system configuration that communicates via a network. When the matching system 10 is of the server-client type, the main body 20 can be composed of one or a plurality of servers, and the display means 80 and the input means 90 can be provided on the terminal side. In this case, 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. does not matter. Depending on the type of service, the terminal may be mainly the terminal of the person in charge of the service, or the terminal of the matching requester (the person who receives the matching service). When the service representative and the matching requester view the same physical screen at the same time on one physical terminal, and when the service representative and the matching requester who are remote from each other share a screen like an online conference on separate terminals There are cases where it is viewed with If the matching requester in a remote location directly inputs his or her project data or directly sees the matching result (scoring result) on the screen, the network should be an external network such as the Internet. , the matching requester's terminal is connected to the main body 20, which is a server. In addition, in any service form, a system administrator's terminal is installed to manage system construction, maintenance, update, and the like.

<Configuration of Processing Means 30>

(Configuration of processing means 30/setting means 31)
The setting means 31 receives input of various setting data from the input means 90 by the system administrator, and executes processing for storing the data in the setting data storage means 41 .

Further, the setting means 31 may receive input of various setting data from the input means 90 by the user (person in charge of service) and execute processing for storing the data in the setting data storage means 41 . In this case, setting data for each user (service staff) is stored in the setting data storage means 41 in association with user identification information (service staff identification information). For example, it is saved in a format such as a setting file for each user. It should be noted that the setting data by the user may be input each time without being saved. will be

Specifically, when the word similarity calculation means 41A determines general similar words (details will be described later, as a result of processing, N5 words (for example, N5=70) are included in the setting data. N3 (for example, N3 = 100) of the model-specific similarities S in each distributed representation model used in ), and N5 general similar words are selected and determined by the word similarity calculation means 41A When obtaining general similarity A by integrating the number of related words N6 (for example, N6=10) and similarity by model S in a plurality of distributed representation models by word similarity calculation means 41A, weighted harmonic average or weighted average , and the coefficients σ and τ in the formula (1) and the coefficient φ in the formula (2) when obtaining the unique similarity B by the word similarity calculation means 41A when using , the weight for the general similarity A and the specific similarity B when using the weighted harmonic average or the weighted average when calculating the final similarity FS using the general similarity A and the specific similarity B by the word similarity calculation means 41A A coefficient value, a weighting coefficient κ(i) in Equation (3) when calculating the degree of relevance CS by the degree-of-relevance calculating means 41B, and There are weighting factors Wtf-idf, Wword2vec, and Wtopic in equation (4). In addition, the number of related cases (processing results of the associative searching means 41), recommended cases (processing results of the matching means 42), and similar cases (processing results of the similar case searching means 43) displayed on the screen by the output means 44 (for example, , within the top 10, etc.) are also set data. Note that these setting data may be fixed values.

Further, the setting means 31 stores the text data for distributed representation model learning acquired and collected in advance in the distributed representation model learning data storage means 52 by the operation of the input means 90 by the system administrator. Also run

(Configuration of Processing Means 30/Distributed Representation Model Learning Means 32)
The distributed representation model learning means 32 uses the text data for distributed representation model learning stored in the distributed representation model learning data storage means 52 to perform morphological analysis and machine learning using various algorithms, and obtains learning results. Data (in this embodiment, a weight matrix W that is a set of high-dimensional word vectors as distributed representations of words) is stored in the distributed representation model storage means 53 (see FIG. 2). The distributed representation model obtained by learning is used in the processing by the associative search means 41 and the process by the similar case search means 43 as shown in FIG. Well, you can prepare them separately. In this embodiment, as an example, they are shared. It should be noted that the disclosed/provided learning result data (weight matrix W) may be acquired and stored in the distributed representation model storage means 53 .

Here, text data of news, text data collected from the Internet such as Wikipedia, and the like are used as the text data for distributed representation model learning, for example. The text data for this distributed representation model learning is data containing N1 (for example, N1=1.8 billion) words, and is a huge amount of data. It contains N2 (for example, N2=1,150,000) unique words without overlapping.

The word vectors stored in the distributed representation model storage means 53 (see FIG. 2) include, for example, continuous bag of words (CBOW), which is a word2vec model, and skip Skip-gram or ELMo by Long Short Term Memory (LSTM), a type of GloVe, fastText, Recurrent Neural Network (RNN) Furthermore, word vectors obtained by algorithms other than neural networks (NN), such as LINE, can also be used.

Specifically, as shown in FIG. 2, a weight matrix W (rows are word vectors), which is a set of word vectors obtained by learning by word2vec or the like, can be used. This weight matrix W is called a word vector lookup table or the like, and is a matrix of the number of vocabularies N2 (for example, N2=1,150,000)×the number of dimensions. In the present embodiment, as an example, three types of dimensions (=100, 200, 400) are set to window size = 12 for all of them to prepare three types of weight matrices W1, W2, and W3. The distributed representation model of the type is stored in the distributed representation model storage means 53 (see FIG. 2). These three types of distributed representation models are considered to be one integrated model because the model-specific similarity S by each distributed representation model is ensembled by a harmonic average or a weighted harmonic average as described later. good too. The number of dimensions is the number of dimensions of a word vector, and in the case of word vectors obtained by a neural network (NN) algorithm such as word2vec, it is equivalent to the number of nodes in the intermediate layer (hidden layer) of the NN.

Note that the number of distributed representation models to be prepared may be only one, and when a plurality of distributed representation models are prepared, the number of types is not limited to three, and may be two or four or more. Four or more types of distributed representation models may be prepared by setting multiple types of window sizes for each of the types of dimensions.

In addition, the word vectors used as distributed representations of words in the present invention are not necessarily word vectors as learning result data obtained by machine learning using various algorithms. For example, learning result data obtained by machine learning Word vectors created by processing, adjusting, or mixing word vectors (such as combining word vectors obtained by machine learning with different algorithms), or created by performing aggregation work, etc. by methods that do not rely on machine learning It may be a word vector or the like.

(Configuration of Processing Means 30/Item Data Preparing Means 33)
The case data preparation means 33 associates the case data of a large number of cases (in this case, companies) prepared for the purpose of receiving the matching service with the case identification information (in this embodiment, case numbers). It executes processing to be stored in the case data storage means 54 . Data used in an existing system that provides a matching service may be used as the item data. Note that the project identification information is not limited to the project number consisting of numbers as in the present embodiment, and may be identification information including characters, symbols, and the like.

The project data includes company information such as company name, location, telephone number, website address, representative name, person in charge and their positions, industry and main products handled, self-promotion (text data in free description format). ), capital, number of employees, annual sales (sales), etc. In addition, as the needs information required of the business matching partner, the desired industry (business type of the partner company), desired area (location of the partner company), and the type of needs correspond to each item prepared in advance by the service company (matching service provider). It includes a needs flag, needs details (text data in a free description format), etc. Of these, important data for executing matching in the present invention are self-PR and needs details, which are text data in a free description format. "Self-data" in the present invention (claims) is data centered on self-PR (it may include self-PR or may be only self-PR), and "needs data" is centered on details of needs. (It is sufficient if the details of the needs are included, or only the details of the needs are sufficient.) This point will be described later with reference to FIGS. 5 and 6. FIG.

As shown in FIGS. 5 and 10, the need flag is data indicating the type of need (data of 1 and 0 indicating whether or not it is applicable), and in this embodiment, seven flags are provided as an example. , the number of needs flags to be set is arbitrary, and may be one or more. In addition, the content and the degree of coarseness can also be set arbitrarily. Need flags such as "buy materials", "buy parts", and "buy materials" may be provided. For example, "sell" and "seeking a buyer" have the same content, and "buy" and "seeking a supplier" have the same content, so the difference in expression is a problem. do not become.

Similarly, a needs flag saying "manufacture at our company" and a needs flag saying "manufacture at your company" may be provided. Need flags such as "manufacture", "manufacture parts", and "manufacture materials" may be provided. In addition, for example, "We will build a system", "Please build a system", "We will transport products/merchandise", "Please transport products/merchandise", "We will undertake advertising , "Please do an advertisement", "We will dispatch personnel", "Please dispatch personnel", "We will sell retail", "We are looking for a retail store", "We will undertake nationwide expansion" , "We would like to expand nationwide", "We will provide real estate", "We are looking for real estate", "We will provide services", "We are looking for services", "Delivery will be on the same day or Need flags such as "It is the next day" and "I want the delivery date to be the same day or the next day" can be set. Also, a need flag of "others" may be provided.

(Configuration of Processing Means 30/Unique Text Data Preparing Means 34)
The unique text data preparation means 34 uses the job data of each job (each company) stored in the job data storage means 54 to prepare each data necessary for the matching process and the accompanying associative search process and similar job search process. It executes a process of preparing unique text data of the case (each company), and executes a process of storing the processed unique text data in the case data storage means 54 in association with the case identification information (case number). .

Note that the unique text data after processing may be stored in a unique text data storage means separately from the case data storage means 54 in association with the case identification information (case number). In this embodiment, the processed unique text data is stored in the item data storage means 54, but the processed unique text data portion stored in the item data storage means 54 is stored as the unique text data storage means. You can call Therefore, the number of databases and the number of tables to be installed to store the item data and the unique text data after processing is arbitrary, and regardless of the data division format, the item identification information (item number) It is sufficient if they are associated with each other.

The unique text data is used in any of the processes of the associative search means 41, the matching means 42, and the similar case search means 43. FIG. In this embodiment, as shown in FIG. 5, for the processing by the matching means 42, combined text data obtained by combining self data and needs data is used as specific text data. On the other hand, as for the processing by the similar item search means 43, as shown in FIG. 10, the self data and the needs data are not combined, and the self data and the needs data in separate states are used as unique text data. However, in the processing by the matching means 42, the self-data and the needs data may be used separately as unique text data without being combined. You may use the combined text data which combined this and needs data as specific text data.

As for the processing by the associative search means 41, in this embodiment, the unique text data (in the state of combined text data) used in the processing by the matching means 42 are shared. However, in the processing by the associative search means 41, either self data or needs data may be used as unique text data. The extraction system of Patent Document 2 developed by the applicant of the present application mentioned above is a system for inputting a keyword and extracting companies related to the keyword, so it can be applied to the processing by the associative search means 41. . At this time, the extraction system of Patent Document 2 was developed mainly as a system for extracting candidate stocks (i.e. companies issuing such stocks) to be constituent stocks of theme-type funds by inputting keywords. Investor Relations information (IR information) sent by each company to investors and homepage information (HP information) of each company are used as unique text data of each company for extraction. When applied to the matching system of No. 2, IR information and HP information are replaced with self data, needs data, or combined text data described for the purpose of matching.

The processing by this unique text data preparation means 34 is executed in the stage of preprocessing (see FIG. 14). Also, when registering a new item (see FIG. 23), upon receiving an instruction from the new item registration means 45, it is executed on the unique text data of the new item as immediate reflection processing of the new item before updating the topic model or the like. be. Furthermore, it is also executed when the time for updating the topic model comes (see FIG. 16), but it is not necessary to repeatedly execute the processing that is executed in the instant reflection processing of the new item.

(Configuration of processing means 30/unique text data preparation means 34: Combined text data creation processing, FIGS. 5 and 6)
The unique text data preparation means 34 performs [α] processing for creating combined text data, [β] processing for removing unnecessary symbols and tags, and [γ] processing for breaking down into words by morphological analysis and extracting only nouns. and [δ] unnecessary word removal processing are executed. The data obtained by executing the process [α] and the data obtained by executing the processes [α] to [δ] are both called combined text data (unique text data). At this time, after the [γ] process, it will be broken down into words. Therefore, for convenience of explanation, it is called combined text data (unique text data). Also, when the combined text data (unique text data) that has undergone the processing up to [δ] is specifically distinguished, it will be referred to as processed combined text data (unique text data after processing).

[α] In the process of creating the combined text data, the unique text data preparation means 34, as shown in FIG. Combined text data is created by combining self-data consisting of text data obtained from the applicant and needs data consisting of text data that combines the desired industry (matching partner industry) and detailed needs (text data in free-description format). . The resulting combined text data is handled as one piece of document data.

At this time, as shown in FIG. 6, the unique text data preparation means 34 connects self-industry, self-PR (free description format), desired industry, and needs details (free description format) with a full stop (.) in between. Here, self-data includes self-industry and needs data includes desired industry. Needs details (free description format) alone may be used as needs data.

Also, since the combined text data created by the unique text data preparation means 34 is text data, when the self-industry or the desired industry are selectable symbols or numbers (1, 2, . . . ), Those industries are converted into text data (electrical machinery, construction/civil engineering, IT, etc.) and then combined. On the other hand, if the self-industry and desired industries are 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, text data is already available. , so it can be combined as it is.

In addition, the industry selected from among the industries prepared in advance by the matching service provider (whether specified by symbols or numbers or described in text does not matter) and the matching requester in a free description format. If the described industry is used in combination (when there is no corresponding industry among the industries prepared in advance, and the applicable industry is described in a free description format), free description Only the industries described in the format may be included in self data and needs data. In this case, only self-industry and desired industries described in free-description form are considered to be information equivalent to self-PR (free-description form) and detailed needs (free-description form).

Although not shown in FIGS. 5 and 6, if there is a column for "others" in a free description format, the text data described in the column for "others" can be used as a self-PR (free description). 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 specific text data preparation means 34 deletes, for example, ☆, <br>, etc. from the combined text data obtained in the process of [α].

[γ] In the process of breaking down into words by morphological analysis and extracting only nouns, as shown in FIG. morphological analysis is performed on the combined text data, and the combined text data is broken down into words (segmented) and spaced. 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 word dictionary storage means and synonym dictionary storage means (not shown). In the word dictionary, words that are not included in existing analysis tool dictionaries, such as "earphones", "instagram", and "organic", are registered. Also, in the synonym dictionary, words with varying notations such as "battery" and "battery" are registered.

[δ] In the process of removing unnecessary words, the unique text data preparation means 34 narrows down the words. First, the words are filtered based on their part of speech and number of occurrences. Specifically, only a part of nouns (eg, "general", "sa-hen connection", etc.) is retained, and other words are discarded. In addition, in the set of combined text data of all projects, words appearing less than three times, for example, are discarded. The relationship between each word and the number of appearances is stored in word appearance number storage means (not shown). Therefore, when the combined text data of a new project increases, the number of occurrences of the words contained therein is added, so that the threshold value of, for example, three times may be exceeded, thereby making use of the discarded words. In some cases, it becomes possible to In addition, when the combined text data of a new case increases, a completely new word appears in the combined text data, and if the number of appearances exceeds a threshold value of, for example, three times, the new word is It is not a word that is thrown away, but a word that is used.

Next, the unique text data preparation means 34 eliminates unnecessary words (noise words) stored in the unnecessary word dictionary storage means 55 . Specifically, for example, words such as "company", "needs", "support", etc., which are considered to appear regardless of the type of business, are eliminated as unnecessary words. For example, domestic place names such as "Kyoto", "Kanto", and "East Japan" are excluded as unnecessary words. However, overseas place names such as "Europe", "China", "Dalian", etc. are left. The words registered as unnecessary words are, for example, 1,000 to 2,000 words.

In the example of FIG. 6, the words remaining after the processing [α] to [δ] by the unique text data preparation means 34 are, in order from the top, "restaurant", "store", "alcohol", "brand ”, “Food and drink”, “Food”, “Alcohol”, “Food”, “Brand”, “Awamori”, and so on. Therefore, as shown in FIG. 6, the relationship between each remaining word and the number of occurrences thereof is obtained, and this relationship provides information necessary for 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 the case number, the case identification information, and the number of the combined text data (document data). The item number does not have to be a consecutive number (although it may be a discontinuous number), but here, for the sake of convenience of explanation, including the explanation of topic estimation that will be described later, the number will be narrowed down.

(Configuration of Processing Means 30/Unique Text Data Preparing Means 34: Processing When Self Data and Needs Data Are Used as Separate Unique Text Data without Combining, FIG. 10)
In the similar case search processing of this embodiment, as shown in FIG. 10, self data and needs data are used as separate unique text data without being combined. This is almost the same as the case of creating combined text data after processing described above, except that [α] is not processed. At this time, since the self data and the needs data are separate unique text data, the processes [β] to [δ] are executed for each. Therefore, since the unique text data after processing exists separately, each processed unique text data is stored in the case data storage means 54 in association with the case identification information (case number). In other words, the processed combined text data, the processed self data, and the processed needs data are stored in the project data storage means 54 as specific text data with different contents for each project.

In the process of creating combined text data, as shown in FIG. 5, the specific text data preparation means 34 combines the self-industry and self-promotion as self-data, and combines the desired industry and detailed needs as needs data. However, as shown in FIG. 10, the unique text data for searching for similar items includes only self-promotion as self-data and only needs details as needs data. However, with regard to the unique text data for searching for similar projects, it is also possible to combine self-industry and self-promotion to make self-data, and combine desired industries and detailed needs to make needs data. Therefore, it can be said that there is processing [α].

(Configuration of processing means 30/topic estimation means 35)
The topic estimation means 35 performs Gibbs sampling using soft clustering or a neural language model using the unique text data of all the projects (companies) created by the unique text data preparation means 34 and stored in the project data storage means 54. etc., the topic value (vertical vector π(i)) indicating the appearance probability of each topic in the unique text data (i) of each project (each company) and the appearance probability of each word in each topic (matrix β) is executed, and π(i) and β obtained by this topic estimation processing are stored in the topic model storage means 56 as a topic model.

At this time, in this embodiment, the case data storage means 54 stores combined text data, which is common text data for matching and associative search, and self data and needs data, which are specific text data for similar case search. Since it is stored, topic estimation processing is executed separately for each of these three types of unique text data, and π(i) and β obtained by each topic estimation processing are used as topic models for each topic model. It is stored in the storage means 56 .

The processing by this topic estimation means 35 is executed as preprocessing (see FIG. 14). It is also executed in topic model update processing (see FIG. 16). At the time of the latter update, the unique text data of the new item (not limited to one item, but may be registered with a plurality of new items since the previous update) is stored in the item data storage means 54. Therefore, the topic estimation process is executed using the unique text data of all issues including the unique text data of the new issue.

For the soft clustering or neural language model when performing the topic estimation process, this embodiment adopts Latent Dirichlet Allocation (LDA) as an example. The "model" of the neural language model here is a concept including algorithms and parameters. On the other hand, the "model" of the topic model stored in the topic model storage means 56 is the parameter ( learning result data). Therefore, an estimator is composed of topic estimation means 35 implemented by an algorithm, topic value prediction means 36 described later, and topic model storage means 56 for storing topic models.

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), etc. can be adopted. For example, when implementing Doc2Vec, an existing library called Gensim can be used.

Here, as shown in FIGS. 18 and 19, the topic value indicating the appearance probability of each topic in the unique text data (i) of each item is a K-dimensional vertical vector π(i). is called the topic distribution. i is the document number, and when i=1 to n, n is the number of documents. In the business matching of the 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 topic values π(i, 1), π(i, 2), π(i, 3), . For example, in the business matching of this embodiment, 20 industries can be assumed as topics, and K=20. However, the number of topics K is not limited to 20, and the topic does not necessarily have to be assumed to be a type of industry. We are simply designing the system based on the assumption of the type of industry that can be used.

18 and 19, the appearance probability of each word in each topic can be represented by a matrix β of K rows and p columns. K is the number of topics. p is the number of words, eg, p=about 3,000 words.

As shown in FIG. 6, the information directly obtained from the unique text data (combined text data in the example of FIG. 6) (i=00001234) handled as one document data is each word contained in the unique text data. and the number of appearances thereof. Then, if the number of appearances of each word is known, by dividing the number of appearances of each word by the sum of the number of appearances of all words, the probability of appearance of each word in one specific text data can be obtained. Known information. The topic estimating means 35 performs Gibbs sampling, etc., using data indicating the relationship between each word and the number of occurrences of each word in each of a large number of unique text data as known information. Topic distribution π(i) in each specific text data (i) (topic value indicating the probability of appearance of each topic with topic number = 1 to K), and probability of appearance of each word in each topic (matrix of K rows and p columns β). Note that when a matrix β of K rows and p columns and a topic distribution π(i) represented by a K-dimensional vertical vector are obtained, each word (1st to p-th word ), the transposed matrix (p rows and K columns) of the matrix β and the topic distribution π(i) (K-dimensional column vector) are multiplied in this order. The above is the same when the specific text data is not the combined text data but the uncombined self data or needs data.

(Configuration of processing means 30/topic value prediction means 36)
The topic value prediction means 36 receives an instruction from the new matter registration means 45 and receives an instruction from the new matter registration means 45 to create a new matter (new company ) and the probability of appearance of each word in each topic (matrix β of K rows and p columns) (matrix β of the topic model before updating) in each topic stored in the topic model storage means 56. Then, predict the topic value (topic distribution represented by the K-dimensional vertical vector π(i)) that indicates the appearance probability of each topic in the unique text data of the new project, and predict the predicted π(i) before updating It executes a process of storing the temporary value in the topic model storage means 56 .

At this time, the topic value prediction means 36 predicts the following three types of topic distributions π(i) for the unique text data (i=n+1) after processing of the new project (new company). That is, the topic distribution π(i) in the combined text data, which is the unique text data used in the matching process, and the topic distribution π(i) in each of the self data and the needs data, which are the unique text data used in the similar case search process Predict.

When performing the processing by the topic value prediction unit 36, as shown in FIG. 6, the relationship between each word in the unique text data of the new project (combined text data in the example of FIG. 6) and the number of appearances thereof ( That is, the appearance probability of each word in the unique text data of a new project treated as one document data) and the appearance probability of each word (1st to p-th words) in each topic (topic number = 1 to K) (K The matrix β) of rows and p columns is known information. Therefore, the topic value prediction means 36, from these known information, generates a topic value (topic distribution represented by a K-dimensional vertical vector π(i)) that indicates the appearance probability of each topic in the unique text data of the new item. Predict.

In addition, in the unique text data of the new project, new occurrences that are 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 there is such a word, it is not possible to make a prediction including the newly appearing word. Therefore, assuming that the newly appearing word does not exist, the topic distribution in the unique text data of the new case is predicted. Therefore, the information of this newly appearing word is reflected in the topic model stored in the topic model storage means 56 in the subsequent topic estimation processing (topic model update processing) by the topic estimation means 35. .

(Configuration of Processing Means 30/Word Importance Vector Calculation Means 37)
The word importance vector calculation means 37 calculates the TFIDF value of each word in the unique text data of each project (each company) created by the unique text data preparation means 34 and stored in the project data storage means 54 or other words. A word importance vector composed of importance index values is calculated, and the obtained word importance vector is stored in the word importance vector storage means 57 in association with the case identification information (case number). . The peculiar text data used for calculating the word importance vector is a set of words remaining after the peculiar text data preparation means 34 has processed up to [δ].

Here, the TFIDF value is used as an example of the word importance index value in this embodiment, but it is not limited to this, and may be Okapi-BM25 or the like. Note that TF (Term Frequency) is the frequency of occurrence of words in a document (here, unique text data). is the value divided by the sum of IDF (Inverse Document Frequency) is the inverse document frequency, and the logarithm (log) is taken for the value obtained by dividing the total number of documents (here, the total number of unique text data) by the number of unique text data containing each word. value.

A word importance vector is a p-dimensional vector, where p is the number of words. This word importance vector is preferably subjected to normalization (for example, L2 norm normalization).

In this embodiment, the item data storage means 54 stores combined text data, which is unique text data used in the matching process, and self data and needs data, which are unique text data used in the similar item search process. , a word importance vector for each of these three types of unique text data is calculated, and each word importance vector is stored in the word importance vector storage means 57 in association with the case identification information (case number). In addition, if the unique text data used in the matching process and the similar case search process is unified with the combined text data, or if the self data and the needs data that are not combined are unified, the matching process and the similar case search process The word importance vector for can be shared.

The processing by this word importance vector calculating means 37 is executed in the stage of pre-processing (see FIG. 14). Also, at the time of registering a new matter (see FIG. 23), in response to an instruction from the new matter registration means 45, as a process for immediate reflection of the new matter before updating the topic model, etc., a word for the unique text data of the new matter Importance vector calculation processing is executed. Furthermore, it is also executed when the time for updating the topic model comes (see FIG. 16), but it is not necessary to repeatedly execute the processing that is executed in the instant reflection processing of the new item.

(Configuration of Processing Means 30/Matching Learning Data Preparing Means 38)
The matching learning data preparation means 38 selects or does not select each of the large number of matching feature data (see FIG. 8) as matching learning data (matching model learning data to be stored in the matching model storage means 60). Annotated data (see Fig. 9) with labels (tags) of is created, and a large number of tagged feature data for matching are associated with a pair of project identification information (project number) to store data for matching learning. It executes processing to be stored in means 59 .

At this time, the matching learning data preparation means 38 prepares a large number of matching feature data. A detailed description will be given with reference to FIG. The selection/non-selection tagging process for each feature data for matching is executed using the matching candidate selection result list stored in the matching candidate selection result list storage means 58 .

The processing by the matching learning data preparation means 38 is executed at the stage of preprocessing (see FIG. 14) for constructing the initial model of the matching model. It is also executed when the matching model is updated (see FIG. 16).

(Configuration of Processing Means 30/Matching Model Learning Means 39)
The matching model learning means 39 prepares the matching learning data created by the matching learning data preparation means 38 and stored in the matching learning data storage means 59 (selected/non-selected data for each of the large number of matching feature data). Annotated data with labels (tags, see FIG. 9) is used to perform learning processing by a supervised classification model, and a matching model obtained (here, it means a parameter as learning result data.) is stored in the matching model storage means 60.

The processing by the matching model learning means 39 is executed at the stage of preprocessing (see FIG. 14) for constructing the initial model of the matching model. It is also executed when the matching model is updated (see FIG. 16).

Strictly speaking, if the number of unique text data is increased and the topic estimation processing by the topic estimation means 35 is executed again, the topic distributions of the respective unique text data will have slightly different values. Each piece of matching feature data that is created by using them is also in a slightly different state. If the matching model learning means 39 performs learning processing using such slightly different matching feature data, the matching model stored in the matching model storage means 60 will be in a slightly different state. However, if the number of specific text data to be added is one or a relatively small number, the topic distribution hardly changes, so the matching models stored in the matching model storage means 60 are used as they are without re-learning. be able to. Therefore, updating the matching model does not necessarily have to be performed at the same time as updating the topic model by re-executing the topic estimation process, and updating the topic model and updating the matching model may be performed at different timings. (See step S14 in FIG. 16).

(Configuration of Processing Means 30/Input Receiving Means 40)
The input reception unit 40 executes processing for receiving input of keywords or project identification information (project number) in the input unit 110 provided on each screen 100, 200, 300 (see FIGS. 11, 12, and 13). It is.

Here, for inputting a keyword to the input unit 110, in addition to inputting a keystroke operation to the input unit 110 of each of the screens 100, 200, and 300, a keyword displayed in the "similar keyword" display unit 210 of the screen 200 in FIG. Selective input (input by click operation, tap operation, etc.) by selecting each word (input keyword and a plurality of related words related to the keyword) is included. When one of the words is selected and input, the selected word is automatically displayed on the input section 110 . In addition, for inputting project identification information (project number) to the input unit 110, in addition to input by keystroke operation to the input unit 110 of each of the screens 100, 200, and 300, the "keyword search list" on the screen 200 of FIG. Selection input by selecting project identification information (project number) of each related project displayed on the display unit 220, and the "recommended project list" display unit 320 and the "similar project list" display unit 330 of the screen 300 in FIG. selection input by selecting the item identification information (item number) of each recommended item and each similar item displayed in . When one of the item identification information (item number) of each related item, each recommended item, and each similar item is selected and input, the selected item identification information (item number) is automatically displayed in the input section 110. there is

The input reception unit 40 determines whether the information input to the input unit 110 is a keyword or item identification information (item number). to execute the associative search process, and if it is judged to be the case identification information (case number), the case identification information is transmitted to the matching means 42 and the similar case search means 43 to perform the matching process and the similar case identification information. Execute the matter search process. In this embodiment, the input of keywords and the input of project identification information (project number) are accepted by the same input unit 110. However, the keyword input unit and the project identification information (project number) ) may be provided separately. However, from the viewpoint of improving the operability and convenience of the user, it is preferable that the same input unit 110 is configured to receive the data.

Further, the input receiving unit 40 causes the search history storage unit 61 to store the log information of the received keyword or item identification information (item number). The search history storage means 61 may be a volatile memory such as a main memory. At this time, if the matching system 10 has a configuration (for example, a network configuration) that can simultaneously accept operations from a plurality of users (service personnel or matching requesters), log information of the search history of each user , is stored in the search history storage means 61 in association with the user identification information. Then, using the log information stored in the search history storage unit 61, the input reception unit 40 uses the "search history" display unit 130 provided on each screen 100, 200, 300 (in FIG. (illustration is omitted)), the search history is displayed (display in which keywords or project identification information are arranged in order of input). The same keyword or the same item identification information (item number) may be entered continuously, but the same output will only be repeated. will be lined up.

(Configuration of processing means 30/associative search means 41)
The associative search means 41 includes word similarity calculation means 41A and relevance calculation means 41B.

When an input of a keyword is received by the input receiving means 40, the word similarity calculating means 41A calculates the difference between the keyword calculated using the distributed representation of the words stored in the distributed representation model storing means 53 and other words. and the unique similarity B between the keyword and other words calculated using the unique text data of each project (each company) stored in the project data storage means 54 A process of calculating the degree of similarity FS and determining words related to the keyword in descending order of the obtained final degree of similarity FS is executed. More detailed contents will be described later.

Here, the unique text data of each case when calculating the unique similarity B is unique text data for associative search, and in this embodiment, combined text data that is common unique text data for matching. .

Note that the word similarity calculation means 41A may use the general similarity A as the final similarity FS instead of using the unique similarity B to determine the related words of the keyword. However, from the viewpoint of performing more effective matching, it is preferable to calculate the final similarity FS using both the general similarity A and the specific similarity B. FIG.

The degree-of-relevance calculation means 41B calculates the appearance of each of the keywords received by the input receiving means 40 and their related words in the unique text data of each project (each company) stored in the project data storage means 54. The number of occurrences C is obtained, and using the obtained word-by-word appearance frequency C and the final similarity FS of the related words, the degree of relevance CS of each case with respect to the keyword is calculated, and the related cases are determined in descending order of the degree of relevance CS. to execute. More detailed contents will be described later.

(Configuration of processing means 30/associative search means 41/word similarity calculation means 41A: calculation processing of general similarity A, FIG. 2)
The word similarity calculation means 41A stores a distributed representation model (in this embodiment, a weight matrix W consisting of a set of word vectors) stored in the distributed representation model storage means 53 (see FIG. 2). W1, W2, and W3) are used to calculate the model-specific similarity S between the keyword received by the input receiving means 40 and other words, and ranking is performed according to the magnitude of this model-specific similarity S. Among the words obtained, a predetermined top N3 (for example, N3=100) words are selected as similar words by model. Although N3 is stored in the setting data storage means 51 as setting data, it may be a fixed value.

More specifically, the word similarity calculation means 41A first stores distributed representation models of types 1 to 3 in the distributed representation model storage means 53 (see FIG. 2), as shown in FIG. Using the weight matrix W (W1 to W3), the word vector V (k, j) for the keyword received by the input receiving means 40 and the word vector V (i, j) for other words A cosine similarity S(i, j) is calculated and stored in the model similarity storage means 62 . The model similarity storage means 62 may be a volatile memory such as a main memory. As shown in FIG. 2, the cosine similarity S(i,j) is the inner product of V(k,j) and V(i,j), the magnitude of V(k,j) and V(i,j). ) is the value divided by the product of the magnitude of This cosine similarity S(i, j) calculation process may be executed using the function of an existing tool such as a word2vec module.

Here, k is the number of the keyword (word) received by the input receiving means 40 . i is the number of each word, i=1, 2, 3, . . . , k−1, k+1, . 1.15 million). The omission of k means that the keyword words are excluded from the N2 words, and the cosine similarity S(i, j) It means to calculate j is the type number of the distributed representation model, and j=1, 2, 3 in this embodiment. Note that the number of distributed representation models is not limited to three, and may be a plurality other than three, and is not limited to a plurality, and may be one. However, if a plurality of distributed representation models are used, it is possible to absorb variations in model-specific similarity S among the distributed representation models and determine more appropriate words as general similar words. When the number of distributed representation models is one, the model-specific similarity S in the distributed representation model becomes the general similarity A, and the top N3 of the model-specific similarities S in the distributed representation model (for example, N3=100 ) is a model-specific similar word and a general similar word.

Therefore, as shown in FIG. 2, for example, for the type 1 distributed representation model, the weight matrix W1 is used to determine the word vector V(k,1) for the keyword word and the word vector V(k,1) for the other words Calculate the cosine similarity S(i,1) with V(i,1). Similarly, for the type 2 distributed representation model, using the weight matrix W2, the word vector V(k, 2) for the keyword word and the word vector V(i, 2) for the other words are A cosine similarity S(i, 2) between is calculated. The same applies to the type 3 distributed representation model.

Then, the word similarity calculation means 41A selects the magnitude Select the cosine similarities S(i,j) within the top N3 (eg, N3=100). , k-1, k+1, . . . , N2-2, N2-1, N2 and j=1, 2, 3. The number of number i is (N2-1) because one word of the keyword number k is excluded from N2 (eg, 1,150,000) words.

For example, for a type 1 model, (N2−1) (eg, (1,150,000−1)) cosine similarities S(i, 1) (i=1, 2, 3, . . . , k−1 , k+1, . The selected cosine similarities S(i,1) are outlined in bold in the table at the bottom of FIG. Similarly, for the type 2 model, (N2−1) (for example, (1,150,000−1)) cosine similarities S(i, 2) (i=1, 2, 3, . . . , k− 1, k+1, . . . , N2−2, N2−1, N2), the cosine similarity S(i,2) within the top N3 ranks (for example, within the 100th rank) is selected. Furthermore, for the type 3 model, (N2−1) (for example, (1,150,000−1)) cosine similarities S(i,3) (i=1, 2, 3, . . . , k−1 , k+1, .

The word number i of S(i,j) within the top N3 (for example, within the 100th) selected by each distributed representation model of types 1 to 3 usually does not match. In other words, as shown in the table at the bottom of FIG. 2, the portion surrounded by the thick frame is shifted. Therefore, when considering the three types as a whole, the number of words selected by at least one distributed representation model, including the words for the deviation when the words selected by each distributed representation model do not match, is N3. (for example, 100) will be larger. Let this number be N4 (eg, 150).

In addition, since N3 (eg, 100) word numbers are shifted and selected, a total of N4 (eg, 150) word numbers are selected, so distributed representation models of types 1 to 3 , the cosine similarity S(i,j) corresponding to (N4-N3) (eg, 150-100=50) word numbers is not selected, so it does not exist. . In this way, for words with no corresponding cosine similarity S(i,j), the cosine similarity S(i,j)=0 is assumed. That is, in reality, the values of cosine similarity S(i, j) exist, but are so small that they do not rank within the top N3 ranks (for example, within the 100th rank), so they are regarded as 0. Become.

For example, in the table at the bottom of FIG. 2, for the type 1 distributed representation model, out of N4 (eg, 150) word numbers i, number i = k-1 S (k-1, 1) does not exist, so we consider S(k-1,1)=0. Similarly, for the distributed representation model of type 2, among N4 (for example, 150) word numbers i, S(2, 2) with number i=2 and S(N −2, 2) does not exist, so it is assumed that S(2, 2)=S(N−2, 2)=0. In addition, for the type 3 distributed representation model, among N4 (eg, 150) word numbers i, S (k-1, 3) with number i = k-1 and S (k-1, 3) with number i = N-2 Since S(N-2,3) does not exist, it is assumed that S(k-1,3)=S(N-2,3)=0.

After that, the word similarity calculation means 41A calculates the cosine value stored in the model-by-model similarity storage means 62 for each of N4 (for example, 150) words with the number i for each of the distributed representation models of types 1 to 3. By obtaining the harmonic average or weighted harmonic average (weighted harmonic average) of similarity S(i,j), general similarity A(i) is calculated by ensemble of three types of distributed representation models. That is, the general similarity A(i) is the harmonic average or weighted harmonic average of S(i,1), S(i,2), and S(i,3). Weighted harmonic mean, for example, when there is a distributed representation model that you want to emphasize regarding dimension or window size, gives a relatively large weight to it, or combines word2vec's CBOW distributed representation model and Skip-gram's distributed representation model. In the case of a mixture of CBOW and Skip-gram, the weight can be changed.

Taking the harmonic average or weighted harmonic average, for the word number i, the cosine similarities S(i,1), S(i,2), S(i,3) of each distributed representation model of types 1 to 3 are When one of the values (at least one value) is zero, general similarity A(i)=0. For example, in the table at the bottom of FIG. 2, since S(2,2)=0 with number i=2 for the distributed representation model of type 2, the general similarity A(2) of the word with number i=2 = 0. Similarly, since S(k-1, 1) of number i = k-1 for distributed representation models of types 1 and 3 = S(k-1, 3) = 0, The word general similarity A(k−1)=0. Also, since S(N-2,2)=S(N-2,3)=0 with number i=N-2 for distributed representation models of types 2 and 3, the word with number i=N-2 , the general similarity A(N−2)=0.

Subsequently, the word similarity calculation means 41A selects words for which the value of the ensembled general similarity A(i) is not zero, determines them as general similar words, and The general similarity A(i) for is stored in the general similarity storage means 63 . The general similarity storage means 63 may be a volatile memory such as a main memory. In addition, words selected as general similar words without the value of general similarity A(i) obtained by ensemble of multiple types of distributed representation models being zero are selected by each of types 1 to 3 distributed representation models. Of the numbers i of the model-specific similarities S(i, j) within the top N3 ranks (for example, within the 100th rank), the number selected in all distributed representation models, that is, the number overlapping in all distributed representation models. is a word. In other words, it is a word that is a model-specific similar word for all distributed representation models. Assume that as a result, the number of words determined as general similar words is N5 (for example, 70). The value of N5 is naturally less than N3 (for example, 100), and N5=N3 is based on the top N3 ranking (for example, within 100th ranking) for all distributed representation models of types 1 to 3. It is only when the combination of words with the degree of similarity S(i, j) matches (the order need not match).

As described above, in the present embodiment, the general similarity A(i) is the harmonic average or weighted harmonic average of S(i,1), S(i,2), and S(i,3). The method of integrating the model-specific similarities S for calculating the general similarity A is not limited to harmonic averaging or weighted harmonic averaging, and may be, for example, simple averaging, weighted averaging, geometric averaging, or the like. In addition, before taking various averages, from the viewpoint of excluding outstanding values, the maximum and/or minimum model similarity S among the model similarities S of the word number i of each distributed representation model is excluded. After that, various averages may be taken.

(Configuration of processing means 30/associative search means 41/word similarity calculation means 41A: processing for calculating inherent similarity B, FIG. 3)
Further, the word similarity calculation means 41A calculates the N5 (for example, 70) general similar words and the keyword words determined as described above for each case stored in the case text data storage means 54. By matching with unique text data (in this embodiment, combined text data of self data and needs data), unique text data containing each general similar word and/or keyword word and its project identification information (project identification information) number). The extracted item identification information (item number) is stored in an unillustrated extracted item storage means (a volatile memory such as a main memory may be used).

Then, as shown in FIG. 3, the word similarity calculation means 41A performs the following processing for all N5 (for example, 70) general similar words. In FIG. 3, any one general similar word out of the N5 general similar words is β.

That is, the word similarity calculation means 41A calculates the number of occurrences of the keyword α and the general similar word β in the set P of all projects that include both the keyword α and the general similar word β in the unique text data. The evaluation value Ps of the set P based on the number of occurrences of (However, the coefficient σ in Equation (1) described later is set to 0, and the evaluation value of the set P based only on the number of occurrences of the general similar word β without using the number of occurrences of the keyword α Ps.), and an evaluation value Qs of the set Q based on the number of occurrences of the general similar word β in the set Q of each case in which the unique text data does not include the keyword α and includes the general similar word β. is used to calculate the inherent similarity B from Ps/(Ps+Qs), and the calculated inherent similarity B is stored in the inherent similarity storage means 64 . The unique similarity storage means 64 may be a volatile memory such as a main memory.

In FIG. 3, for example, the projects containing the keyword α (for example, “alcohol”) are companies A, B, C, H, and J, and the projects containing the general similar word β (for example, “shochu”). are A company, B company, C company, D company, E company, F company, and G company. In this case, a set P of cases including both the keyword α and the general similar word β is composed of A company, B company, and C company. Also, a set Q of projects that do not contain the keyword α but contain the general similar word β are composed of D company, E company, F company, and G company. Company H and company J are not involved in the calculation of the unique similarity B for the general similar word β with respect to the keyword α.

The evaluation value Ps of the set P is calculated by the following formula (1) using a function f of the number of appearances of α and the number of appearances of β for each project (company) belonging to the set P.

Ps=f (number of appearances of α and number of appearances of β for each case belonging to set P)
=(Aα+Bα+Cα)×σ+(Aβ+Bβ+Cβ)×τ (1)

Here, Aα is the number of appearances of the keyword α in company A's unique text data. Similarly, Bα and Cα are the numbers of appearances of the keyword α in the unique text data of B company and C company, respectively. Aβ, Bβ, and Cβ are the numbers of occurrences of the general similar word β in each of the unique text data of company A, company B, and company C, respectively. FIG. 3 is an example, so if more projects belong to the set P, the number of appearances of the keyword α and the general similar word β in the specific text data of all those projects will be added. .

Also, the coefficient σ can take a value of σ=0. In that case, the unique text data of each case belonging to the set P includes both the keyword α and the general similar word β. The number of appearances of the keyword α is not used. Although the coefficients σ and τ are stored as setting data in the setting data storage means 51, they may be fixed values.

The evaluation value Qs of the set Q is calculated by the function g of the number of occurrences of β for each project (company) belonging to the set Q, using the following equation (2).

Qs=g (number of appearances of β for each project (company) belonging to set Q)
= (Dβ+Eβ+Fβ+Gβ)×φ (2)

Here, Dβ, Eβ, Fβ, and Gβ are the numbers of appearances of the general similar word β in the specific text data of D company, E company, F company, and G company, respectively. Since FIG. 3 is an example, if more projects belong to the set Q, the number of occurrences of the general similar word β in the unique text data of all those projects is added. Although the coefficient φ is stored in the setting data storage means 51 as setting data, it may be a fixed value.

For example, in data aggregation example 1 in the table of FIG. ) is set to φ=1, Qs=(2+1+2+4)×1=9. Therefore, the intrinsic similarity B is Ps/(Ps+Qs)=13/(13+9)=0.591.

In data aggregation example 1 in the table of FIG. When the coefficient φ of 2) is set to 1, Ps=6 and Qs=9, so the inherent similarity B is Ps/(Ps+Qs)=6/(6+9)=0.4.

In addition, in the data aggregation example 1 in the table of FIG. 3, when calculating the evaluation value Ps of the set P, the expression (1 ) of σ = τ = 0.5 and the coefficient φ of Equation (2) = 1, Ps = 6.5 and Qs = 9, so the inherent similarity B is Ps / (Ps + Qs )=6.5/(6.5+9)=0.419.

Furthermore, if the value of the number of appearances of a and β in the unique text data of a certain project is outstanding, there is a possibility that the impact of that one project will become excessive. Therefore, for example, a threshold value (for example, 10 times) may be set, and when the number of appearances exceeds the threshold value, upper limit adjustment may be performed such as changing to the threshold value. For example, if Aβ=13 when the threshold is 10 times, the upper limit is adjusted to Aβ=10. Also, conversion adjustment such as obtaining the square root of the number of appearances may be performed. For example, if Aβ=2, it is converted and adjusted to Aβ=√2=1.414.

Therefore, the unique similarity B should be calculated using the unique text data of each project (each company). Therefore, the inherent similarity B (referred to as the text data reference relevance Pword (relevance rate) in Patent Document 2) is obtained by the method described in the extraction system of Patent Document 2 developed by the applicant of the present application. ) may be calculated. In the extraction system of Patent Literature 2 described above, the unique similarity B is obtained by the number of cases (companies) belonging to the set P and The unique similarity B is calculated using the number of projects (companies) to which the project belongs. Therefore, according to the method of the extraction system of Patent Document 2 described above, since the unique similarity B is the ratio of the cases including the keyword α among the cases including the general similar word β, the data tabulated in the table of FIG. In Example 1, the unique similarity B is 3 companies/(3 companies+4 companies)=3 companies/7 companies=0.429.

In addition, in the situation of data aggregation example 2 in the table of FIG. In the situation of aggregation example 3, if the coefficients σ = 0 and τ = 1 in formula (1) and the coefficient φ = 1 in formula (2) are set, Ps = 6 and Qs = 8, so there is inherent similarity The degree B is Ps/(Ps+Qs)=6/(6+8)=0.429, which is the same value as the unique similarity B calculated by the method of the extraction system of Patent Document 2 described above. Therefore, when the number of occurrences of the words α and β becomes a special situation (equal in each case), the calculated value of the inherent similarity B by the method of Patent Document 2 considering the number of cases and the number of appearances are considered. The calculated value of the unique similarity B by the method of this embodiment matches.

(Configuration of Processing Means 30/Associative Search Means 41/Word Similarity Calculation Means 41A: Final Similarity FS Calculation Processing)
After that, the word similarity calculation means 41A calculates the general similarity A(i) for each general similar word stored in the general similarity storage means 63 and each general similarity stored in the specific similarity storage means 64. By obtaining the harmonic average or weighted harmonic average (weighted harmonic average) with the unique similarity B (i) for the word, the final similarity FS (i) for each general similar word is calculated, and this final similarity Among the general similar words ranked by the magnitude of FS(i), a predetermined number of top words or words selected by the magnitude of the final similarity FS(i) with respect to the threshold are selected as related words , and stores the determined related words and the final similarity FS(i) of the related words in the final similarity storage means 65 . In this embodiment, top N6 (for example, 10) general similar words are selected and determined as related words. Although N6 is stored in the setting data storage means 51 as setting data, it may be a fixed value.

Here, i is the number of words determined as general similar words, i=1, 2, 3, . Therefore, here the number i is packed.

When calculating the final similarity FS(i) using the general similarity A(i) and the specific similarity B(i), simple average, weighted average, A geometric mean, a root mean square (RMS), or the like may be obtained, but a harmonic mean or a weighted harmonic mean is preferable from the viewpoint of averaging the ratios.

(Configuration of Processing Means 30/Associative Search Means 41/Relevance Degree Calculation Means 41B: Relevance Degree CS Calculation Process, FIG. 4)
The degree-of-relevance calculation means 41B selects each related word (in this embodiment, general similar words Among β, the following processing is performed for each specific text data containing either words with the highest final similarity FS (N6 (eg, 10) words) or keyword α. Since each unique text data to be processed is naturally included in a set of unique text data containing either each general similar word β or keyword α, each unique text data to be processed here The case identification information (case number) of the unique text data is within the range of the case identification information (case number) extracted by the word similarity calculation means 41A and stored in the extracted case storage means (not shown).

That is, as shown in FIG. 4, the degree-of-association calculation means 41B calculates the final degree of similarity FS for the top N6 ( For example, 10 words) and each word of the keyword α, the number of occurrences C(i) for each word appearing in each unique text data to be processed as described above is obtained, and the following equation (3) is obtained. A value obtained by multiplying the word-specific appearance frequency C(i) for each related word γ (γ1, γ2, ...) by the final similarity FS(i) for each related word γ, or a weighting factor for each related word γ to this value sum the value multiplied by κ(i) and the number of occurrences per word C(0) for the keyword α or the value obtained by multiplying this number of occurrences per word C(0) by the weighting factor κ(0) of the keyword α Thus, the degree of relevance CS of each project (each company) with respect to the keyword α is calculated, and the calculated degree of relevance CS is associated with the project identification information (project number) and stored in the degree of relevance storage means 66 .

CS=Σi{κ(i)×FS(i)×C(i)} (3)

Here, i is the number of the word, i=0, 1, 2, 3, . is the number of words determined as related words γ, and N6 (eg 10) is the number of related words γ. Therefore, here the number i is packed. The final similarity FS for the keyword α(i=0) is FS(0)=1.

κ(i) is a weighting factor for each word (keyword α, each related word γ), and both are stored in the setting data storage means 51 as setting data. In principle, κ(i)=1 may be set for the weighting factors of all words (i=0, 1, . . . , N6). This is because the word appearance count C(i) is multiplied by FS(i) as a weight. Therefore, κ(i), which is a value other than 1, can be used when further weighting is performed with respect to the final similarity FS(i). For example, the largest FS(i) or largest C(i) is multiplied by κ(i)=3, and the second largest FS(i) or second largest C(i) is multiplied by κ Further weighting can be done, such as multiplying by (i)=2. Also, κ(i) may be used for relative weighting of the keyword α (i=0) and each related word γ (i=1 to N6), in which case κ(0) and other κ(i) (i=1 to N6) may be set to different values. At this time, the final similarity for the keyword α (i=0) is set to FS(0)=1 and further multiplied by the weighting factor κ(0), so κ(0)×FS(0) is one It may be considered a weighting factor, or FS(0) itself may be considered a weighting factor instead of being 1 without using κ(0). Note that the value of the degree of association CS obtained by Equation (3) is a numerical value for relative evaluation for ranking projects (companies). ) can be a numerical value that can be relatively weighted, so it can be a value of any level (for example, a range of 0 to 1, a range of 0 to 100, and any range It means that it can be a number.), and it doesn't have to be an integer either.

More specifically, in the example of FIG. 4, the degree-of-association calculation unit 41B uses keyword α (for example, “alcohol”) and each related word γ (for example, γ1=“Shochu”, γ2=“Awamori”, etc.) is determined. As shown in FIG. 4, in the unique text data of company A, "sake" (keyword α) appears three times, "shochu" (related word γ1) appears twice, and "awamori" (related word γ2) appears twice, the number of occurrences C(i) (i=0, 1, . . . , N6) is obtained.

Subsequently, the degree-of-association calculation means 41B acquires the final degree of similarity FS(i) (i=1 to N6) for each related word γ stored in the degree-of-similarity storage means 65, and obtains the final degree of similarity FS (i), FS (0) = 1 for the keyword α (i = 0), and the number of appearances for each word in the unique text data of company A obtained as described above C (i) (i = 0, 1, . . . , N6) and weighting coefficients κ(i) (i=0, 1, . Calculate the degree of relevance CS of the company.

Then, the degree-of-association calculating means 41B similarly executes the above processing for the unique text data of the cases to be processed other than the company A. As a result, the degree of relevance CS of each project (each company) with respect to the keyword α is obtained. Acquire the Item data of the related Item. At this time, it is also possible to select and determine the cases with the highest number of related degrees CS as the related cases, or to determine the related cases by setting a threshold value for the related degrees CS.

(Configuration of processing means 30/matching means 42)
The matching means 42 includes matching feature data creating means 42A and matching score calculating means 42B. This matching means 42 includes the matching system of the above-mentioned Patent Document 1 developed by the applicant of the present application or its modified system (instead of using the combined text data described in Patent Document 1 as unique text data, the unbound When using state self data and needs data), the present invention mainly develops the peripheral part of the matching system of US Pat. Therefore, the effectiveness of the matching process of the present invention is as described as experimental results in paragraph number [0211] of Patent Document 1, and the precision rate, recall rate, F1 value, and accuracy are all high. numerical values are obtained.

When the input receiving means 40 receives the input of the item identification information (item number), the matching characteristic data creating means 42A sets the item of the received item identification information (item number) as the item to be subjected to matching processing, and creates a topic model. Using the topic distribution (the topic value indicating the probability of appearance of each topic) in the unique text data of each case stored in the storage means 56, a plurality of candidate cases consisting of the target case and each candidate case to be matched with this target case is executed to create matching feature data for each of the pairs. More detailed contents will be described later.

The matching score calculating means 42B uses each of the plurality of matching feature data created by the matching feature data creating means 42A as input data, and is pre-learned by two-class classification as to whether or not it is selected, and stored in the matching model storage means 60. Processing to calculate a score (matching score) that indicates the probability (probability) that each candidate project will be selected for the target project by performing classification processing using a supervised classification model using the matching model stored in is executed. More detailed contents will be described later.

(Configuration of processing means 30/matching means 42/matching feature data creation means 42A: FIGS. 5, 7, and 8)
As shown in FIG. 5, the matching feature data creating means 42A generates the topic distribution π(i) (topic number=1 to K of each topic in the unique text data for matching (combined text data in this embodiment). topic value indicating the probability of occurrence), a word importance vector for matching (TFIDF vector as an example in this embodiment) for unique text data for matching, a matching target case (target company), and Using the needs flag and annual sales for each candidate project (each candidate company), a process of creating matching feature data that serves as input data to the discriminator consisting of the matching score calculation means 42B and the matching model storage means 60 is performed. Run. However, what is described in the lower part of FIG. 5 is the matching feature data of one side of the pair consisting of the matching target case and the candidate case (therefore, the state before creating the synthetic variables, the word importance state before calculating the inner product of the degree vectors).

More specifically, as shown in FIG. 8, the matching feature data creating means 42A, as a process during operation (during user search), extracts the target case X and each candidate case stored in the topic model storage means 56. Topic distribution (topic value indicating the appearance probability of each topic) in unique text data for matching of Y (combined text data in this embodiment) and word importance vector calculated by the word importance vector calculation means 37 are stored. Matching word importance vector (TFIDF vector, etc.) for matching specific text data of target case X and each candidate case Y stored in means 57, and target case stored in case data storage means 54 Using the need flags and annual sales included in the project data of X and each candidate project Y, matching feature data for a plurality of pairs each composed of the target project X and each candidate project Y is created.

Here, the topic distribution (topic value indicating the appearance probability of each topic) in the matching specific text data of the target case X and each candidate case Y stored in the topic model storage means 56 is obtained by the topic estimation means 35. In addition to the topic distribution in the unique text data for matching of each case obtained in the topic estimation process (at the time of building the initial model of the topic model and at the time of updating the model), the topic value by the topic value prediction means 36 depending on the timing of the search Topic distribution (provisional values before updating the topic model) in unique text data for matching new projects obtained by prediction may be included. "Depending on the timing of the search" means that the topic distribution by the topic value prediction means 36 after updating the topic model (including after building a new model) and at the timing when no new project is registered after the update. (temporary value) does not exist, and on the other hand, when there is a registration of a new project after the update of the topic model and the timing of the next update has not arrived, the topic distribution (temporary value) by the topic value prediction means 36 (temporary value) exists.

Also, in FIG. 5, the matching feature data for each case identification information (each case number) (however, FIG. 5 shows the state before creating the composite variable, the state before calculating the inner product of the word importance vector) is , topic distribution, word importance vector (for example, TFIDF vector), needs flag, and annual sales. , annual sales are not indispensable for creating matching feature data, and it is possible to choose not to use them. Also, although not shown in FIG. 5, the document length may be used to create the matching feature data. As the document length, for example, [α] the unique text data created by the unique text data preparation means 34, or [β] the unique text data after unnecessary symbols and tags have been removed. (that is, the number of characters in the unique text data before it is broken down into words by the morphological analysis of [γ]), or the total number of words in the unique text data after processing up to [δ] etc. can be used.

Furthermore, as shown in FIG. 7, it is effective to include synthetic variables in creating matching feature data for proper scoring.

As shown in FIG. 7(A), when creating a composite variable from each topic value, a composite variable consisting of the product of the topic distribution π(X) of the target project X and the topic distribution π(Y) of the candidate project Y A distribution of variables λ(XY) can be determined and included in the matching feature data. In FIG. 7A, the notation λ(XY)=π(X)*π(Y) is used, but the inner product of vectors is not obtained, but the product of corresponding elements of vectors is obtained. It means. That is, since it is the product of topic values for the same topic, the product of topic values of topic number = 1, the product of topic values of topic number = 2, ..., the product of topic number = K (K is the number of topics) Multiply topic values together. In more precise notation, 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, from the topic distribution .pi.(X), which is a K-dimensional vector, and the topic distribution .pi.(Y), which is a K-dimensional vector, the composite variable distribution .lambda.(XY), which is a K-dimensional vector, is obtained.

Also, as shown in FIG. 7(A), distribution μ(XY) of a composite variable consisting of the absolute value of the difference between the topic distribution π(X) of the target project X and the topic distribution π(Y) of the candidate project Y can be obtained and included in the matching feature data. The reason why the difference is not the difference but the absolute value of the difference is to eliminate the directivity, 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 vectors. 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, . Calculate the absolute value of the difference between topic values of =K (K is the number of topics). In more precise notation, 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 composite variable distribution μ(XY), 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. 7A, both the composite variable distribution λ(XY) consisting of the product of the topic distributions and the composite variable distribution μ(XY) consisting of the absolute values of the differences of the topic distributions are matched. Although it is preferable to include it in the feature data, either one of the product distribution λ(XY) and the difference absolute value distribution μ(XY) may be used.

Furthermore, as shown in FIG. 7(B), when creating a synthesized variable by each needs flag, the logical sum (OR), logical product (AND), and Exclusive logical sum (XOR), negative logical sum (NOR), negative logical product (NAND), negative exclusive logical sum (NXOR), arithmetic sum, etc. can be employed. In the example of FIG. 7(B), composite variables by logical sum (OR), logical product (AND), and arithmetic sum are shown. This example does not mean that all of multiple types (three types) of composite variables such as logical sum (OR), logical product (AND), arithmetic sum, etc. are used, but one type of composite variable is selected and used. It is described with the intention that it should be done. Multiple types of combined variables may be included in the matching feature data.

When creating a combined variable for each need flag, the combined variable may be created across different types of need flags instead of between corresponding need flags. For example, if there are needs flags that are related in terms of content, such as (1) sell, (2) buy, among the need flags, using the need flags (1) and (2), the above A composite variable may be created by a logical sum (OR) or the like.

Also, as shown in FIG. 7(C), when creating a combined variable based on annual sales, a ratio of both annual sales can be used. In the case of a composite variable M(XY) based on the value of the ratio, let the annual sales of applicant X be S(X) and the annual sales of applicant Y be S(Y). )=S(X)/S(Y), or M(XY)=S(Y)/S(X). In order to eliminate 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. If directionality is given, M(XY) and M(YX) will have different values.

Furthermore, although not shown in FIG. 7, when creating a combined variable based on the document lengths, a ratio of both document lengths can be used. In the case of a composite variable based on a ratio value, the directionality may be waited or may be eliminated, as in the case of the annual sales described above.

As for the word importance vector (such as the TFIDF vector), the inner product of the word importance vector of the target case X and the word importance vector of the candidate case Y is obtained. Since it is an inner product, it has no directionality.

In this embodiment, the unique text data for matching is combined text data, but the uncombined self data and needs data may be used as unique text data. There are topic distribution π1(X) about data and topic distribution π2(X) about needs data. ).

At this time, in order to obtain a composite variable consisting of the product of topic distributions, the self-data of the target project X and the needs data of the candidate project Y are combined to give λ1(XY)=π1(X)*π2(Y). In addition, it is possible to combine the self-data of candidate project Y and the needs data of target project X to obtain λ2(XY)=π1(Y)*π2(X). As described above, "*" means the product of corresponding elements of vectors, not the inner product of vectors. Similarly, composite variables consisting of the absolute values of differences in topic distributions are μ1(XY)=|π1(X)−π2(Y)|, μ2(XY)=|π1(Y)−π2(X)| can do. Therefore, the four composite variables λ1(XY), λ2(XY), μ1(XY), μ2(XY) can be included in the matching feature data. Note that when no composite variable is used, the four topic distributions π1(X), π2(X), π1(Y), and π2(Y) are directly included in the matching feature data.

Also, by using the obtained two synthesized variables λ1(XY) and λ2(XY), two-step synthesis is performed to find one synthesized variable λave(XY) consisting of the average value of the corresponding elements of these vectors. you can go Similarly, using two synthesized variables μ1(XY) and μ2(XY), two-step synthesis is performed to find one synthesized variable μave(XY) consisting of the average values of corresponding elements of these vectors. may When two-step synthesis is performed in this manner, the two synthesis variables λave(XY) and μave(XY) are included in the matching feature data.

Further, the matching feature data creating means 42A determines whether π1(X) and π2(Y) are included or their composite variables λ1(XY) and μ1 First matching feature data including (XY) and second matching feature data including π1(Y) and π2(X) or their composite variables λ2(XY) and μ2(XY) , and these first and second matching feature data are separately input to a discriminator consisting of matching score calculation means 42B and matching model storage means 60, so that the first and second matching scores are calculated as After that, an average value or the like of these first and second matching scores may be calculated to calculate an integrated matching score.

In the case of calculating the matching score after such integration, the matching learning data preparation means 38 selects two pieces of matching learning data from one pair of the case number X and the case number Y. 1. Create second matching feature data. That is, in FIG. 9, two pieces of matching feature data tagged with "1 (selected)" are created from one pair case that has been selected, and two matching feature data are created from one pair case that has not been selected. (Non-selected)” is prepared for all pair cases, and the matching model learning means 39 executes the learning process to construct the matching model.

(Configuration of Processing Means 30/Matching Means 42/Matching Score Calculation Means 42B)
The matching score calculation means 42B performs classification processing (classification algorithm) by a "supervised classification model" using a matching model that has been pre-learned and stored in the matching model storage means 60 in two-class classification based on whether or not it is selected. and constitutes a classifier for matching together with the matching model storage means 60 . The input to this discriminator is the matching feature data created by the matching feature data creating means 42A, and the output from the discriminator is the probability of selecting a pair of cases consisting of a target case and a candidate case (that is, It is a score (matching score) that indicates the probability that a candidate project is selected for a target project. This score is output for each candidate case (each paired case consisting of a target case and a candidate case). For example, when the score is output in terms of likelihood, a value in the range of 0 to 1 is output.

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 60 is the parameter obtained by learning (learning result data) only.

In addition, the “classification model with supervision” may be any machine learning method as long as it is a method capable of performing machine learning based on two-class classification of whether or not to be 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 machine classifier (SVC: Support Vector Classifier ), Random Forest Classifier, Multi Layer Perceptron Classifier (MLP Classifier), Neural Network, etc. can be employed. This embodiment employs a random forest classifier.

Note that the classifier configured by the matching score calculation means 42B and the matching model storage means 60 inputs the same input data (matching feature data) to a plurality of types of classifiers as described above, and outputs from each classifier. The integrated score may be output by calculating the average value of the scores output from each classifier, or weighting may be performed during the integration (the integrated score ).

Also, the matching score calculation means 42B calculates the score (matching score) between the target case and each candidate case (all cases other than the target case), and then determines the recommended cases in descending order of the calculated score. , the item data of the determined recommended item is acquired from the item data storage means 54 . At this time, it is possible to select and determine the proposals with the highest score of a predetermined number as the recommended proposals, or to determine the recommended proposals by setting a threshold for the score.

(Configuration of Processing Means 30/Similar Matter Search Means 43)
The similar case search means 43 includes a case similarity degree calculation means 43A and a similar case extraction means 43B.

The case similarity calculation means 43A uses the unique text data (self data and needs data that are not combined in this embodiment) for searching for similar cases of each case stored in the case data storage means 54 to calculate the case A project similarity (PR score and needs score in this embodiment) indicating the degree of similarity between 68 to store the data. More detailed contents will be described later.

When the input receiving unit 40 receives the input of the item identification information (case number), the similar case extraction unit 43B extracts the case similarity calculated by the case similarity calculation unit 43A and stored in the case similarity storage unit 68. This is to execute a process of extracting similar cases to the input target case using the degree. More detailed contents will be described later.

(Configuration of processing means 30/similar case search means 43/case similarity degree calculation means 43A: FIGS. 2 and 10)
As shown in FIG. 2, the case similarity calculation means 43A calculates all the words (word 1, word 2, word 3, word 4, word 5, word 5, word 1, word 2, word 3, word 4, word 5, ) is obtained from the distributed representation model storage means 53, and the average vector of the word vectors of all the words included in one case (a vector consisting of the average values of the corresponding elements of each word vector) is calculated and stored in the text vector storage means 67 as the text vector of the item in association with the item identification information (item number).

At this time, even if the same word appears multiple times in the specific text data of the item, it is regarded as a different word, and the same word vector is prepared for the number of occurrences, and the average value is calculated. For example, even if word 1 and word 4 are the same word, a word vector of word 1 and a word vector of word 4 are prepared, respectively, and an average value is calculated. Then, this process is executed for all cases.

In the present embodiment, as unique text data for retrieving similar projects, unbound self data and needs data are used, so two sentence vectors are created for each project and stored in the sentence vector storage means 67. However, the combined text data may be used to create one sentence vector per issue. Also, as shown in FIG. 10, the self-data does not include the self-industry but only the self-PR, but the self-industry may be included. Needs data includes only detailed needs, but may include the desired industry (the industry of the partner company).

2 and 10, the case similarity calculation means 43A uses the sentence vectors of each case stored in the sentence vector storage means 67 to obtain any pair of cases X and Y (any two The similarity Sword2vec (cosine similarity in this embodiment) between the text vectors of the case numbers X and Y) is calculated. Arbitrary pair projects X and Y are combinations (two out of n combination to choose). In the present embodiment, two sentence vectors are created for one case, one for the self data and the other for the needs data, so the similarity Sword2vec between the sentence vectors is calculated for each case. In this embodiment, word2vec is used as the distributed representation, so the term is used for the suffix. is not limited to

Furthermore, as shown in FIG. 10, the case similarity calculation means 43A uses the word importance vectors (such as the TFIDF vector) stored in the word importance vector storage means 57 to calculate any pair cases X, Y ( A similarity Stf-idf (a cosine similarity in this embodiment) between word importance vectors of two arbitrary case numbers X and Y) is calculated. In this embodiment, the word importance vector storage means 57 stores two word importance vectors for self data and needs data as word importance vectors for searching for similar cases. A similarity Stf-idf between degree vectors is calculated.

10, the item similarity calculation means 43A calculates the topic distribution π(i) (K-dimensional vertical vector, K is the number of topics, and i is the number of items) stored in the topic model storage means 56, as shown in FIG. is used to calculate the similarity Stopic (cosine similarity in this embodiment) between the topic distribution vectors of an arbitrary pair of cases X and Y (a combination of two arbitrary cases numbers X and Y). In this embodiment, the topic model storage means 56 stores two topic distribution vectors for self-data and needs data as topic distribution vectors for searching for similar projects. Calculate the degree Stopic. At this time, in this embodiment, as shown in FIG. 10, the topic distribution vector in the needs data is combined with the needs vector consisting of a set of needs flags stored in the item data storage means 54, and the topic distribution vectors ( The similarity Stopic between the combination vector with the needs vector) is calculated.

Furthermore, as shown in FIG. 10, the case similarity calculation means 43A calculates the similarity Stf-idf between word importance vectors, the similarity Sword2vec between sentence vectors, the similarity Stopic between topic distribution vectors, and their weights. Using the coefficients Wtf-idf, Wword2vec, and Wtopic, calculate the case similarity of an arbitrary pair case X and Y (a combination of any two case numbers X and Y) according to the following formula (4), The calculated case similarity is associated with the case identification information (case numbers X, Y) of pair cases X, Y and stored in the case similarity storage means 68 . In this embodiment, a PR score using self-data and a needs score using needs data are calculated as the project similarity. Wtf-idf, Wword2vec, and Wtopic are stored in the setting data storage means 51 as setting data, but may be fixed values.

Proposal similarity (PR score, needs score)
=Stf-idf×Wtf-idf+Sword2vec×Wword2vec+Stopic×Wtopic (4)

The processing by the case similarity calculation means 43A is executed at the stage of pre-processing (see FIG. 14). Also, when registering a new matter (see FIG. 23), upon receiving an instruction from the new matter registration means 45, the new matter and all existing matters ( all projects other than new projects). Furthermore, it is also executed when the time for updating the topic model comes (see FIG. 16), but it is not necessary to repeatedly execute the processing that is executed in the instant reflection processing of the new item.

(Configuration of Processing Means 30/Similar Matter Search Means 43/Similar Matter Extraction Means 43B)
During operation (during a search by the user), the similar case extraction means 43B receives the input of case identification information (case number) from the case similarity storage means 68 when the input reception means 40 receives an input of case identification information (case number). Acquire the memorized matter similarity associated with the combination of the matter number and all matter numbers other than the matter number (combination of the matter numbers of the paired matter). Similar cases are extracted for the case of the accepted case number), and the case data of the extracted similar case is acquired from the case data storage means 54 . At this time, cases with a predetermined number of higher order similarities may be extracted as similar cases, or similar cases may be extracted by setting a threshold value for the case similarity.

In this embodiment, since the PR score and the needs score are stored as the project similarity in the project similarity storage means 68, the average value or weighted average value of the PR score and the needs score is calculated. It is also possible to calculate an integrated score and extract similar projects based on the integrated score. good too. The integrated score can be obtained by various methods such as simple average, weighted average, harmonic average, weighted harmonic average, and geometric average. It may be stored in the storage means 51 or may be a fixed value.

Further, a higher value or a lower value may be selected from the PR score and the needs score, and similar projects may be extracted based on the selected score value.

(Configuration of processing means 30/output means 44)
When the input receiving means 40 receives the input of the keyword, the output means 44 inputs the project data of the related project decided by the associative search means 41 or the related word in addition to the related project on the display means 80. Displayed on the same screen (see screen 200 in FIG. 12) as the input unit 110 of the keyword or matter identification information (case number) by the receiving means 40, and the input of the matter identification information (case number) is accepted by the input receiving means 40 In this case, the item data of the candidate item with the high score (matching score) calculated by the matching means 42 is displayed as the item data of the recommended item on the same screen as the input unit 110 (see screen 300 in FIG. 13). At the same time, the same screen as the input unit 110 (see screen 300 in FIG. 13) is displayed on the same screen as the input unit 110 (see screen 300 in FIG. 13). It executes the processing shown above.

In addition, the output unit 44 outputs related items displayed on the screen (see the screen 200 in FIG. 12 and the screen 300 in FIG. Item identification information for one or more items selected by the user (person in charge of service or customer who registers a new item) from recommended items and similar items as a reference item when registering a new item A memo input (not shown) of (case number) is accepted, and the case identification information (case number) of the received reference case is stored in the reference case storage means 69 at the time of registration.

(Configuration of processing means 30/new matter registration means 45)
The new item registration means 45 accepts an input for registering item data of a new item from the input unit 90 operated by a user (a customer who registers a new item or a service representative acting for the input). It executes a process of automatically assigning the item identification information (case number) for the new item, and storing the received item data in the item data storage means 54 in association with the assigned item identification information (case number). The contents of the item data are as already explained in detail in the description of the item data preparation means 33 .

At this time, the new case registration unit 45 acquires the case identification information (case number) of the reference case stored in the reference case storage unit 69 at the time of registration, and uses the acquired case identification information of the reference case to store the case data. The item data of the reference item is acquired from the storage means 54 and displayed on the new item registration screen (not shown) on the display means 80 . While looking at the display, the user performs input work for registering the item data of the new item (own item). Also, even if the reference case identification information (case number) of the reference case is not stored in the reference case storage means 69 at the time of registration, if the user remembers the case number of the reference case or notes it on paper, If the case number is entered, it is accepted by the new case registration means 45, and the case data of the reference case is acquired from the case data storage means 54 and displayed on the new case registration screen (not shown).

Further, after the new case data is registered, the new case registration means 45 uses the case data of the new case stored in the case data storage means 54 to perform immediate reflection processing of the new case (see FIG. 23). Run. Specifically, an instruction is issued to the unique text data preparation means 34 to prepare the unique text data of the new item (combined text data for matching and associative search, and uncombined self data and needs data for similar item retrieval). Execute the processing to be prepared (processing [α] to [δ], see FIGS. 5, 6, and 10), and associate the post-processing specific text data with the matter identification information (case number) of the new matter to create the matter. Store in the data storage means 54 .

Furthermore, the new case registration means 45 instructs the word importance vector calculation means 37 to instruct the new case specific text data (for matching purposes) created by the specific text data preparation means 34 and stored in the case data storage means 54. and combined text data for associative search, and non-combined self data and needs data for similar item search), calculate and obtain a word importance vector consisting of the TFIDF value or other word importance index value of each word The obtained word importance vector is stored in the word importance vector storage means 57 in association with the case identification information (case number) of the new case. At this time, when the word importance vector is set to the TFIDF value, the TF value can be calculated only with the unique text data of the new project, but the calculation of the IDF value requires the unique text data of all the projects. Since it is necessary, strictly speaking, the TFIDF values of the existing projects other than the new projects will also change slightly. However, the TFIDF value of the existing project may be left unchanged, and only the TFIDF value of the new project may be calculated and additionally stored.

In addition, the new item registration means 45 issues an instruction to the topic value prediction means 36 so that the unique text data after the processing of the new item (combined text data for matching and associative search, and non-combined self text data for similar item search). data, needs data) and the probability of appearance of each word in each topic (matrix β of K rows and p columns) (matrix β of the topic model before update) in each topic stored in the topic model storage means 56, to create a new Predict the topic value (topic distribution represented by the K-dimensional vertical vector π(i)) that indicates the appearance probability of each topic in the item-specific text data, and predict π(i) of the new item before updating. It is stored in the topic model storage means 56 as a temporary value.

Furthermore, the new case registration means 45 issues an instruction to the case similarity calculation means 43A to create a sentence vector of the new case, store it in the sentence vector storage means 67, and furthermore, add a new case and all the existing cases. The calculated project similarity (PR score, needs score in this embodiment) is calculated, and the calculated project similarity is used as the project identification information (project number) of each pair of projects consisting of a new project and each existing project. The item similarity storage means 68 stores the items in association with each other.

The new item registration unit 45 also accepts correction of item data. In the case of minor corrections, the project identification information (project number) will be maintained and the contents will be accepted for correction input. handling.

<Structure of Storage Means 50>

The setting data storage means 51 stores various setting data by the system administrator or the user (person in charge of service) whose input is received by the setting means 31 . The details of this setting data are as already explained in detail in the explanation of the setting means 31 .

The distributed representation model learning data storage means 52 stores text data for distributed representation model learning. The details of the text data for learning the distributed representation model have already been described in the description of the distributed representation model learning means 32 .

As shown in FIG. 2, the distributed representation model storage means 53 stores the distributed representation model as learning result data obtained by the learning process by the distributed representation model learning means 32 . In this embodiment, this distributed representation model is a weight matrix W that is a set of high-dimensional word vectors as a distributed representation of words, and three types of W1, W2, and W3 with different dimensions are prepared.

The case data storage means 54 stores the case data of each case in association with case identification information (case number). The details of this case data are as already explained in detail in the explanation of the case data preparation means 33 .

The unnecessary word dictionary storage means 55 stores unnecessary words (noise words) used in processing [δ] by the unique text data preparation means 34 . The details of this unnecessary word have already been explained in detail in the description of the unique text data preparation means 34 .

The topic model storage means 56 stores π(i) shown in FIG. 18 and FIG. column vector consisting of topic values indicating probabilities) and β (matrix consisting of the appearance probabilities of each word in each topic). Further, when the immediate reflection processing (see FIG. 23) of the new matter by the new matter registration means 45 is executed, π(i) of the new matter obtained by the processing by the topic value prediction means 36 is Store as a temporary value.

The word importance vector storage means 57 stores the word importance vector composed of the TFIDF value or other word importance index value of each word in the unique text data of each case calculated by the word importance vector calculation means 37 for case identification. It is stored in association with information (issue number).

The matching candidate selection result list storage means 58 stores a matching candidate selection result list used when the matching learning data preparation means 38 performs selection/non-selection tagging processing for each piece of matching feature data. .

The matching learning data storage means 59 stores the matching learning data created by the matching learning data preparation means 38 (the data for learning of the matching model, and the matching feature data with labels (tags) indicating selection/non-selection. ) are stored in association with a pair of item identification information (item number).

The matching model storage means 60 stores the matching model obtained by the learning process by the matching model learning means 39 . Together with the matching score calculation means 42B, it constitutes a classifier for matching.

The search history storage unit 61 stores log information of keywords or project identification information (project number) received by the input reception unit 40 .

The model similarity storage means 62 stores the model similarity S calculated by the word similarity calculation means 41A.

The general similarity storage means 63 stores the general similarity A calculated by the word similarity calculation means 41A.

The unique similarity storage means 64 stores the unique similarity B calculated by the word similarity calculation means 41A.

The final similarity storage means 65 stores the final similarity FS calculated by the word similarity calculation means 41A.
is stored.

The degree of association storage means 66 stores the degree of association CS calculated by the degree of association calculation means 41B in association with the case identification information (case number).

The text vector storage means 67 stores the text vector obtained by the case similarity calculation means 43A in association with the case identification information (case number).

The case similarity storage unit 68 stores the case similarity (PR score, needs score) calculated by the case similarity calculation unit 43A in association with the case identification information (case number) of the pair case.

The registration reference case storage unit 69 stores the case identification information (case number) of the reference case at the time of new case registration selected by the user and accepted by the output unit 44 .

<Flow of search: Fig. 11, Fig. 12, Fig. 13>

In FIG. 11, a user (a service representative, a customer who newly registers, or a registered customer who requests introduction of a matching partner) enters a keyword or project identification information (project number) in the input section 110 of the screen 100. input. At this time, if the customer (customer who already has his/her case number) or the person in charge of the service requesting the introduction of the matching partner, he/she (or his/her own customer's) case number is entered in the input section 110. However, in that case, the screen 300 is displayed, and the details thereof will be described later with reference to FIG. 13 . Here, when a customer who newly registers (customer who does not have his/her case number yet) or its service representative registers his/her (or his/her own customer's) information as a new case, he or she can What kind of description content is in the project data of (here, business content), or what kind of description content is recommended by matching for projects with content similar to yours First, a keyword (a word related to one's own business) is input to the input unit 110 in order to confirm whether the

When a keyword (K1, for example, "alcohol") is input to the input unit 110, it is received by the input receiving means 40, and as shown in FIG. A screen 200 shown in FIG. 12 is displayed.

12, the screen 200 is provided with an input section 110, a "Search" button 120 for executing a search, and a "search history" display section . Note that the keyword (K1) that was input immediately before is still displayed in the input section 110 . The keyword (K1) entered first is also displayed in the “search history” display section 130 .

Further, the screen 200 of FIG. 12 is provided with a "similar keyword" display section 210, and this display section 210 displays each word (an input keyword and a plurality of related words related to the keyword). 211 are provided, and when these display portions 211 are selected by an operation such as clicking or tapping, the selected word (inputted keyword or related word) is automatically input to the input portion 110. . A keyword, a related word, or other words can also be input to the input unit 110 by keystroke input without performing a selection operation on each display unit 211 . If the same keyword is input/selected again, the same processing result will be displayed.

Further, the screen 200 of FIG. 12 is provided with a "keyword search list" display section 220. This display section 220 includes a display section 221 of the degree of relevance CS for each related matter and a term (specific to the relevant matter). related words included in the text data), a display portion 222 of the project number (project identification information) of the related project (related company), selection information of "sell/buy" of the needs flag, company's industry, handling There are provided respective display sections for goods, name of area, details of desired area, details of desired industry, self-promotion text data, and text data of needs details. Calculation of the degree of relevance CS, processing of determining the related item, and processing of acquiring the item data of the related item from the item data storage means 54 using the item number of the related item are executed by the associative search means 41. The process of displaying the information received from the display means 80 on the screen is executed by the output means 44 .

When a related case number display section 222 is selected by an operation such as clicking or tapping, the selected related case number is automatically input to the input section 110 . In addition, it is also possible to input the project number of a related project or other project numbers to the input unit 110 by keystroke input without selecting each display unit 222 .

12, a related word is input as a new keyword by a selection operation on the related word display section 211, or another keyword (K2, K3, . . . ) is input by keystroke. 11, the associative search means 41 performs the process again, and the output means 44 displays a screen 200 shown in FIG. On the other hand, in the input section 110 of the screen 200, the matter number (M1, for example, "5678") of the related matter is input by selecting the matter number display portion 222 of the related matter, or the matter number (M1 ) is input, as indicated by solid lines in FIG. In FIG. 11, dotted line arrows indicate the flow of keywords (words), and solid line arrows indicate the flow of case numbers.

In FIG. 13 , screen 300 is provided with input section 110 , “Search” button 120 , and “search history” display section 130 . Note that the input unit 110 keeps displaying the case number (M1, for example, "5678") that was input immediately before. Also, in the "search history" display section 130, the case number (M1) is displayed as the last input information.

The screen 300 of FIG. 13 also includes a "target project" display section 310 that displays the project with the project number (M1) input to the input section 110, and a selection section 311 that determines whether or not to display information on the target project. and are provided. In the "target project" display section 310, a display section 312 of the project number (project identification information) of the target project, selection information of "sell/buy" of the needs flag, company's industry, products handled, area name, desired area details , desired industry details, self-promotion text data, and needs details text data.

Further, the screen 300 of FIG. 13 is provided with a “recommended case list” display section 320 and a selection section 321 for selecting whether or not to display information on recommended cases by matching processing. In the "recommended case list" display section 320, a display section 322 of the case number (case identification information) of the recommended case, the selection information of "sell/buy" of the needs flag, the company's industry, the product handled, the area name, the desired area Details, desired industry details, self-PR text data, need details text data, and score (matching score) display sections are provided. Calculation of the matching score, processing for determining the recommended case, and processing for acquiring the case data of the recommended case from the case data storage means 54 using the case number of the recommended case are executed by the matching means 42 and received from the matching means 42. A process of displaying information on the screen of the display means 80 is executed by the output means 44 .

When the case number display section 322 of the recommended case is selected by an operation such as clicking or tapping, the case number of the selected recommended case is automatically input to the input section 110 . Alternatively, the item number of the recommended item or another item number can be input to the input unit 110 by keystroke input without performing a selection operation on each display unit 322 .

Further, the screen 300 of FIG. 13 is provided with a “similar case list” display section 330 and a selection section 331 for selecting whether or not to display information on similar cases. In the "similar case list" display area 330, a display area 332 of the case number (case identification information) of the similar case, the selection information of "sell/buy" of the needs flag, the industry of the company, the product handled, the name of the area, the desired area Details, desired industry details, self-PR text data, needs details text data, and project similarity (PR score, needs score) are provided. Similar case search processing 43 executes similar case search processing 43 to calculate the degree of case similarity, extract similar cases, and acquire the case data of similar cases from the case data storage means 54 using the case numbers of similar cases. The process of displaying the information received from the process 43 on the screen of the display means 80 is executed by the output means 44 .

When the similar case number display section 332 is selected by an operation such as clicking or tapping, the case number of the selected similar case is automatically input to the input section 110 . In addition, it is also possible to input the case number of a similar case or another case number to the input unit 110 by keystroke input without performing a selection operation on each display unit 332 .

Then, in the input unit 110 of the screen 300 of FIG. 13, the item number of the recommended item is input by the selection operation of the item number display unit 322 of the recommended item, or by the selection operation of the item number display unit 332 of the similar item. When the case number of a similar case is entered, or another case number (M2, M3, . Each process by the search process 43 is executed, and the screen 300 shown in FIG. 13 is displayed by the output means 44 . On the other hand, when a keyword is input to the input section 110 of the screen 300 in FIG. 13, as indicated by the dotted line in FIG. 200 is displayed.

A customer who newly registers (a customer who does not have his/her case number yet) or the person in charge of the service refers to the screen 200 of FIG. 12 and the screen 300 of FIG. By selecting a reference case at the time of registration of a new case (my case) and entering a memo, it is stored in the reference case storage means 69 at the time of registration, or the case number of the reference case is memorized, or a paper make a note of it. Then, while looking at the project data of the reference project displayed on the screen by the new project registration means 45, when inputting for registration of the project data of the new project (his own project) is performed, the new project registration means 45 registers the new project. A matter number is automatically assigned to the matter, and registration processing is executed.

Subsequently, after registering the new project, the customer who made the new registration or the person in charge of the service inputs the project number (Mnew) of the new project to the input section 110 (screen 100), as shown in FIG. , 200, and 300), and by repeating the search, a recommended case for a new case (your own case), a recommended case for a similar case to the new case, and a recommended case for the new case. Display the item data such as items similar to the item on the screen.

<Overall Flow of Processing by Matching System 10: FIGS. 14 and 15>

In FIG. 14, first, the system administrator prepares a large number of existing project data that have already been created, associates them with project identification information (project number), and stores them in the project data storage means 54 (step S1). Data can be migrated from an existing system in operation. This processing is executed by the case data preparation means 33 .

Next, preprocessing for associative search (see FIG. 17 for details) is executed (step S2). Subsequently, preprocessing for matching (construction of topic model, initial model of matching model, etc., see FIG. 18) is executed (step S3). Further, preprocessing for searching for similar items (similar companies) (see FIG. 19 for details) is executed (step S4). These pre-processing steps S2 to S4 can all be executed in real-time processing during searching (at the time of operation), but the reaction speed becomes slower.

In FIG. 15, after the pre-processing (steps S1 to S4) is completed, when the operation stage is entered, the search processing by the user is performed. In the search process, first, the input reception unit 40 receives input of a keyword or a case number (case identification information) by the user at the input unit 110 (see FIG. 11) of the screen 100 (step S5). From the second time onwards (when returning from step S11, which will be described later), the input section 110 of the screens 200 and 300 accepts input (including selection input).

Next, the input reception means 40 determines whether or not the information received by the input unit 110 is a keyword (step S6). 20) is executed (step S7), and if it is not a keyword (that is, if it is a case number), matching processing (see FIG. 21 for details) is executed by the matching means 42 (step S8), A similar project (similar company) search process (see FIG. 22 for details) is executed by the similar project search means 43 (step S9).

Subsequently, the output means 44 displays the result of the associative search process (see FIG. 12) on the screen, or displays the results of the matching process and the similar job (similar company) search process (see FIG. 13) on the screen (see FIG. 13). step S10). Then, if the search is to be continued (step S11), the process returns to step S5 to accept input (including selection input) of a keyword or project number (project identification information) by the user. On the other hand, if the search is not to be continued, the search processing is terminated as it is, or if there is a memo input of the matter number of the reference matter at the time of registering the new matter, the memo input is accepted by the output means 44. End the search process.

Then, the user (the company that newly registers the item data, or the person in charge of the service) looks at the item data of the reference item displayed on the screen, and performs a new item registration process for registering his or her own item data. (reception of registration of a new matter, assignment of a matter number to the new matter, immediate reflection processing, see FIG. 23) is executed (step S12). This processing is executed by the new case registration means 45 . Then, the user who has completed the registration of the new project can immediately reflect the new project before the timing of updating the topic model including the new project data. , the process returns to step S5, the case number of the new case is entered in the input unit 110, and the search is repeated. After that, when it is time to update the topic model including the new project data, the process proceeds to the next update process (steps S13 to S15) in FIG.

As shown in FIG. 16, the timing of updating the topic model is (1) when the number of new item data registrations reaches a predetermined number, (2) each time new item data is registered, (3) Every day (every night), every week, every month, etc., provided that new project data is registered during the most recent period; Either is fine. In the case of (2), if update processing is to be executed immediately after registering new matter data, there is no need to immediately reflect the new matter by the new matter registration means 45 (see FIG. 23). However, after registering new project data, update processing is not executed immediately. For example, when update processing is executed in the nighttime batch of the same day, there is a time lag between registration processing and update processing, so new project data Immediate reflection processing of the new matter by the registration means 45 is performed.

In FIG. 16, when it is time to update, first, update processing for associative search is executed (step S13). This processing is the same as the pre-processing for associative search in step S2 (see FIG. 17 for details). However, since it is not necessary to update the distributed representation model frequently, it may be omitted.

Next, update processing for matching is executed (step S14). This processing is the same as the preprocessing for matching in step S3 (see FIG. 18 for details). At this time, although the topic model is updated, the updating of the matching model may be omitted. Therefore, the update of the topic model and the update of the matching model do not need to be performed with the same frequency, nor need to be performed at the same time.

Subsequently, an update process for searching similar projects (similar companies) is executed (step S15). This processing is the same as the pre-processing for searching for similar items (similar companies) in step S4 (see FIG. 19 for details). At this time, the processing that is immediately reflected at the time of registration of the new item may be omitted.

<Flow of preprocessing for associative search in step S2: Fig. 17>

In FIG. 17, first, when the system administrator operates the input means 90 and inputs setting data for associative search (N3=for example, 100, N6=for example, 10, number of related items to be displayed, etc.), the setting means 31 , this setting data is accepted and stored in the setting data storage means 51 (step S201). The setting means 31 also accepts and saves input of setting data for matching (the number of recommended projects to be displayed, etc.) and setting data for searching similar projects (Wtf-idf, Wword2vec, Wtopic, the number of similar projects to be displayed, etc.). You can go here.

Next, the system administrator operates the input means 90 to save the previously acquired and collected text data for learning the distributed representation model in the data storage means 52 for learning the distributed representation model (step S202). . This processing is performed by the setting means 31 .

Subsequently, the distributed representation model learning means 32 executes distributed representation model learning processing using the text data for distributed representation model learning stored in the distributed representation model learning data storage means 52, and obtains The learning result data (weight matrix W composed of word vectors) is stored in the distributed representation model storage means 53 (step S202). It should be noted that the disclosed/provided learning result data (weight matrix W) may be acquired and stored in the distributed representation model storage means 53 .

In the present embodiment, the distributed representation model for associative search and for similar item search are shared, but even if they are not shared, the distributed representation models for both may be prepared in the process of step S202.

<Flow of preprocessing for matching in step S3: FIG. 18>

In FIG. 18, the unique text data preparation means 34 acquires self-data (own industry, company PR) and needs data (desired industry, details of needs) for all projects (in this case, companies) from the project data storage means 54. Then, unique text data (combined text data in this embodiment) for matching and for associative search (common) is created (step S301). In the case of model update processing (step S14 in FIG. 16), unique text data may be additionally created for a new case.

Subsequently, the unique text data preparation means 34 removes unnecessary symbols and tags from all the unique text data (here, combined text data) created in the process of step S301, and breaks it down into words by morphological analysis. , extraction of only nouns, and removal of unnecessary words are executed, and the unique text data after processing is stored in the case data storage means 54 in association with the case identification information (case number) (step S302). The contents of these processes have already been described in detail in the description of the inherent text data preparation means 34, so detailed description thereof will be omitted here. In the case of the model update process (step S14 in FIG. 16), these processes should be executed only for the unique text data of the new item.

Next, the topic estimating means 35 performs soft clustering or a neural language model (here, , LDA), and the topic value π(i) that indicates the appearance probability of each topic (topic number = 1 to K) in the unique text data (i = 1 to n) of each project (each company) , and the appearance probability β of each word in each topic, and store the obtained π(i) and β for matching in the topic model storage means 56 (step S303). The content of this topic estimation process has already been explained in detail in the explanation of the topic estimation means 35, so a detailed explanation is omitted here. In the case of the model update process (step S14 in FIG. 16), the topic estimation process is started with the number of projects (number of companies) n corresponding to the number of documents increased by the number of newly registered projects. executed.

After that, the word importance vector calculation means 37 uses all of the processed unique text data (here, combined text data) (i=1 to n) for matching stored in the case data storage means 54. , the word importance vector (here, the TFIDF vector) is calculated, and the calculated word importance vector for matching is stored in the word importance vector storage means 57 in association with the case identification information (case number) (step S304). ). The details of this process have already been explained in the explanation of the word importance vector calculating means 37, so the detailed explanation is omitted here. In the case of model update processing (step S14 in FIG. 16), the TD value can be calculated using only the unique text data of the new project, but the calculation of the IDF value requires the unique text data of all projects. Therefore, the word importance vector is calculated for all projects including new projects.

Subsequently, the matching learning data preparation means 38 treats pairs existing in the matching candidate selection result list stored in the matching candidate selection result list storage means 58 as positive examples (correct data), Pairs other than the pairs to be matched are treated as negative examples (incorrect answer data), and a large number of pairs are prepared for creating matching feature data (see FIG. 9) as matching learning data (step S305). The contents of this process have already been described in detail in the description of the matching learning data preparation means 38, so detailed description thereof will be omitted here. In the case of the model update process (step S14 in FIG. 16), even if the topic model is updated by executing the topic estimation process, it is not always necessary to update the matching model, so this process is omitted. You may When the topic model is updated, the topic distribution of each project will change slightly, so strictly speaking, the feature data for matching will also change. It is from.

Then, the matching learning data preparation means 38 stores each topic value π(i) from the topic model storage means 56 for all pair cases prepared for creating the matching learning data in the process of step S305. Needs flags and annual sales are obtained from the project data storage means 54, and composite variables (product of each topic value, absolute value of difference, logical sum of needs flags, ratio of annual sales, etc.) are calculated. (Step S306). Further, the word importance vector for matching stored in the word importance vector storage means 57 (here, , TFIDF vector) is calculated, matching feature data (see FIG. 9) is created as matching learning data, and is associated with the case identification information (case number) of the pair case, and the matching learning data storage means 59 (step S307). The contents of these processes have already been described in detail in the description of the matching learning data preparation means 38, so detailed description thereof will be omitted here. In the case of model update processing (step S14 in FIG. 16), the matching model update processing may be omitted, as in step S305 above.

Subsequently, the matching model learning means 38 uses the matching learning data stored in the matching learning data storage means 59 to perform two-class classification of selected and unselected by a supervised classification model (for example, random forest). is obtained, and the obtained matching model is stored in the matching model storage means 60 (step S308). The contents of this process have already been described in detail in the description of the matching model learning means 38, so a detailed description will be omitted here. In the case of the model update process (step S14 in FIG. 16), the matching model update process may be omitted as in the above steps S305 to S307.

<Flow of pre-processing for searching similar projects (similar companies) in step S4: Fig. 19>

In FIG. 19, the unique text data preparation means 34 acquires self-data (own industry, company PR) and needs data (desired industry, details of needs) for all projects (in this case, companies) from the project data storage means 54. Then, two types of unique text data for searching for similar items (in this embodiment, self data and needs data are not combined) are prepared (step S401). At this time, as the self data, only the company's PR may be used without including the own industry, and as the needs data, only the detailed needs may be used without including the desired industry. See Figure 10). In the case of model update processing (step S15 in FIG. 16), unique text data may be additionally created for a new case.

Subsequently, the unique text data preparation means 34 removes unnecessary symbols and tags from all the two types of unique text data (non-combined self data and needs data) for searching similar cases prepared in the process of step S401. etc., breaking down into words by morphological analysis, extracting only nouns, and removing unnecessary words. (step S402). The contents of these processes have already been described in detail in the description of the inherent text data preparation means 34, so detailed description thereof will be omitted here. In the case of the model update process (step S15 in FIG. 16), these processes should be executed only for the unique text data of the new item.

Next, the topic estimating means 35 uses all of the two types of processed unique text data (i=1 to n) for searching for similar cases stored in the case data storage means 54 to obtain two types of respective text data. Separately, topic estimation processing by soft clustering or neural language model (here, LDA) is executed, and each topic ( The topic value π(i) indicating the probability of appearance of topic numbers = 1 to K) and the probability of appearance β of each word in each topic are obtained, and the obtained two types of π(i) and β for similar matter search are obtained. (That is, .pi.(i) and .beta. for self data, .pi.(i) and .beta. for needs data) are stored in topic model storage means 56 (step S403). The content of this topic estimation process has already been explained in detail in the explanation of the topic estimation means 35, so a detailed explanation is omitted here. In the case of the model update process (step S15 in FIG. 16), the topic estimation process is started with the number of projects (number of companies) n corresponding to the number of documents increased by the number of newly registered projects. executed.

After that, the word importance vector calculation means 37 performs two types of processed unique text data (here, unbound self data and needs data) (here, unbound self data and needs data) (i = 1 to n), the word importance vector (here, the TFIDF vector) is calculated, and two types of word importance vectors for searching similar projects (that is, self data and needs data) are calculated. ) is associated with the case identification information (case number) and stored in the word importance vector storage means 57 (step S404). The details of this process have already been explained in the explanation of the word importance vector calculating means 37, so the detailed explanation is omitted here. In the case of model update processing (step S15 in FIG. 16), the TD value can be calculated using only the unique text data of the new project, but the calculation of the IDF value requires the unique text data of all projects. Therefore, two types of word importance vectors are calculated for all projects including new projects.

Subsequently, the project similarity calculation means 43A uses all of the two types of processed unique text data (i=1 to n) for searching for similar projects to calculate each unique text data (each own data, each needs data) ), as shown in Fig. 2, the average value of the distributed representation of all the words that make up each unique text data (here, word vector by word2vec) is calculated, and these are used as the unique text data of each case. The text vector is stored in the text vector storage means 67 in association with the item identification information (item number) (step S405). At this time, for each project (each company), the average value (sentence vector) of the distributed representation of the self data and the average value (sentence vector) of the distributed representation of the needs data are calculated separately. The content of this process has already been described in detail in the description of the case similarity degree calculation means 43A, so a detailed description thereof will be omitted here.

Then, the project similarity calculation means 43A calculates the degree of similarity Stf-idf between the word importance vectors stored in the word importance vector storage means 57 for each pair in the round robin of all projects (companies), the sentence vector The similarity Sword2vec between sentence vectors stored in the storage means 67 and the similarity Stopic between topic distribution (topic value) vectors stored in the topic model storage means 56 are calculated, and each of these similarities is used as a weighting factor. (Wtf-idf, Wword2vec, Wtopic) to calculate the project similarity that indicates the degree of similarity between projects (companies). ) and stored in the case similarity storage means 68 (step S406). At this time, as the degree of project similarity, for each pair project, a PR score indicating the degree of similarity with respect to self data and a needs score indicating the degree of similarity with respect to needs data are separately calculated. The content of this process has already been described in detail in the description of the case similarity degree calculation means 43A, so a detailed description thereof will be omitted here.

<Flow of associative search processing in step S7: Fig. 20>

In FIG. 20, first, the word similarity calculation means 41A inputs and calculates each distributed representation model (in this embodiment, a set of three types of word vectors with different dimensions) stored in the distributed representation model storage means 53. Calculate the model-specific similarity S such as the cosine similarity between the selected and input keyword and other words, and for each distributed representation model, the words whose model-specific similarity S is within the top N3 (for example, within the 100th) are selected as similar words by model, and the similarity by model S of each similar word by model is stored in the similarity by model storage means 62 (step S701). The details of this process have already been explained in the explanation of the word similarity calculation means 41A, so the detailed explanation is omitted here. Note that the model-specific similar words are shifted in each distributed representation model.

Next, for all of the words (N4) selected by at least one distributed representation model, the word similarity calculation means 41A calculates the model-specific similarity S of the word of the distributed representation model for which the corresponding word is not selected. is zero (see FIG. 2), and by obtaining the harmonic average or weighted harmonic average of the similarity S for each model of all distributed representation models (here, three types of distributed representation models), multiple types of variance The general similarity A is calculated by ensemble of the expression model, the words for which the calculated general similarity A is not zero are determined as general similar words (N5 pieces), and the general similarity A of each general similar word is the general similarity stored in the degree storage means 63 (step S702). The details of this process have already been explained in the explanation of the word similarity calculation means 41A, so the detailed explanation is omitted here.

Subsequently, the word similarity calculation means 41A calculates all of the unique text data for associative search (self data and needs data in this embodiment) for each general similar word (N5) and the input/selected keyword. ) to extract the project number of the project (here, company) associated with the unique text data containing at least one of these words (step S703). Further, the word similarity calculation means 41A compares each general similar word (N5) and the specific text data of each matter extracted by matching with the input/selected keyword, and compares the input/selected keyword with , and each of the general similar words (N5) is calculated, and the calculated inherent similarity B is stored in the inherent similarity storage means 64 (step S704). The contents of these processes have already been described in detail in the description of the word similarity calculation means 41A, so detailed description thereof will be omitted here.

Then, the word similarity calculation means 41A calculates the final similarity FS by the harmonic average or the weighted harmonic average of the general similarity A and the specific similarity B for each general similar word (N5). Words whose FS are within the top N6 (for example, within the top 10) are determined as related words, and the final similarity FS of each related word is stored in the final similarity storage means 65 (step S705). The details of this process have already been explained in the explanation of the word similarity calculation means 41A, so the detailed explanation is omitted here.

After that, the degree-of-association calculation means 41B converts each related word (N6=10) and the input/selected keyword into unique text data for associative search (combined text data in this embodiment). is multiplied by the final degree of similarity FS(i), and these values are summed for each word to obtain the relation Calculate the degree CS, determine the related cases in descending order of the degree of association CS, store the degree of association CS of each related matter in the degree of association storage means 66, and use the matter number of the related matter to store the matter data storage means. The item data of the related item is acquired from 54 (step S706). The details of this process have already been explained in the explanation of the degree-of-association calculation means 41B, so detailed explanation is omitted here.

<Flow of matching processing in step S8: Fig. 21>

In FIG. 21, first, the matching feature data creation means 42A selects and inputs target projects (companies) into the input unit 110 (see FIGS. 11 to 13) and all other projects that are candidates for recommendation. For each pair with (company), each topic value π(i) is acquired from the topic model storage means 56, a needs flag and annual sales are acquired from the item data storage means 54, and a combined variable (each topic value product, absolute value of difference, logical sum of needs flags, ratio of annual sales, etc.) is calculated (step S801). Furthermore, the matching feature data creating means 42A calculates the word importance for each pair of the target case (company) input/selectively input to the input unit 110 and all the other cases (company) that are candidates for recommendation. After calculating the inner product of the word importance vectors (here, TFIDF vectors) stored in the vector storage means 57 (step S802), matching feature data (see FIG. 8) is created (step S803). The contents of these processes have already been described in detail in the description of the matching feature data creation means 42A, so detailed description thereof will be omitted here.

Next, the matching feature data created in the process of step S803 is used as a classifier for matching (matching score calculation means 42B and matching model storage means 50) by a supervised classification model (for example, random forest). discriminator), and outputs a score (matching score) indicating the probability that each candidate project (candidate company) will be selected for the target project (company) input/selected to the input unit 110. Recommended items (recommended companies) are determined in descending order of score, and the item data of the recommended items are obtained from the item data storage unit 54 using the item numbers of the determined recommended items (step S804).

<Flow of similar project (similar company) search processing in step S9: Fig. 22>

In FIG. 22, the similar project extracting means 43B extracts the target project (company) of the project number input/selectively input by the input unit 110 (see FIGS. 11 to 13) and all others from the project similarity storage means 68. acquires the project similarity (in this embodiment, the PR score and the needs score) with the project (company) (step S901).

Subsequently, the similar case extraction means 43B extracts similar cases in descending order of case similarity, and acquires the case data of the similar cases from the case data storage means 54 using the case numbers of the extracted similar cases ( step S902).

<Flow of New Matter Registration Processing in Step S12: FIG. 23>

In FIG. 23, the new matter registration means 45 receives an input for registering the matter data of the new matter from the user (step S1201), automatically assigns the matter identification information (case number) to the new matter, and accepts it. The project data of the new project is stored in the project data storage means 54 in association with the assigned project number of the new project (step S1202). At this time, the new case registration means 45 acquires the case data of the reference case from the case data storage means 54 using the case identification information (case number) of the reference case stored in the reference case storage means 69 at the time of registration. , is displayed on the screen of the display means 80 .

After that, according to an instruction from the new item registration means 45, the following immediate reflection processing of the new item before updating is executed.

First, the inherent text data preparation means 34, which receives an instruction from the new matter registration means 45, acquires self data and needs data from the matter data storage means 54 for the new matter (company), and then prepares the following inherent text data. Prepare (step S1203). Specifically, unique text data used in matching processing and associative search processing (text data combined with self data and needs data) and unique text data used in similar case search processing (non-combined self data, needs data) are prepared. .

Subsequently, the unique text data preparation means 34 receiving the instruction from the new item registration means 45 prepares the unique text data of the new item (company) (combined text data for matching and associative search (shared), similar item search). (unconnected self-data, needs data), remove unnecessary symbols and tags, break down into words by morphological analysis, extract only nouns, remove unnecessary words, and each unique text after processing The data is stored in the case data storage means 54 in association with the case identification information (case number) (step S1204).

Next, the word importance vector calculation means 37, which receives the instruction from the new case registration means 45, uses the unique text data (i=1 to n+1) after processing all the cases including the new case (company). , calculates the next word importance vector (here, the TFIDF vector) for the unique text data (i=n+1) after processing of the new matter, associates it with the matter identification information (case number), and stores the word importance vector storage means 57 (step S1205). That is, the word importance vector used in the matching process (the word importance vector of the combined text data) and the word importance vector used in the similar case search process (the word importance vectors of the self data and the needs data) are calculated. .

Then, the topic value prediction means 36 receiving the instruction from the new matter registration means 45 predicts the next topic distribution π(i) for the unique text data (i=n+1) after processing of the new matter (company). , the obtained topic distribution π(i) is stored in the topic model storage means 56 in association with the item identification information (item number) (step S1206). That is, the topic distribution π(i) of the combined text data used in the matching process and the topic distribution π(i) of each of the self data and the needs data used in the similar item search process are predicted.

Subsequently, the project similarity calculation means 43A, which receives an instruction from the new project registration means 45, uses the processed unique text data of the new project (company) to distribute all the words that make up the unique text data. An average value of the expressions (here, word vectors based on word2vec) is calculated as a sentence vector, and the obtained sentence vector is associated with the case identification information (case number) and stored in the sentence vector storage means 67 (step S1207). . At this time, the average value (sentence vector) of the distributed representation of the self data and the average value (sentence vector) of the distributed representation of the needs data are calculated separately.

Furthermore, by the project similarity calculation means 43A that receives an instruction from the new project registration means 45, the word importance vector storage means 57 stores the new project (company) and all the existing projects. Similarity between word importance vectors Stf-idf, similarity between sentence vectors stored in sentence vector storage means 67 Sword2vec, similarity between topic distribution (topic value) vectors stored in topic model storage means 56 By calculating the degree Stopic and integrating each of these similarities using weighting coefficients (Wtf-idf, Wword2vec, Wtopic), the project similarity indicating the similarity between projects (between companies) is calculated and calculated The obtained case similarity is associated with the case identification information (case number) of the pair case and stored in the case similarity storage means 68 (step S1208). At this time, as the project similarity, the PR score indicating the similarity of the self data and the needs score indicating the similarity of the needs data are separately calculated.

<Effects of this embodiment>

According to this embodiment, the following effects are obtained. That is, in the matching system 10, when the user (the person receiving the matching service or the person in charge of the service) enters a keyword into the input unit 110 (see FIGS. 11 to 13), the associative search means 41 performs processing. , the item data of related items and related words can be displayed on the same screen as the input unit 110 . On the other hand, when the case identification information (case number) is input to the input unit 110, the processing by the matching unit 42 and the processing by the similar case search unit 43 are executed, and the target case (inputted case) is displayed on the same screen as the input unit 110. It is possible to display the item data of the item recommended as a matching partner for the item with the selected item number) and the item data of the item similar to the target item.

At this time, the three processes by the associative search means 41, the matching means 42, and the similar case search means 43 are all the unique text data of each case, that is, the same data However, in any process, self data and needs data are used.), so in addition to being able to realize processing suitable for matching in all three processes, the processing contents are related to each other. be able to. Therefore, the matching system 10 is not simply a combination of the three existing systems, but can share data in each process and standardize the contents of the process.

Further, the result of each process by the associative search means 41, the matching means 42, and the similar case search means 43 can be utilized in the next process by these means 41, 42, 43. , 43 can be repeated, allowing the user to perform a variety of searches.

More specifically, when a keyword is input to the input unit 110 (see FIGS. 11 to 13), and a related project related to the keyword is obtained through processing by the associative search means 41, the related project identification information ( case number) into the input unit 110, it is possible to obtain recommended cases as matching partners for the related case, and to obtain similar cases to the related case. By inputting the project identification information (project number) of the obtained recommended project to the input unit 110, it is possible to obtain a recommended project as a matching partner for the recommended project, and obtain a similar project to the recommended project. can be done. Further, by inputting the project identification information (project number) of the obtained similar project to the input unit 110, it is possible to obtain a recommended project as a matching partner for the similar project, and obtain a similar project for the similar project. can be done. In the matching system 10, it is possible to repeatedly search across related projects, recommended projects, and similar projects in this way, and during that time various keywords can be input at arbitrary timing (see FIG. 11).

For this reason, if the user is a person who registers a new project or a service representative who assists, guides, and guides the user, he or she may be in the same position as the customer he/she is in charge of (for example, a certain product or merchandise). You can easily understand the content of the project data (self-data and needs data included in it) about companies that wish to match for the purpose of selling products), and if you register what kind of project data, Since it is possible to easily grasp what kind of project data a project has to be recommended as a matching partner (that is, what kind of project data has a high degree of matching accuracy), The information referred to in those searches can be utilized when registering the item data of a new item. Therefore, the user can register suitable job data from the viewpoint of matching and receive effective matching service.

In addition, since a person who registers a new item can search and refer to such item data by himself/herself, it is possible to reduce the burden on the person in charge of service. Furthermore, for example, even when a service representative conducts a search at a service counter or at a customer's site, it is possible to directly show the contents of project data of several reference projects obtained by the search to the person who registers a new project. Therefore, it is possible to reduce the burden on the person in charge of service (trouble in providing assistance, guidance, and guidance).

In addition, the matching means 42 does not update the topic model obtained by the topic estimation process or the matching model obtained by learning using the matching feature data. Since it is output, the same item is always output as a recommended item with the same order (same score) for a certain item. And if this becomes a selection result and is reflected in the updating of the matching model, it will lead to promoting biased introductions. However, in the matching system 10, by performing the processing by the similar case search means 43, it is possible to grasp the similar cases of the target case. Since the projects can be output, not only recommended projects for the target project but also recommended projects for projects similar to the target project can be output, and the service staff can introduce them. Similarly, by repeating the search, it is possible to output similar cases of the recommended cases to the target case and similar cases of the recommended cases to the similar cases of the target case, and the service personnel can introduce them. Therefore, since the service staff can find a project to introduce from various angles, even a staff member with little know-how can introduce a suitable project as a matching partner to the matching requester, and the introduction project Therefore, it is possible to prevent the matching model from being updated with biased learning data.

Furthermore, the associative search means 41 calculates the final similarity FS using the general similarity A and the specific similarity B calculated using the specific text data, and determines related words in descending order of the final similarity FS. Therefore, the related words can be not just general related words (words with a high general similarity A to the input keyword), but can be related words reflecting the specific text data. Therefore, from the viewpoint of matching, a more effective related word can be input as a new keyword. In other words, in the processing by the associative search means 41, since the expression (3) for obtaining the degree of relevance CS used for determining the related case includes the frequency of occurrence C(i) for each word in the specific text data, this relation In addition to this, the final similarity FS used for determining related words used in determining related cases is also calculated using the unique text data. Since the calculated unique similarity B is involved, more effective processing can be realized from the viewpoint of matching.

In addition, since the matching system 10 is provided with the topic value prediction means 36, it is possible to reflect the unique text data of the new item in the search without updating the topic model (immediate reflection of the new item in FIG. 23). processing). Therefore, even before the topic model is updated by the topic estimation process including the unique text data of the new issue, the matching process can be performed with the new issue as the target issue. Therefore, before the topic model is updated, a search can be performed by repeating each process by the associative search means 41, the matching means 42, and the similar case search means 43, including new cases. In other words, even before the topic model is updated, the item identification information (item number) of the new item can be input to the input unit 110 (see FIGS. 11 to 13). can be a related item, a recommended item, or a similar item displayed on the screen as a processing result.

Further, the similar project search means 43 constructs a topic distribution vector composed of topic values indicating the appearance probability of each topic in the unique text data of each project and the unique text data of each project, as shown in Equation (4). Using a sentence vector consisting of the average value of the distributed representation of each word and a word importance vector (TFIDF vector in this embodiment) for the unique text data of each case, The project similarity (PR score, needs score) is calculated, and similar projects to the target project are extracted in order of the high project similarity. It is possible to realize the process of extracting similar cases that accurately reflect the content and are useful for matching.

Further, as shown in FIG. 3, the associative search means 41 is configured to calculate the unique similarity B using the number of occurrences of words. A reflected inherent similarity B can be calculated, a final similarity FS can be calculated using the inherent similarity B, and related words can be determined using the final similarity FS. Therefore, from the viewpoint of matching, more appropriate related word determination processing can be performed.

Further, the input receiving means 40 receives input of keywords and input of project identification information (project number) by the same input unit 110 (see FIGS. 11 to 13), and automatically determines which input is received. Therefore, the configuration of the screen on which the input unit 110 is arranged and the search results are displayed is simplified, and the operability for the user can be improved.

Then, using the log information stored in the search history storage unit 61, the input reception unit 40 uses the "search history" display unit 130 ( 11, the illustration of the display unit 130 is omitted.), so that the user can see what kind of keyword or item identification information (item number) and what kind of It is possible to easily grasp whether the user has input the information in the appropriate order and performed the search. Therefore, it is possible to improve the user's search work efficiency.

In addition, the output means 44 receives a memo input (not shown) of the matter identification information (case number) of the reference case and stores it in the reference case storage means 69 at the time of registration. By using the case identification information of the reference case stored in the reference case storage means 69, the case data of the reference case is displayed on the new case registration screen (not shown) displayed on the display means 80. Therefore, the user can perform the input work for registering the item data of the new item (own item) while viewing the display. Therefore, appropriate item data can be easily registered.

<Mode of deformation>

It should be noted that the present invention is not limited to the above-described embodiments, and includes modifications within the scope of achieving the object of the present invention.

(Modification of search history display: FIG. 24)
For example, in the above-described embodiment, the "search history" display section 130 (illustration of the display section 130 is omitted in FIG. 11) provided on each screen 100, 200, 300 (see FIGS. 11 to 13). , the search history is displayed by arranging keywords or project identification information in the input order, but the search history may be displayed as shown in FIG.

In FIG. 24, a screen 400 is provided with an input section 110 and a "Search" button 120 in the same manner as the screens 100, 200, and 300 (see FIGS. 11 to 13) of the above embodiment. A "search history" display section 430 different from the "search history" display section 130 (illustration of the display section 130 is omitted in FIG. 11) is provided.

In the "search history" display area 430, keywords or item identification information (item numbers) are arranged in order of input from the bottom line upward. Therefore, the example of FIG. 24 indicates that the input to the input unit 110 is performed in the order of the keyword "alcohol", the case number "5678", the case number "6006", . Note that the input order may be downward from the top line.

In addition, the “search history” display section 430 is provided with three columns. Keywords are displayed in the left column, the project identification information (project number) on the same side as the first input project identification information (project number) is displayed in the center column, and the first input is displayed in the right column. The project identification information (project number) and the project identification information (project number) on the opposite side are displayed. "Same side" and "opposite side" are side information distinguished from the viewpoint of whether it is the side for which matching is requested or the matching partner side. Note that the order in which the keyword, the item identification information on the same side, and the item identification information on the opposite side are arranged in the horizontal direction (arrangement position of each column) is not limited to the example in FIG. should be distinguished by In the example of FIG. 24, the vertical direction indicates the order of input (time information), and the horizontal direction indicates the keyword, and the same side/opposite case identification information. The relationship with the direction may be exchanged.

In the "search history" display section 430, the case of the first input case identification information (case number) becomes the starting case (company), and sorted display is performed between the center column and the right column. In the example of FIG. 24, the case number "5678" is the starting point. When a similar item (similar company) is selected and input to the input unit 110, it is displayed in the same column, and when a recommended item (recommended company) is selected and input to the input unit 110, it is displayed in the opposite column. Therefore, the matching partner (recommended case) of the matching partner (recommended case) returns to the original column. Once the keyword is input to the input unit 110, this relationship is reset, and thereafter, the project with the first input project identification information (project number) becomes the new starting project. In the example of FIG. 24, the case number "18777" is the new starting point. It is also possible to key in the case number of a case that is neither a similar case nor a recommended case to the input unit 110, but in this case as well, the relationship up to that point is reset, and the keyed in case number becomes a new starting point. becomes.

Further, in the "search history" display section 430, a character display is also performed to indicate which case number of the related case, the similar case, and the recommended case has been selected and input. For example, the characters "similar" in "6006/similar (sales)" indicate that it is a similar case with the case number "5678" that was input immediately before. In addition, the character "(sale)" in "6006/similar (sale)" is the information of the need flag, and indicates that the need flag of the case with the case number "6006" is "sell".

In the "search history" display section 430, the information of the need flag (in this case, by trade), and the side information by the sorting display of the center column and the right column according to the distinction between the input of the similar item and the input of the recommended item. It also displays the consistency information for That is, if the needs flag of the first input project number is "Sell", the recommended project for that project becomes the "Buy" project, and the recommended project for that recommended project returns to the "Sell" project. Therefore, it is natural to have the "sell" deal numbers in the middle column and the "buy" deal numbers in the right column. If the needs flag of the project with the first input project number is "Buy", it is natural that all of these relationships are reversed. Therefore, if such a natural state is maintained for the needs flag of the project (starting project) of the first input project number, normal background color (e.g. white) and normal color text (e.g. black) ) is displayed. On the other hand, it is clear that there are projects whose needs flag is “buy” (both “sell” and “buy”), and that self-data and needs data, which are text data, are not necessarily “sell” and “buy”. Since the information is not divided into two, the above natural state may not be maintained. display is performed.

Also, whether the matter (company) of the matter number entered by the user first is the matter of the user (the same side as the user) or the matter of the matching partner (the opposite side of the user) depends on the former case. Although there are many, the system cannot accurately grasp it unless it accepts the user's expression of intention. However, in any case, that is, even if the user's search intention (which side was selected first) is unknown, the case number that was entered first remains the starting point. It is possible that the needs flag of this starting point is "trading", but in that case, the user's search intention is unknown from the beginning even from the information of the needs flag, so all subsequent displays (reset ) is displayed with a yellow background or yellow letters to alert the user that the display of consistency information does not work.

As described above, the display of the "search history" display section 430 by the input receiving means has the following functions (1) to (5).

(1) There is a basic function that indicates the type of input and the order of input, i.e., whether the input is a keyword or item identification information (item number).

(2) Recommendation/similarity discrimination for distinguishing whether input of project identification information for a recommended project for the target project (project with the input project number) has been received, or whether input of project identification information for a project similar to the target project has been received It has an information display function. In the example of FIG. 24, the characters "recommended" and "similar" are used. However, the recommendation/similarity distinguishing information may be realized by, for example, color-coded display, presence/absence of hatching, presence/absence of a frame, etc., instead of character display.

(3) There is a function of displaying side information for distinguishing whether each case for which the input of the case identification information (case number) is received is on the same side or on the opposite side. It is not only possible to tell whether a project is on the same side or the opposite side as viewed from the starting point project, but it is also possible to see from any project whether other projects are on the same side or the opposite side. In the example of FIG. 24, it can be seen that "6667/similar (buy)" and "5295/recommended (buy)" are on the same side, and "6667/similar (buy)" and "11223/recommended (sell)" are on the same side. is the opposite side. In the example of FIG. 24, this side information display function is realized by sorting display of the center column and the right column. However, the side information may be realized by, for example, color-coded display, presence/absence of shading, presence/absence of frame enclosure, etc., instead of display by sorting columns. Note that this function of displaying side information is different from the function of displaying recommendation/similarity distinguishing information in (2) above, since the recommended items of the recommended items are displayed on the same side.

(4) There is a function of displaying information on needs flags. Needs flag indicating whether to buy or sell, whether to buy or not, whether to be a worker or an employer, whether to be a man or a woman, and other paired information in the item data of each item (in the example of Fig. 24, whether to buy or sell) is assumed to be included. In the example of FIG. 24, they are indicated by characters "(sell)", "(buy)", and "(buy)". However, the information of the needs flag may be realized by, for example, color-coded display, presence/absence of hatching, presence/absence of frame enclosure, etc., instead of character display.

(5) Need flag information (in the example of FIG. 24, whether it is trading) and side information that distinguishes whether it is on the same side or on the opposite side (in the example of FIG. 24, distinction between input of similar projects and input of recommended projects) There is a function to display consistency information with the distribution display of the center column and right column according to ). In the example of FIG. 24, a normal display in which normal-colored characters (for example, black characters) are displayed on a normal background color (for example, white), a warning display indicating the occurrence of reversal by making a red background or red characters, etc., and a yellow background. Alternatively, there is a warning display that indicates that the display of consistency information does not work by using yellow letters or the like.

By performing search history display processing with these functions (1) to (5), the user can , it becomes easier to grasp what you have done so far, what you are doing now, and what you should do from now on, so you can perform efficient searches.

(Modified form of display of reference cases when registering new cases: Fig. 25)
In the above-described embodiment, the output means 44 is configured to receive a memo input (not shown) of the item identification information (item number) of the reference item. Sometimes, the case data of the reference case may be displayed.

In FIG. 25, a user (a customer (company) who registers a new item, or a person in charge of the service) is searching for related items displayed on the screen 200 (see FIG. 12), The recommended items and similar items displayed on the screen 300 (see FIG. 13) can be selected as reference items. The selection operation may be a selection input operation using check boxes or radio buttons provided corresponding to each related item, each recommended item, and each similar item. An operation of dragging and dropping a provided icon to the drop area may be used, and in short, it is sufficient to directly select an item displayed on the screen during a search on the screen. When the screen is switched during the search and the already selected item is displayed again, the check box, radio button, etc. of the item are displayed in a state in which the selection input operation has already been performed. In addition, the case number of the selected case is displayed in the peripheral part of the drop area (lower part in the example of FIG. 25). A drop area for canceling the selection may be provided, and the icon may be dragged and dropped there when canceling the selection.

Then, the case identification information (case number) of the selected reference case is stored in the reference case storage means 69 at the time of registration by the output means 44 in the same manner as in the memo input (not shown) of the above-described embodiment. Thereafter, when the user registers a new item, the new item registration means 45 displays a new item registration screen 500 as shown in FIG.

In FIG. 25, the new project registration screen 500 is provided with a self PR input section 510 , a detailed needs input section 520 , and a reference project display section 530 . The reference case display section 530 includes a case number display section 531, a needs flag display section 532, and a self-PR text data of the reference case for automatically copying and pasting into the self-PR input section 510 of the new case. A select button 533, a select button 534 for automatically copying and pasting the self-PR text data of the reference project to the detailed needs input section 520 of the new project, and inputting the detailed needs text data of the reference project to the self-PR of the new project. A selection button 535 for automatically copying and pasting to the section 510 and a selection button 536 for automatically copying and pasting the detailed needs text data of the reference project to the detailed needs input section 520 of the new project are provided. ing. The reason why the selection buttons 533 to 536 are provided so that they can be cross-moved is that there are projects whose needs flags are set to "buy and sell" (projects specifying both paired information). This is because the reference case selected by the user is not necessarily the case on the same side as the user himself (the person who performs the new registration).

The user can select and operate selection buttons 533 to 536 to copy and paste the text data of the reference case, and then edit the text data for his own case. Also, by performing the selection operation of the selection buttons 533 to 536 multiple times, multiple text data can be automatically copied and pasted into the self-PR input section 510 of the new project. In this case, the selected text data are added to the self-PR input section 510 of the new item one after another in the order of selection, instead of overwriting. In the example of FIG. 25, a selection button 533 for self-PR text data "XXXXXXXXX" of the reference project (project number=15542) and a selection button 535 for the text data "ZZZZZZZZ" of the detailed needs of the reference project (project number=8275) are displayed. are selected in this order, so that “XXXXXXXXX” and “ZZZZZZZZ” are pasted in self-PR input section 510 . Likewise, multiple text data can be automatically copied and pasted one after another to the detailed needs input section 520 of the new project. By performing normal copy and paste operations on the screen 500 instead of selecting the selection buttons 533 to 536, part or all of the text data of the reference case (selected copy range) can be newly created. It may be pasted into the self-PR input section 510 or the detailed needs input section 520 of the project.

In this way, if the project data (including self-promotion and needs details) of the reference project is displayed when registering a new project so that it can be used to edit the self-promotion and needs details of the user's own project, A user who registers a new project can easily input appropriate project data (including self-promotion and detailed needs) as data for matching.

(Modified form that can receive simultaneous input of multiple item identification information: FIG. 26)
In addition, the matching system of the present invention may be configured so that a plurality of items can be input at the same time, as shown in FIG. In this case, all of the multiple projects become the target projects, and a target project group is composed of the multiple target projects. Then, a recommended case group including a plurality of recommended cases is displayed for the target case group. This recommended project group is a set of recommended projects determined by an integrated matching process for the target project group. is not a simple collection, but depending on the ensemble matching algorithm, it may result in the same display state as such a simple collection. The purpose of performing the integrated matching process in this way is to introduce a recommended case group. business card exchanges, group meetings, etc., group matchmaking with multiple men (target project group) and multiple women (recommended project group), joint dining parties, etc., so one-on-one This is because it is not enough just to consider the matching of . On the other hand, similar projects are displayed from the viewpoint of improving search convenience and diversity. Similar projects are displayed for each target project. As a result, it is possible to perform a search by replacing at least some of the multiple target cases that constitute the target case group with respective similar cases.

In FIG. 26, screen 600 has an input section 610, a "Search" button 620 for executing a search, a "search history" display section 630, and an input section 640 for the number of items constituting a recommended item group. is provided. The title of the screen 600 is Business Matching, but it is not limited to this, and may be Group Coupling of Men and Women, for example.

In the input section 610, the item identification information (item number) of a plurality of target items forming the target item group is simultaneously input by comma separation, semicolon separation (example in FIG. 26), blank insertion, or the like. For example, if five project numbers are simultaneously input to the input unit 610 and "5" is input to the input unit 640, a recommended project group consisting of five projects is displayed for five projects (five companies or five people). . However, the number of target projects that make up the target project group and the number of recommended projects that make up the recommended project group do not necessarily have to match. A recommended item group may be displayed. Note that the matching process of displaying one recommended case group of 10 cases and the matching process of displaying two recommended case groups of 5 cases (group numbers = 1 and 2) for a target case group of 5 cases are: It's a different process.

In addition, the screen 600 of FIG. 26 includes a "target case" display area 650, a "recommended case group" display area 660, a "similar case list" display area 670, and the number of similar cases displayed per target case. An input section 671 is provided. The “target project” display area 650 displays the project data (including the project number, self PR, and need details) of each of the plurality of target projects forming the target project group.

At least one recommended case group is displayed in the "recommended case group" display portion 660, and a group number is assigned to each recommended case group. Project data (including project number, self-promotion, and details of needs) of each of the plurality of recommended projects constituting each recommended project group is displayed. The number of recommended item groups to be displayed is input by the system administrator and stored as setting data in the setting data storage unit 51, but may be a fixed number. Alternatively, the user may specify it on the screen 600 . It should be noted that the input to the input unit 640 is the number of items forming one recommended item group, so it is different from the number of recommended item groups referred to here. In addition, there is no overlap between a plurality of recommended cases forming a certain recommended case group and a plurality of recommended cases forming another recommended case group.

The "similar case list" display section 670 displays the case data (including the case number, self PR, and need details) of a plurality of similar cases for each of the plurality of target cases that constitute the target case group. The item number of the target item is also displayed so that it is possible to know which item is similar to the target item. The number of similar cases to be displayed for one target case is specified by the user by inputting it into the input unit 671, but it may also be entered by the system administrator as setting data and stored in the setting data storage means 51, or as a fixed number. good too.

Note that the method of determining the recommended case group is described in the above-mentioned Patent Document 1 (matching system developed by the applicant of the present application), so that algorithm can be adopted. This group matching algorithm determines the recommended project group so that the satisfaction of all the multiple target projects (companies, etc.) that make up the target project group is high, but it does not necessarily perform optimization processing in a strict sense. No need. In short, matching scores between one case and one case are obtained for all pair cases, and using those scores, within each group matching algorithm, recommend cases so that the total value of the score becomes large. Group decisions can be made.

For example, it is assumed that a recommended case group of 10 cases (Nc cases) is determined for a target case group of 5 cases (Nt cases) of A company, B company, C company, D company, and E company. In this case, a group of candidate projects consisting of (Na-Nt) projects obtained by excluding 5 projects (Nt projects) of companies A, B, C, D, and E from all registered projects (Na projects) think. Calculate the total value of 5 (Nt) matching scores between any 1 project in this candidate project group and 5 projects (Nt projects) of Company A, Company B, Company C, Company D, and Company E do. This total value is calculated for all (Na-Nt) projects in the candidate project group. Among the projects in the candidate project group, the projects with the largest total value, the second, third, . number (Nc) of recommended projects, and 10 projects that make up a group of recommended projects with a group number of 2 from the 11th ((Nc+1)th) to the 20th (2×Ncth) total value (Nc) recommended cases, and the total value from the 21st ((2×Nc+1)th) to the 30th (3×Ncth) cases constitutes the recommended case group with the group number=3. The number (Nc) of recommended cases is determined, and the matching process of . . . is executed.

As another group matching algorithm, for example, first, five projects (Nt projects) of company A, company B, company C, company D, and company E are selected from a candidate project group consisting of (Na-Nt) projects. 5 projects (Nt projects) with the highest matching score with each of the . In other words, the case with the highest matching score with company A, the case with the highest matching score with company B, and so on are selected. If there is no overlap among the five selected cases, all of the five cases are determined as recommended cases, and the remaining cases that should be decided as recommended cases are five cases. was decided as a recommended project, and the remaining projects are 6 projects. Next, five projects (Nt projects) with the second highest matching score with each of the five projects (Nt projects) of A company, B company, C company, D company, and E company are selected. That is, the case with the second highest matching score with company A, the case with the second highest matching score with company B, and so on are selected. If there is no overlap among the selected 5 projects, all of the 5 projects are additionally determined as recommended projects, and the remaining project to be determined as a recommended project is 1 project. Additional proposals are determined as recommended proposals, and the remaining proposals are two proposals. Subsequently, five projects (Nt projects) with the third highest matching score with each of the five projects (Nt projects) of A company, B company, C company, D company, and E company are selected. That is, the case with the third highest matching score with company A, the case with the third highest matching score with company B, and so on are selected. Since there are two projects remaining, the projects with the second highest matching score among the selected five projects are selected and additionally determined as recommended projects. Then, in the determination of the group number of recommended projects after group number = 2, the matching scores with each of the five projects (Nt projects) of Company A, Company B, Company C, Company D, and Company E are 4th, 5th, and 6th. , , and so on (Nt cases) are sequentially selected.

As shown in FIG. 26, when a configuration is adopted in which a plurality of case numbers can be simultaneously input, the similar cases for each of the plurality of target cases constituting the target case group are extracted by the similar case search means 43, Since the item data of these similar items are displayed in the "similar item list" display section 670, the item group to be introduced can be determined from various points of view. For this reason, even service personnel with little know-how can perform appropriate referral work, and it is possible to prevent the matching model from being updated by learning data including biased referral results.

That is, the matching means 42 does not update the topic model obtained by the topic estimation process or the matching model obtained by learning using the matching feature data. Since output is performed, the same recommended case group is always output for a certain target case group. However, when the display as shown in FIG. 26 is performed, at least one of the plurality of target cases that constitute the target case group is changed to a similar case to search or recommend. It is possible to change at least one item out of a plurality of recommended items that constitute the item group to an item similar to those recommended items and introduce the item.

As described above, the matching system and program of the present invention are suitable for use in, for example, business matching systems that combine companies.

10 Matching System 36 Topic Value Predicting Means 40 Input Receiving Means 41 Associative Search Means 41A Word Similarity Calculating Means Constituting Associative Searching Means 41B Relevance Calculating Means Constituting Associative Searching Means 42 Matching Means 42A For Matching Constituting Matching Means Characteristic data creation means 42B Matching score calculation means constituting matching means 43 Similar case search means 43A Case similarity calculation means constituting similar case search means 43B Similar case extraction means constituting similar case search means 44 Output means 45 New cases Registration means 53 Distributed representation model storage means 54 Item data storage means 56 Topic model storage means 60 Matching model storage means 68 Item similarity storage means 69 Reference item storage means at registration

Claims (9)

A matching system configured by a computer that executes matching processing that combines natural persons, legal persons, non-juristic bodies, or other cases,
Self-data consisting of text data describing the characteristics of each of a large number of projects, and needs data consisting of text data describing needs for matching partners, or obtained by combining these self-data and needs data Using the combined text data obtained as a large number of unique text data that describe information specific to each project, and performing topic estimation processing by soft clustering or neural language models, in the unique text data of each project topic model storage means for storing a topic value indicating the probability of occurrence of each topic in association with item identification information, and storing the probability of occurrence of each word in each topic obtained by executing the topic estimation process;
input receiving means for executing processing for receiving input of keywords or project identification information;
When the input receiving means receives the input of the item identification information, the item of the received item identification information is set as the item to be subjected to matching processing, and each item in the unique text data of each item stored in the topic model storage means is matched. Using the topic value that indicates the appearance probability of the topic, create matching feature data for each of a plurality of pairs consisting of the target case and each candidate case that is the matching partner of this target case, and create a plurality of created matching feature data By using each of the feature data as input data and using a matching model that has been pre-learned by two-class classification of whether or not to be selected and stored in a matching model storage means, classification processing by a supervised classification model is performed, a matching means for executing a process of calculating a score indicating the probability that each candidate project is selected with respect to the target project;
When the input receiving means receives the input of the project identification information, in conjunction with the processing by the matching means, the project between the target project calculated using the unique text data of each project and other projects Similar case search means for executing a process of extracting cases similar to the target case in descending order of similarity;
final similarity is the general similarity between the keyword and other words calculated using the distributed representation of the word stored in the distributed representation model storage means when the input of the keyword is received by the input receiving means; Or, using this general similarity and the unique similarity between the keyword and other words calculated using the unique text data of each case, calculate the final similarity, and obtain Related words of the keyword are determined in descending order of the final similarity obtained, and the number of appearances of each word in which the keyword and the related words appear in the specific text data of each project is obtained. an associative search means for executing a process of determining related cases in descending order of the degree of relevance of each case to the keyword calculated using the number of appearances for each word and the final similarity of the related words;
When an input of a keyword is received by the input receiving means, the related item determined by the associative search means or the related word in addition to the related item is added to the keyword or item identification information by the input receiving means. displayed on the same screen as the input unit, and when the input receiving means receives the input of the project identification information, the candidate project with the high score calculated by the matching means is the same as the input unit as the recommended project. and output means for displaying the similar cases extracted by the similar case search means on the same screen as the input unit. The unique text data of the new issue using the unique text data of the new issue not used in the topic estimation process and the appearance probability of each word in each topic stored in the topic model storage means a topic value prediction means for executing a process of predicting a topic value indicating the appearance probability of each topic in
The matching means is
A topic value indicating the appearance probability of each topic in the specific text data of the new project obtained by the topic value prediction means, and a candidate project to be matched with the new project stored in the topic model storage means. creating matching feature data for a plurality of pairs of the new case and each candidate case using the topic value indicating the appearance probability of each topic in the unique text data, and creating a plurality of created matching features; Each of the data is used as input data, and using a matching model that has been pre-learned by two-class classification of whether or not it is selected and stored in the matching model storage means, classification processing by a supervised classification model is performed, and executing a process of outputting a score indicating the probability that each of the candidate projects is selected for the new project,
The similar case search means is
a process for extracting similar cases including the degree of similarity between the new case and the other cases calculated using the unique text data of the new case and each other case,
The associative search means is
Obtaining the number of occurrences of each word in which each of the keywords and the related words appears in the unique text data of each item including the new item, and calculating the final similarity of the obtained number of occurrences of each word and the related words. 2. The matching system according to claim 1, wherein a process for determining related cases is executed in descending order of the degree of relevance of each case to the keyword calculated using the keyword. The similar case search means includes:
A topic distribution vector consisting of a topic value indicating the appearance probability of each topic in the unique text data of each matter stored in the topic model storage means, and an average of the distributed representation of each word constituting the unique text data of each matter. and a word importance vector consisting of the TFIDF value or other word importance index value of each word for the unique text data of each project. 3. The matching system according to claim 1 or 2, wherein a process of extracting similar cases to the target case is executed in descending order of the degree of similarity between the case and other cases. . Search history storage means for storing the keyword or project identification information received by the input reception means and log information including the input order thereof,
The input receiving means is
Using the log information stored in the search history storage means, in addition to indicating the input type and input order of whether the input of the keyword or the item identification information was accepted,
Recommendation/similarity distinguishing information for distinguishing whether input of item identification information of the recommended item for the target item has been received or input of item identification information of the similar item of the target item has been received;
Whether the input of the item identification information of the recommended items and the item identification information of the similar items is accepted repeatedly, and whether the item whose item identification information is input is the item on the matching requesting side or the matching partner's side. From this point of view, the recommended case is a case on the opposite side of the case for which the input of the case identification information is received, and the similar case is a case on the same side as the case for which the input of the case identification information is received. With respect to the above-mentioned recommended projects for the projects, by making the projects on the same side as the projects for which the input of the project identification information is received, it is possible to determine whether the projects for which the input of the project identification information is on the same side or on the opposite side. Execute a search history display process including at least one of the side information that distinguishes the
Alternatively, if the project data of each project includes a needs flag indicating whether it is a sale/buy, whether it is a worker/employer, whether it is a male/female, or other paired information, the recommendation/recommendation 4. The configuration according to any one of claims 1 to 3, wherein a search history display process including at least one of the similarity distinguishing information and the side information and including the needs flag is executed. Matching system described in. a new matter registration means for accepting input of matter data of a new matter, assigning matter identification information, and storing the received matter data in the matter data storage means in association with the given matter identification information;
The output means is
executing a process of storing the item identification information of the related item, the recommended item, or the similar item selected as a reference item when registering a new item or the item data thereof in a reference item storage means upon registration,
The new matter registration means is
obtaining the item data of the reference item from the item data storage means using the item identification information of the reference item stored in the reference item storage means at the time of registration, or acquiring the item data of the reference item stored in the reference item storage means at the time of registration; and acquiring the project data of the reference project that is stored in the system, and displaying the acquired project data of the reference project on the same screen as the input unit for the project data of the new project. The matching system according to any one of claims 1 to 4. The associative search means is
configured to execute a process of calculating the final similarity using the general similarity and the specific similarity, and
by the number of occurrences of the general similar word in a set P of all cases in which both the keyword and the general similar word selected according to the level of general similarity are included in the specific text data, or Alternatively, an evaluation value Ps of the set P based on the number of appearances of the keyword and the number of appearances of the general similar word in the set P,
Using the evaluation value Qs of the set Q based on the number of occurrences of the general similar word in the set Q of each case in which the keyword is not included in the unique text data and the general similar word is included,
6. The matching system according to any one of claims 1 to 5, wherein the process of calculating the unique similarity is executed by Ps/(Ps+Qs). The input receiving means is
After the input of the keyword and the input of the item identification information are accepted by the same input unit, the information input to the input unit is configured to determine whether the information is the keyword or the item identification information. The matching system according to any one of claims 1 to 6, characterized in that The input receiving means is
It is configured to be able to receive simultaneous input of multiple project identification information,
The matching means is
When simultaneous input of a plurality of item identification information is accepted by the input receiving means, each item of the received plurality of item identification information is grouped into one item group to be subjected to matching processing and stored in the topic model storage means. Using the topic value that indicates the appearance probability of each topic in the unique text data of each project, each of the plurality of target projects that make up the target project group and each matching partner of each of these multiple target projects Matching feature data is created for each of a plurality of pairs consisting of candidate cases, and each of the created plurality of matching feature data is used as input data, and is learned in advance by two-class classification of whether or not to be selected. Accuracy that each candidate case is selected for each of a plurality of target cases that constitute the target case group by performing classification processing by a supervised classification model using the matching model stored in the matching model storage means. Calculate a score that indicates
Selecting a plurality of candidate projects in which the total value of each score between each of the plurality of target projects constituting the target project group and each of the plurality of candidate projects to be matching partners is large, and selecting the plurality of selected candidates a process for determining a project as a plurality of recommended projects constituting a recommended project group,
The similar case search means is
when simultaneous input of a plurality of item identification information is received by the input receiving means, the unique text data of each item for each of the plurality of target items forming the target item group, together with the processing by the matching means; is configured to execute a process of extracting similar cases of the target case in descending order of the case similarity between the target case and other cases calculated using
The output means is
displaying a plurality of recommended items constituting the recommended item group determined by the matching unit on the same screen as the input unit when the input receiving unit receives simultaneous input of a plurality of items of item identification information; In addition, a process of displaying the similar cases for each of the plurality of target cases constituting the target case group extracted by the similar case search means on the same screen as the input unit is executed. The matching system according to any one of claims 1 to 7, characterized in that: A program for causing a computer to function as the matching system according to any one of claims 1 to 8.

Images