JP6941802B1 - Search system, search method and search program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the burden on a user in searching for a business partner and effectively list candidate companies suitable as a business partner. SOLUTION: A management server 20 of a search system 1 stores a feature word database 223 that stores a plurality of feature words representing the characteristics of a company together with their scores for each company, and a distributed expression of the feature words obtained by performing machine learning. It is equipped with a learning model database 224 that stores each feature word as a learning model. When the management server 20 acquires a keyword input from an external input device, the management server 20 searches the feature word database 223 for a company having a feature word corresponding to the keyword, and a distributed expression based on the feature word of the company and a feature corresponding to the keyword. Calculates similarity to word-based distributed expressions and generates a list of companies including feature words, scores and similarity. [Selection diagram] Fig. 3

Description

The present invention relates to a search system, a search method, and a search program that generate a company list for the purpose of searching for a business partner.

In recent years, product development in companies often utilizes not only their own technology but also the excellent technology of external companies. For example, many large and medium-sized companies have small and medium-sized venture companies with excellent technology as business partners.
When searching the Internet for a new business partner with whom there has been no transaction, the search user enters a keyword on the Web browser displayed on the screen of a personal computer or the like, and the companies that hit the keyword are listed as candidates for the business partner. .. Most of the keywords used in this case are words related to products and materials such as products and their characteristics, but since they are selected based on the experience and knowledge of the search user, the difference in user skills and experience is the search result. It often appears in. Therefore, there is known a technique of extracting a keyword technique related to an input keyword from company information, calculating the similarity between the two, and then searching for a company (see, for example, Patent Document 1).

Japanese Unexamined Patent Publication No. 2019-133367

However, the companies listed by the conventional keyword search through trial and error of the user are not given superiority or inferiority as to whether or not they are suitable as business partners at the time of listing. Therefore, if the user does not confirm and judge the company information of all the listed candidate companies, there is a possibility that a more appropriate candidate company will be selected and leaked, which requires labor and human cost after listing. ..

The present invention has been made in consideration of such circumstances, and is a search system, a search method, and a search program that can reduce the burden on the user in searching for a business partner and effectively list candidate companies suitable as business partners. The purpose is to provide.

In order to solve the above problems, the present invention provides a search system for generating a list of companies for the purpose of business partners search, a plurality of feature words representing the characteristics of the company together with the score based on the current frequency output of the characteristic words Learning to memorize the characteristic word storage means to be memorized for each company and the distributed expression of the characteristic word obtained by performing machine learning to predict one characteristic word from the plurality of characteristic words as a learning model. A model storage means, an acquisition means for acquiring a keyword input from an input device, a search means for searching a company having a feature word corresponding to the keyword from the feature word storage means, and the search means searched by the search means. A similarity calculation means for calculating the similarity between a distributed expression based on a plurality of characteristic words possessed by a company and a distributed expression based on a characteristic word corresponding to the keyword, and a plurality of features possessed by the company searched by the search means. It is characterized by comprising a list generation means for generating a company list including a word, a score of the characteristic word, and a similarity.

Further, in the search system according to the present invention, the similarity calculation means corresponds to a synthetic distributed expression obtained by synthesizing each distributed expression of a plurality of characteristic words of the company searched by the search means and the keyword. It is characterized by calculating the degree of similarity with the distributed expression of the feature word.

Further, in the search system according to the present invention, the similarity calculation means calculates a feature word similar to the feature word corresponding to the keyword based on the learning model, and the feature word corresponding to the keyword and the feature. It is characterized in that the similarity between a first vector containing a plurality of feature words similar to words and their scores and a second vector containing the feature words owned by each company and their scores is calculated.

Furthermore, in the search system according to the present invention, the similarity calculation means is based on the sum of the products of the scores between the first vector and the same feature words included in the second vector. Is characterized by calculating.

Furthermore, in the search system according to the present invention, the similarity calculation means has a feature in which the score of each feature word is higher than that of the distributed expression based on a plurality of feature words of the company searched by the search means. A word is a distributed expression in which a word is weighted more than a feature word having a lower score, and the similarity with a distributed expression based on the feature word corresponding to the keyword is calculated.

Furthermore, in order to solve the above problems, the present invention is a search method in a search server that generates a company list for the purpose of searching for a business partner, and includes an acquisition step of acquiring a keyword input from an input device, and the above-mentioned. the company with characteristic word corresponding to the keyword, the search step of searching a plurality of feature words representing the characteristics of該企industry from the feature word database stored in each company along with the score based on the current frequency output of the feature words, the plurality Searched in the search process by referring to a learning model database that stores the distributed expression of the feature word obtained by performing machine learning that predicts one feature word from the feature words of the above for each feature word and uses it as a learning model. A similarity calculation process for calculating the similarity between a distributed expression based on a plurality of characteristic words possessed by the company and a distributed expression based on the characteristic words corresponding to the keyword, and a plurality of the distributed expressions possessed by the company searched in the search process. It is characterized by having a list generation step of generating a company list including the feature word, the score and the similarity of the feature word.

Furthermore, in order to solve the above problems, the computer-readable search program according to the present invention acquires a keyword input from an input device to a search server that generates a company list for the purpose of searching for a business partner. A search procedure for searching a company having a characteristic word corresponding to the keyword from a characteristic word database that stores a plurality of characteristic words representing the characteristics of the company together with a score based on the frequency of appearance of the characteristic word for each company. In the search procedure, the search procedure refers to a learning model database that stores the distributed expressions of the feature words obtained by performing machine learning that predicts one feature word from the plurality of feature words for each feature word and uses the learning model as a learning model. The similarity calculation procedure for calculating the similarity between the distributed expression based on a plurality of characteristic words of the searched company and the distributed expression based on the characteristic word corresponding to the keyword, and the company searched by the search procedure It is characterized in that a list generation procedure for generating a company list including a plurality of feature words, scores and similarities of the feature words is executed.

According to the present invention, it is possible to reduce the burden on the user in searching for a business partner and effectively list candidate companies suitable as business partners. For example, a user-friendly list of companies that includes a score that is an indicator of how well each feature represents the characteristics of the company, and a similarity that is an indicator of how relevant the company is to the input keywords. Can be provided to the user.

It is a block diagram of the search system 1 which concerns on one Embodiment of this invention. It is a block diagram which shows the structure of the user terminal 10 in the search system 1 which concerns on one Embodiment of this invention. It is a block diagram which shows the structure of the management server 20 in the search system 1 which concerns on one Embodiment of this invention. It is a figure which shows the data structure example of the characteristic word database 223 which concerns on one Embodiment of this invention. It is a figure which shows the data structure example of the learning model database 224 which concerns on one Embodiment of this invention. It is a flowchart for demonstrating the whole process in the management server 20 of the search system 1 which concerns on one Embodiment of this invention. It is a flowchart for demonstrating the detail of the generation processing (step S12) of the new keyword candidate (feature word) in the management server 20 of the search system 1 which concerns on one Embodiment of this invention. It is a figure which shows the display example on the user terminal 10 of the candidate company list which concerns on one Embodiment of this invention. It is a flowchart for demonstrating the detail of the similarity calculation process (step S15) in the management server 20 of the search system 1 which concerns on one Embodiment of this invention. It is a figure which shows the display example on the user terminal 10 of the candidate company list including the degree of similarity which concerns on one Embodiment of this invention.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.
FIG. 1 is a configuration diagram of a search system 1 according to an embodiment of the present invention. As shown in FIG. 1, the search system 1 according to the present embodiment is composed of a plurality of user terminals 10 and a management server 20, and is connected to each other so as to be communicable by a wired or wireless network 30 such as the Internet. Below, the above search system 1 is introduced to a company that provides a company matching service including the search for business partner companies (hereinafter referred to as "management company"), and is detailed below by a SaaS (Software as a Service) type cloud service. The search for business partner candidate companies and various other processes described above shall be performed.

The user terminal 10 is, for example, a personal computer of a manager of a department that searches for a business partner of a management company or a person in charge belonging to the department. When the management company receives a request to search for a business partner from a client company (hereinafter referred to as "ordering company"), the person in charge of the management company conducts interviews with the person in charge of the ordering company about the technology expected of the business partner company. , Consider the appropriate keywords for your search. Then, the user of the user terminal 10 (the person in charge of the management company or another person) searches for the candidate company using the examined keywords and creates a candidate company list of the business partner. Specifically, the user of the user terminal 10 accesses the website managed by the management server 20 and inputs a keyword. The management server 20 searches for candidate companies based on the input keywords and generates a list of candidate companies for business partners. At this time, the list can be displayed on the website accessed from the user terminal 10.

Here, the user terminal 10 includes not only the management company, but also the ordering company that receives the search service of the business partner, the ordering company that the ordering company wishes to interview based on the candidate company list, and the like (hereinafter, "candidate company"). It may be introduced in). In this case, the person in charge of the ordering company uploads the company information of the company and the information related to the search for the business partner (details of the search for the business partner, questions to the candidate company for ordering, etc.) to the management server 20 from each user terminal 10. , You can browse the list of candidate companies searched by the users of the management company. In addition, the person in charge of the ordering company can upload the company information of the company and the information related to the search for the business partner (appeal statement to the ordering company, etc., answer to the question from the ordering company, etc.) to the management server 20 from the user terminal 10.

FIG. 2 is a block diagram showing a configuration of a user terminal 10 in the search system 1 according to the embodiment of the present invention. As shown in FIG. 2, the user terminal 10 displays data such as an input unit 11 such as a keyboard or touch panel on which the user inputs keywords and commands, and a list of candidate companies generated by the input keywords and the management server 20. It includes a display unit 12 such as a monitor, a processing unit 13 that performs various processes, and a communication unit 14 that performs communication processing with a management server 20 and other devices (not shown) via a network 30. The processing unit 13 is composed of a CPU, a ROM storing a program, a RAM, and the like, and performs various processes according to the program stored in the ROM. In the present embodiment, the user terminal 10 will be described as a personal computer, but other than that, a communication terminal such as a smartphone or a tablet may be used.

FIG. 3 is a block diagram showing the configuration of the management server 20 in the search system 1 according to the embodiment of the present invention. As shown in FIG. 3, the management server 20 includes a communication unit 21 that performs communication processing with a user terminal 10 and other external devices (not shown) via the network 30, and a storage unit that stores information such as programs and various data. 22, Search unit 23 that searches for business partner candidate companies based on the keywords entered on the user terminal 10, Similarity calculation unit 24 that calculates the similarity between the entered keywords and the characteristics of the searched companies, Calculation It is provided with a list generation unit 25 that generates a list of candidate companies for business partners using the similarities.

In addition, the management server 20 acquires information on the technology of the candidate company and technical tag information for indicating the characteristics of the technology of the company by tagging from the user terminal 10 and other external devices (not shown), and obtains a database. A database construction unit 26 for storing in the storage unit 22, a feature word generation unit 27 for generating feature words indicating the characteristics of the company, and a learning model generation unit 28 for generating a learning model for the feature words are provided. In addition, the storage unit 22 stores a company database 221 that stores information related to the technology of the candidate company, a technology tag database 222 that stores technology tag information, a feature word database 223 that stores feature words, and a learning model that stores a learning model. It has a database 224.

In the present embodiment, the management server 20 realizes the processing function of each of the above-mentioned units (communication unit 21 to learning model generation unit 28) on one server, but realizes this function on a plurality of servers. You may do so. Further, each server may be configured by a single computer, or may be configured by a plurality of physically different computers or the like.

The database construction unit 26 is a user terminal 10 or other external device (not shown) of an ordering company or an order candidate company selectively extracted from a company's homepage or a paid / free database published on the Internet. When the company information, etc. (hereinafter referred to as "public company information") related to the company outline, business outline, owned technology, etc. is acquired, it is stored in the storage unit 22 in association with the company ID given to each company, so that the company database 221 Is newly constructed or the stored information is updated. For example, in the public company information, the company name (company name), the location (address) of the head office, the explanation of the company outline, the explanation of the business outline, the explanation of the owned technology, the address (URL) of the company homepage, etc. are disclosed. Contains information. The gazetteed company information also includes a sentence in which the person in charge of the management company drafts a company outline, a business outline, or an explanation of the owned technology based on the public information.

In addition, the database construction unit 26 uses the user terminal 10 or other external device (not shown) to provide entry information (information on candidacy from the order candidate company and its candidacy) regarding past and present business partner search project projects in the management company. (Additional information such as appeal statements, etc.), information on questions and answers asked between the ordering company and the candidate company in the process of searching for past and present business partners, and the technology possessed by the candidate company. When information about technology (seeds technology) that can be provided to external companies (hereinafter referred to as "private company information") is acquired, the company database 221 is newly created by storing it in the storage unit 22 in association with the company ID. Build or update stored information.

Further, the database construction unit 26 uses tag information (hereinafter, not shown) regarding international standards and domestic standards regarding products and services such as ISO (registered trademark) owned by the company from the user terminal 10 or other external device (not shown). , "Standard tag"), tag information on equipment / equipment for product processing / production (hereinafter referred to as "equipment tag"), award history received from national / local public organizations and various industry organizations When the technical tag information including the tag information related to the above (hereinafter referred to as "award history tag") is acquired, the technical tag database (technical tag dictionary) 222 is newly constructed or the stored information is updated by storing the technical tag information in the storage unit 22. ..

The acquisition of public company information and private company information by the database construction unit 26, the storage / update process in the company database 221, the acquisition of technical tag information, and the storage / update process in the technical tag database 222 are performed by a keyword search described later. And may be performed together with the processing related to the generation of the candidate company list, or may be performed independently of these processing. For example, the database construction unit 26 acquires public company information, private company information, and technical tag information from an external device or the like on a regular basis or at an arbitrary timing, such as once every few months, and obtains the company database 221 and the technical tag. The stored information in the database 222 may be updated. The public company information and private company information in the company database 221 and the technical tag information in the technical tag database 222 are text data in order to facilitate processing in the feature word generation process and the learning model generation process described later. It is remembered.

The feature word generation unit 27 acquires public company information and private company information for each company from the company database 221 to generate combined text data, and obtains the combined text data through morphological analysis and unnecessary word removal processing. The technically related terms are generated as feature words, and are stored in the storage unit 22 in association with the company ID to newly store or update the information in the feature word database 223. FIG. 4 is a diagram showing an example of a data structure of the feature word database 223 according to the embodiment of the present invention. In the present embodiment, the feature word database 223 stores the feature words generated by the feature word generation unit 27 in association with the score for each company. Further, as the number of characteristic words possessed by each company, the characteristic words having a higher score are stored together with the score for each company, but in the present embodiment, 500 characteristic words are stored together with the score as an example.

Here, the score of the feature word will be described. In the present embodiment, the feature word score is determined by the ratio of the appearance frequency of the feature words generated from the combined text data in each company to the appearance frequency of each feature word in all the text data stored in the company database 221. It is indicated by a numerical value multiplied by a coefficient. In this embodiment, in order to effectively search for business partners especially in the manufacturing industry, many words peculiar to the manufacturing industry are stored as characteristic words, and the search system includes a dictionary for the manufacturing industry. There is.

The learning model generation unit 28 describes characteristic words contained in public literature information (hereinafter referred to as “technical documents”) related to technology such as patent documents, corporate research and development reports, and technology-related news for each manufacturing company. Deep learning is performed to predict one feature word from a plurality of feature words in the above, a distributed expression in which each feature word is expressed by a high-dimensional vector is generated, and the feature word of the distributed expression is stored in the learning model database 224 as a learning model. do. FIG. 5 is a diagram showing an example of a data structure of the learning model database 224 according to the embodiment of the present invention. In the present embodiment, in the learning model database 224, the distributed expression of each feature word generated by the learning model generation unit 28 based on the feature words stored in the feature word database 223 has a vector dimension for each feature word. It is shown numerically in order and stored. In this embodiment, as will be described later, since each feature word is constructed by a distributed representation of a high-dimensional vector (specifically, a 500-dimensional vector) having at least several hundred dimensions or more, 500 numbers are stored. However, the number of dimensions of the vector is not limited to this, and the system can be designed so that it can be changed as appropriate. The details of the generation process of the learning model database 224 will be described later.

Hereinafter, a list generation process of business partner candidate companies by keyword search on the management server 20 according to the embodiment of the present invention will be described with reference to the drawings. In the following embodiment, the search system 1 is introduced in the management company that provides the company matching service, and the user of the management company uses the in-house user terminal 10 in response to the order of the company search service from the ordering company. A series of search processes in the management server 20 that performs a keyword search and generates a list of candidate companies will be described.

FIG. 6 is a flowchart for explaining the overall processing in the management server 20 of the search system 1 according to the embodiment of the present invention. In the search system 1, the user of the user terminal 10 performs new construction of the company database 221, the technical tag database 222, the feature word database 223, and the learning model database 224 in the storage unit 22, and the storage information update process (step S11). It is possible to perform the keyword search process at the same time as, or before and after the keyword search process, respectively, but in the following embodiments, in order to facilitate the explanation, the keyword search process is performed prior to the keyword search process by the user of the user terminal 10. explain.

[Construction / update of corporate database 221]
Prior to the keyword search by the user terminal 10, the database construction unit 26 of the management server 20 acquires the public company information and the private company information of the candidate company from the user terminal 10 or other external device (not shown), and the like. Information is newly stored in the company database 221 for each company or the existing information is updated.

In the present embodiment, the public company information is obtained by acquiring the technical information of each company in text data format from the homepage, press release, news article, etc. of each company accessible via the Internet. When the publicly available company information is data such as HTML format, it can be converted into text data by removing unnecessary words such as headers and footers by a known method or the like. Further, there is no particular limitation on the acquisition timing, and the data is uploaded to the management server 20 from the administrator's user terminal 10 or other external device (not shown) at all times, periodically, or at any timing, and is uploaded to the database construction unit 26. Updates the data newly stored or stored in the corporate database 221.

Further, in the present embodiment, the person in charge of the candidate company accesses the management server 20 from each user terminal 10 and obtains the data of the private company information during the period determined for each project of the business partner search. By inputting, it is uploaded to the management server 20. For example, after the project information of the business partner search project is presented to the candidate company, a period of one month is set, and the answer from the order candidate company regarding the question from the ordering company only within that period (that is, private company information). ) Can be accepted to enable data entry. In addition to this, the period may be the entire period in which each business partner search project is being carried out, or it may be possible to input at any time without setting a special period, and the administrator can enter it at any time. It may be possible to adjust whether or not input is possible. This process is controlled by the database construction unit 26 of the management server 20, and when the private company information input by the user terminal 10 is uploaded to the management server 20, it is newly stored in the company database 221 or the existing information is updated. Will be done. Here, the private company information to be input includes, for example, the company's original information such as the latest technology of the candidate company and the statement of appeal of the technical ability to the ordering company, and the answer information to the question from the ordering company. ..

[Construction / update of technical tag database 222]
The database construction unit 26 of the management server 20 further provides the user terminal 10 or other technical tag information that is the basis for tagging potential technical features and the like possessed by each company prior to the keyword search by the user terminal 10. It is acquired from an external device (not shown) and newly stored or updated in the technical tag database 222. In the present embodiment, the technical tag information is acquired at the same time as or before and after the construction of the corporate database 221 to construct the technical tag database 222. As described above, the technical tag information includes a standard tag, an equipment tag, and an award history tag.

By making it possible to display the information related to each company ID and the above technology tag in association with each other as the attached information of the generated candidate company list, the user can use the international standard owned by the company to narrow down the candidate companies. It is possible to compare the potential technological capabilities of candidate companies at a glance by referring to the items of technology tags including equipment / equipment or award history. For example, a company whose standard tag item displays various international standards and awards is a company that is considered to be paying attention to the quality of products and services at a glance, and other organizations. It can be imagined that the company has high technical capabilities so that it can be commended from the above, and companies that display various equipment / equipment names in the equipment tag item can make prototypes or mass-produce them just by looking at the equipment / equipment names. It is possible to infer whether the company is capable of dealing with such issues. Further, for example, when a business partner company is required to own a clean room, "clean room" is included as a feature word, but when the score is low, the feature is displayed on the screen of the user terminal 10 when it is listed. Words may not be displayed. However, even in such a case, the user can confirm at a glance that he / she owns a clean room by referring to the equipment tag. In addition, since the international standard, award history, etc. are displayed, the user can infer the production system and reliability of the company. Moreover, even if the user who searches does not have specialized knowledge, the characteristics of the company can be easily confirmed by just looking at the item of the technical tag.

[Construction / update of feature word database 223]
The feature word generation unit 27 of the management server 20 generates feature words and their scores for each company from the company information (text data) of the company database 221 and newly stores them in the feature word database 223 or updates existing information. do.

Here, the details of the feature word generation process in the feature word generation unit 27 will be described. In this embodiment, a feature word is generated from the company information or the like stored in the company database 221 prior to the processing such as the keyword search described later, and is stored in the feature word database 223. Although this process may be performed at the time of keyword search described later, in this embodiment, this process is a process independent of the keyword search process, and the feature words are stored in the feature word database 223 in advance. In addition, the feature words stored in the feature word database 223 are sequentially updated independently.

The feature word generation unit 27 first calls the company information (public company information and private company information) of the text data from the company database 221. Next, the feature word generation unit 27 removes unnecessary words from the called company information and extracts the feature words. Specifically, the feature word generation unit 27 uses a technical dictionary (not shown) that summarizes technical terms, an attribute dictionary (not shown) that summarizes grammatical attributes of the word, and the like to make text into a morpheme. Perform morphological analysis to decompose and determine, and extract technical terms from the results. At this time, synonyms and fluctuation words (for example, "motor" and "motor") should be combined as one word. In removing unnecessary words, an unnecessary word dictionary (not shown) is prepared, and particles, non-technical terms (nouns), etc. are removed as unnecessary words from the morphemeized text. The word after removing unnecessary words is assumed to be a technical term, and the feature amount (score described later) is calculated. In addition, the feature word generation unit 27 is an abstract word (for example,) that is commonly used and generalized as a word representing a technical feature in business partner search with respect to the extracted word (technical term). , "Device" of "○○ device") is removed as an abstract word, and the remaining word (for example, "○○" part of "○○ device") is stored in the feature word database 223 as a feature word (newly). Update memory or information).

In the present embodiment, each feature word is given a feature amount (scalar amount) called a score as an index indicating the feature of the text (feature word). In this embodiment, a TF-IDF (Term Frequency --Inverse Document Frequency) value is adopted as this index. The TF-IDF value is the product of the TF value and the IDF value, the TF value is the frequency of appearance of each word in the text data of the company, and the IDF value is the text data of all companies including each word for all companies. It is the logarithmic value of the reciprocal of the ratio of the number of companies. This score indicates the importance of the words contained in the document, and can distinguish the characteristics of the document. Specifically, it can be understood that "words that appear many times in a document and words that do not appear in many documents are important words that characterize the document." Here, if the words are rearranged in descending order of TF-IDF value, words other than technical words are also ranked high, so prepare a technical word dictionary as described above and set the factor only for the words included in the dictionary. Multiply and store in the feature word database 223.
The score index and the number of dimensions are not limited to this, and may be appropriately changed depending on the nature and purpose of the system to be implemented.

[Construction / update of learning model database 224]
The learning model generation unit 28 of the management server 20 performs deep learning to predict the feature words contained in the technical document related to the manufacturing industry from the surrounding feature words, and constructs each feature word by the distributed representation of the high-dimensional vector. A model (learned model) is generated, and the learning model database 224 is newly stored or the stored information is updated for each feature word.

Here, as a distributed expression of a high-dimensional vector of a feature word, a word distributed expression in which characters / words are embedded in a vector space and regarded as one point in the space can be used. In this embodiment, the word vector space is expressed by the learned embedded space using Word2Vec. The vector space representation is not limited to Word2Vec, and other vector space models such as Doc2Vec and TF-IDF may be used.

Further, deep learning in the learning model generation unit 28 is performed using a neural network having a multi-layer structure of an input layer, an intermediate layer, and an output layer, and is performed for each word in a sentence (referred to as a "central word"). Is given a vector in which the words around the central word (called "peripheral words") are converted into one-hot expressions, and the output layer is given a vector in which the central words are converted into one-hot expressions. Learn to predict the central word from. In this embodiment, for example, learning is performed by using the CBOW (Continuous Bag of Words) method, which is a supervised learning method for predicting a central word from peripheral words, giving peripheral words as input, and outputting the prediction of the central word. To learn what kind of words are likely to appear around a certain word through learning. After learning, a distributed expression of each word can be obtained. Here, the distributed representation is a weight matrix to the middle layer. The dimension of the intermediate layer is generally selected from 100 to 1000 dimensions, but in this embodiment, 500 dimensions are adopted.

As described above, in the search system 1 according to the present embodiment, learning is performed using public document information on technology such as patent documents, corporate research and development reports, and technology-related news as technical documents. On the other hand, documents such as general news articles that are likely to be misused with technical terms, and technical terms and expressions that only experts in the field can understand, such as academic papers, are included. Documents are excluded from the technical documents to be studied. In other words, by limiting the technical documents to be learned as described above so that the system is suitable for searching for business partners in the manufacturing industry, which is the so-called "manufacturing" field, optimal machine learning can be performed. There is.

The learning model generation process may be performed immediately before the similarity calculation process described later, but in the present embodiment, this process is a process independent of the similarity calculation process, and a learning model is generated and learned in advance. It is stored in the model database 224, and the learning model stored in the learning model database 224 is updated periodically (for example, once a quarter).

Further, the system may include a plurality of trained models so that the learning model database can be selectively used according to the user's purpose. By doing so, for example, it is possible to deal with the difference in nuance of the characteristic word "motor" used in the automobile industry and the material / metal industry. In addition, the manufacturing industry is further classified into several detailed industries and a separate database is set up for each detailed industry, or a database for multiple industries including the manufacturing industry is set up, depending on the content of the business partner search. The database to be used may be appropriately selected.

[Keyword candidate (characteristic word) presentation processing]
When listing business partner candidate companies based on keyword search, it is usually necessary to enter multiple different keywords at the same time or sequentially to narrow down the search results so that appropriate candidate companies are listed. appear. Therefore, in the present embodiment, the user who lists the business partner candidate companies inputs the first keyword on the user terminal 10, and the text data matching the keyword is stored in the company database 221 as the company information. List the companies that are included or whose feature words corresponding to the keyword are held in the feature word database 223 (step S12). At this time, the management server 20 generates a list of candidate companies based on the input keywords, and the user narrows down the search for the candidate companies by including a plurality of characteristic words of the candidate companies in the list. The keyword candidate can be suggested to the user via the user terminal 10. FIG. 7 is a flowchart for explaining the details of the generation process (step S12) of new keyword candidates (feature words) in the management server 20 of the search system 1 according to the embodiment of the present invention.

The user of the search system 1 inputs a predetermined keyword using the user terminal 10 and transmits the information to the management server 20 as a search instruction. In the present embodiment, first, a case where "resin molding" is input as the first keyword will be given as an example. The management server 20 acquires the keyword “resin molding” input by the user terminal 10 via the communication unit 21 (step S121).

Next, the search unit 23 searches for a company having the same characteristic word "resin molding" as the acquired keyword "resin molding" from the characteristic word database 223 in the storage unit 22, and searches for the company ID of the hit company. Acquire (step S122). For example, in the present embodiment, the company IDs "5398, 36588, 34589, ..." Of a plurality of companies having the characteristic word "resin molding" shown in FIG. 4 are hit, and the characteristic word "resin molding" is possessed. The company ID of the company that does not hit does not hit.

Next, the list generation unit 25 generates a candidate company list including the feature words having a high score among the feature words owned by the company with the company ID hit by the search unit 23 and the score (step S123). The candidate company list generated by the list generation unit 25 is displayed on the screen of the user terminal 10 that has accessed the website managed by the management server 20. FIG. 8 is a diagram showing an example of displaying the candidate company list according to the embodiment of the present invention on the user terminal 10. As shown in FIG. 8, on the user terminal 10, in addition to the input keyword "resin molding", characteristic words that can be other keywords such as "liquid crystal", "sheet metal", and "infrared ray" are displayed together with the score. .. Then, the user can select another feature word "prototype" as the second keyword for searching by multiplying the feature word "resin molding". When presenting the company list generated in step S123 (when displaying on the screen of the user terminal 10), in addition to the company ID and feature words, the list generation unit 25 is used for the convenience of the user. The database 221 may be referred to, and the company name, address, URL, business outline, etc. associated with the company ID may be included in the company list.

At this time, another characteristic word (“prototype” in the above example) displayed as the first keyword (“resin molding” in the above example) may or may not be similar. .. The user of the user terminal 10 may appropriately select the second keyword from the feature words displayed according to the search purpose such as what kind of viewpoints to search from, and at that time, the keywords are similar to each other. It does not consider whether it is similar or not.

In the present embodiment, as an example, the top 10 feature words with the highest scores for each company are displayed in descending order from the feature words with the highest score. This allows the user to know unexpected keyword candidates at the time of the first search. Here, if the score of the same feature word as the keyword "resin molding" first entered by the user is low and lower than the top 10, this feature word "resin molding" is not displayed and another top 10 Individual feature words are displayed as feature words of the company.

Further, for the feature words displayed on the user terminal 10, the number of feature words to be displayed for each company is set (for example, the score is set to the top X feature words (X is a natural number)), or the threshold value is set. By setting to display all the feature words having the above scores, it is possible to reduce the complicated state of the display screen.
By the above process, the user decides to list the candidate business partner companies using the two keywords "resin molding" and "prototype".

In addition to the above-described embodiment, in the process of step S122, the search unit 23 is included in the text data (public company information and private company information) of each company stored in the company database 221 in the storage unit 22. In addition, a company having the same text data as the acquired keyword may be searched for each company, and the company ID of the hit company may be acquired. In the present embodiment, since the number of characteristic words possessed by each candidate company is limited to 500, the above-mentioned characteristic words having a low frequency of appearance and a small score are not retained as data. Therefore, even if a keyword included in the text data of the company database 221 is input as a keyword on the user terminal 10, although it is not retained in the feature word database 223, it should be used in the next and subsequent searches. Can present multiple promising keyword candidates. In this case, the list generation unit 25 is the order candidate company searched by the search unit 23 that the text data matching the keyword input by the user terminal 10 is included, as in the process of step S123 described above. Will generate a list of companies that include multiple feature words owned by, but the keywords are not included in the feature words. However, the user of the user terminal 10 can easily select one or a plurality of keywords that are considered to be appropriate for the next search from the various feature words listed.

By the processing so far, the company database 221 stores the company information including the public company information and the private company information, and the technology tag database 222 stores the technology tag information including the equipment tag, the standard tag and the award history tag. The feature word database 223 stores the feature words and their scores for each company, and the learning model database stores the trained model of the feature words. Then, when the user of the user terminal 10 inputs one keyword, the above processing is performed on the management server 20, and other keyword candidates are displayed on the user terminal 10, so that the user can further select a business partner candidate company. You can select more appropriate keywords to search.

Therefore, in the following embodiment, a process of obtaining a list of business partner candidate companies by a user inputting two keywords and performing an AND search will be described. Here, in the present embodiment, it is assumed that the AND search of the above two keywords is performed. For example, the user of the user terminal 10 inputs "resin molding prototype" and two keywords with a space between them in the input unit 11.

When the management server 20 acquires the two keywords input by the user terminal 10 (step S13), the search unit 23 features a company having the same characteristic words as the acquired keywords, as in the process of step S122 described above. The company ID of the company is acquired by searching from the database 223 (step S14). As a process at this time, for example, a company having both the same characteristic words as the two input keywords (resin molding, trial production) is searched. For example, in the present embodiment, the company IDs "5398, 34589, ..." Of a plurality of companies having both the feature words "resin molding" and "trial production" shown in FIG. 4 are hit, and the feature The company ID "36588, ..." of a company that possesses the word "resin molding" but does not possess the characteristic word "prototype" (assuming that it is not included in the 500 stored as the characteristic word) does not hit.

[Similarity calculation processing]
Next, the similarity calculation unit 24 calculates the similarity between the two acquired keywords and the searched company (step S15). FIG. 9 is a flowchart for explaining the details of the similarity calculation process (step S15) in the management server 20 of the search system 1 according to the embodiment of the present invention.

The similarity calculation unit 24 acquires a distributed representation (high-dimensional vector representation) of the feature words of each company from the learning model database 224 for each hit company and generates a synthetic distributed representation (step S151). In the present embodiment, the synthetic distributed representation is obtained by adding the respective high-dimensional vector representations given to the feature words owned by each company. Specifically, for each hit company, a synthetic distributed expression is created by adding all the distributed expressions of the characteristic words owned by that company (500 characteristic words are given to each company in this embodiment). Generate. At this time, the distributed expressions of each feature word may be simply added together, and the feature word having a higher score is considered to be more representative of the characteristics of the company, and is more than the feature word having a lower score. High weighting may be applied and added together to generate a composite distributed representation.

Next, the similarity calculation unit 24 acquires the distributed representation of the feature words corresponding to the two keywords “resin molding” and “prototype” input by the user terminal 10 from the learning model database 224, and uses each input keyword as the input keyword. Similarly, a synthetic distributed representation is generated from the distributed representation of the corresponding feature word (step S152). In the composition process at this time, both distributed expressions may be simply added, or the second or later (when three or more keywords are input) in the feature word corresponding to the first keyword. ) May be added by giving a higher weight than the feature word corresponding to the keyword.

Then, the similarity calculation unit 24 calculates the similarity between the composite dispersion expression vector of the company generated in step S151 and the composite vector of the keywords generated in step S152 (step S153). In this embodiment, cosine similarity is used as the calculated similarity.

Then, the list generation unit 25 generates a company list including the company name of the company searched by the search unit 23 and the similarity of the company calculated by the similarity calculation unit 24 (step S16). The generated company list is output in a format that can be displayed by a Web browser displayed on the user terminal 10 that has accessed the management server 20. FIG. 10 is a diagram showing an example of displaying a candidate company list including a degree of similarity according to an embodiment of the present invention on the user terminal 10. As shown in FIG. 10, the candidate company name is displayed on the user terminal 10 together with the degree of similarity. As a result, when the candidate companies are listed, it is possible to confirm whether or not they are suitable as business partners based on the degree of similarity. Therefore, the user does not need to check the listed candidate companies company-wide as in the past, but by checking only the companies with the highest similarity or the companies with the similarity above the threshold value, the list is listed. Later labor and cost can be reduced. In addition, the search results (output information) can be standardized, and plus alpha company information can be provided in consideration of the user's search purpose. When displaying the company list, as shown in FIG. 8, for the convenience of the user, the list generation unit 25 refers to the company database 221 and refers to the company name, address, and the company name and address associated with the company ID. The company list may be generated by including the URL, the business outline, and the like in the company list.

In addition, if the requirements (requirement) from the ordering company in the search for a business partner company include items related to owned standards, owned equipment, or whether or not a specific award has been awarded, those are important factors in the search. Therefore, after the candidate companies are listed, the list generation unit 25 can refer to the technical tag information and perform a process of excluding the companies that do not have those technical tags from the list. be. In addition, when listing the companies that own the owned equipment and the like described in the recruitment, it is possible to perform processing such that the companies are displayed at the top as being the companies that match the case.

[Acquisition of private company information and storage processing in company database 221]
Here, there is a possibility that the reliability as a business matching system is low only by the rating with the above-mentioned similarity as it is. This is because the similarity is viewed only by keywords such as public information. Therefore, in the present embodiment, the input information in the project project acquired in the past business partner search and the additional information such as the fact that the ordering company and the order candidate company are actually proceeding to the interview are used as non-public company information. A search system with higher reliability is constructed by storing it in the company database 221 and generating feature words from the combined text data including them.

As described above, in the search for business partner candidate companies in the search system 1 according to the present embodiment, not only public company information but also private company information is in an important position.
Here, the details of the acquisition of private company information by the database construction unit 26 and the storage processing in the company database 221 will be described.

In the present embodiment, the management server 20 of the search system 1 of the management company is accessed from the user terminal of the ordering company or the user terminal 10 of the order candidate company, and the private company information is input or uploaded, respectively, only at a predetermined time. be able to. For example, a business partner search project project starts from the user terminal 10 of the order candidate company, and the person in charge of the order candidate company responds to the project only for one month, which is the period for recruiting the order company, on the management server 20. Allows input via a Web browser managed by. In addition to this, the period may be the entire period in which each business partner search project is being carried out, or it may be possible to input at any time without setting a special period, and the administrator can enter it at any time. It may be possible to adjust whether or not input is possible. The input answer information and the like are stored in the company database 221 by the database construction unit 26 as non-public company information, and the information is newly stored or the existing information is updated.

In general, public company information is general information such as business outlines and service explanations, and often does not contain information on the core technical capabilities of the company. On the other hand, in the answer information such as questions in each business partner search project, the content of the proposal including the appeal statement from the order candidate company that matches the recruitment outline of the search project is written, and it is like each company etc. More specialized technical information about the company, taking into account circumstances, technology and development trends, etc., is entered as text. Therefore, for the feature words generated from this information (private company information), the information obtained in the search for similar business partners in the past can be obtained by increasing the weighting compared to the feature words generated from the public company information. It is possible to build a search system that makes effective use of it. Specifically, a system may be used in which a higher weight is given to the feature words corresponding to the texts used many times in the answer information.

As described above, according to the search system 1 according to the present embodiment, all corporate information such as public information is not treated as equivalent text data, but a smart search that considers various circumstances and trends in the past business partner search. The processing can be performed. As a result, the results of the past business partner search can be effectively utilized, the burden on the user in the business partner search can be reduced, and the candidate companies suitable as business partners can be effectively searched.

[Other processing]
In the above-described embodiment, the keyword input for the company search (the keyword acquired by the management server 20 in step S121) and the acquired keyword are searched in step S122, and are similar to the acquired keyword in step S123. The similarity with the characteristic words owned by the candidate company for which the degree is calculated is taken into consideration, but the similar words with the peripheral keywords (keywords similar to the input keyword) of the input keyword are not considered. Therefore, instead of the similarity calculation process (step S123), another embodiment for calculating the similarity between the peripheral keywords of the input keyword and the feature words owned by the candidate company can be considered.

Specifically, in the similarity calculation unit 24, for each company hit by the search, the distributed expression of the feature words corresponding to the keywords input by the user terminal 10 and acquired by the management server 20 and the keywords with high similarity are used as the basis. The cosine similarity between the obtained vector (vector 1) and the vector (vector 2) based on the characteristic words of each company is calculated.

A vector (vector 1) based on a keyword having a high degree of similarity to the distributed expression of the feature word corresponding to the acquired keyword has a cosine similarity with the distributed expression of the feature word corresponding to the acquired keyword (for example, “resin molding”). The feature words of the high learning model are calculated by the similarity calculation unit 24, the feature words corresponding to the acquired keywords and a plurality of other feature words similar to the feature words (high cosine similarity) are acquired, and they are obtained. It is a vector composed of feature words and their respective scores. When two or more keywords are input, for example, the calculated distributed representation of the feature words is combined to generate the vector 1. When synthesizing the distributed expression, the first keyword is regarded as an important keyword that is the axis rather than the second and subsequent keywords, and the distributed expression of the feature word corresponding to the first keyword is the second. The composition may be performed with a larger weight than the distributed expression of the feature words corresponding to the subsequent keywords.

Further, the vector (vector 2) based on the characteristic word of each company is a matrix (company matrix) having a combination of the characteristic word of each company and the TF-IDF value as an element. In the present embodiment, each company has 500 feature words as shown in FIG. 4, and the company matrix is a vector composed of those feature words and their respective scores.

Then, the similarity calculation unit 24 calculates the similarity between the vector 1 and the vector 2, but at this time, the product is calculated only for the scores between the same feature words included in both vectors, and the products between the different feature words are calculated. The sum of the products of the scores between the feature words is calculated without considering the product of, and the similarity is calculated based on the size.
In the candidate company list generation through the cosine similarity calculation (step S153) described above, since the characteristic words of each company are not weighted, it is effective when it is desired to list a wide range of companies. On the other hand, the generation of a candidate company list through similarity calculation using a vector based on the characteristic words of each company described above is effective when searching for a company with a higher matching probability.

The list generation unit 25 includes the company in the company list when a predetermined condition such as a case where the similarity calculated by the similarity calculation unit 24 is equal to or higher than a predetermined threshold value is satisfied. You may.
Further, when the feature word or the learning model corresponding to the keyword input by the user terminal 10 is not stored in the storage unit 22, the management server 20 can be set to return an error to the user terminal 10. In such a case, the keyword in which the error occurred can be preferentially stored as a feature word, and the process of generating the learning model can be performed. For example, if the input keyword is not registered in the learning model, the corpus (text) including the keyword is additionally learned as input.

Further, in the process of step S12 described above, the AND operation may be performed by inputting two or more keywords to suggest the third and subsequent keywords to the user. AND may be the same as the above processing. Further, depending on the search purpose, the AND search is performed by two keywords in the process of step S13, but even if three or more keywords are used, the same principle may be used, and any of the keywords is included. An OR (or) search may be performed as long as it is performed.

Further, in the present embodiment, an example in which all or a part of the functions provided by the management server 20 are realized by the SaaS type cloud service has been described, but those functions are described in PaaS (Platform as a Service) and IaaS (Platform as a Service). It may be realized by a service provided via Infrastructure as a Service) or API (Application Programming Interface).

The configuration of the search system 1 and the configuration of the user terminal 10 and the management server 20 described in the present embodiment are examples, and may be changed within the scope of the present invention. Further, the processing flow of the management server 20 and the like is also an example, and it is possible to delete unnecessary processing steps, add new processing steps, and replace processing steps within the range not exceeding the scope of the present invention.

1 Search system 10 User terminal 20 Management server 21 Communication unit 22 Storage unit 23 Search unit 24 Similarity calculation unit 25 List generation unit 26 Database construction unit 27 Feature word generation unit 28 Learning model generation unit 30 Network 221 Corporate database 222 Technical tag database 223 Feature word database 224 Learning model database

Claims (7)

A search system that generates a list of companies for the purpose of searching for business partners.
A feature word storage means for storing for each company a plurality of characteristic words with scores based on the current frequency output of the feature words representing the characteristics of the companies,
A learning model storage means that stores the distributed expression of the feature word obtained by performing machine learning that predicts one feature word from the plurality of feature words for each feature word and uses it as a learning model.
An acquisition method for acquiring keywords input from an input device,
A search means for searching a company having a characteristic word corresponding to the keyword from the characteristic word storage means, and
A similarity calculation means for calculating the similarity between a distributed expression based on a plurality of characteristic words of the company searched by the search means and a distributed expression based on the characteristic words corresponding to the keyword.
A search system including a plurality of feature words of the company searched by the search means, and a list generation means for generating a company list including the score and similarity of the feature words. The similarity calculation means calculates the similarity between the synthetic distributed expression obtained by synthesizing the distributed expressions of the plurality of feature words of the company searched by the search means and the distributed expression of the feature words corresponding to the keyword. The search system according to claim 1, wherein the search system is performed. The similarity calculation means calculates a feature word similar to the feature word corresponding to the keyword based on the learning model, and the feature word corresponding to the keyword, a plurality of feature words similar to the feature word, and a plurality of feature words thereof. The search system according to claim 1, wherein the similarity between the first vector including the score, the feature word owned by each company, and the second vector including the score is calculated. The third aspect of the present invention is that the similarity calculation means calculates the similarity based on the sum of the products of the scores between the same feature words included in the first vector and the second vector. Described search system. The similarity calculation means weights a feature word having a higher score for each feature word more than a feature word with a lower score for a distributed expression based on a plurality of feature words possessed by the company searched by the search means. The search system according to any one of claims 1 to 4, wherein the distributed expression is obtained by calculating the similarity with the distributed expression based on the feature word corresponding to the keyword. A search method on a search server that generates a company list for the purpose of searching for business partners.
The acquisition process to acquire the keywords input from the input device, and
The company with characteristic word corresponding to the keyword, the search step of searching a plurality of feature words representing the characteristics of該企industry from the feature word database stored in each company along with the score based on the current frequency output of the characteristic words,
In the search step, referring to a learning model database that stores the distributed expression of the feature word obtained by performing machine learning that predicts one feature word from the plurality of feature words for each feature word and uses it as a learning model. A similarity calculation process for calculating the similarity between the distributed expression based on a plurality of characteristic words possessed by the searched company and the distributed expression based on the characteristic words corresponding to the keyword.
A search method comprising a plurality of feature words of the company searched in the search step, a list generation step of generating a company list including the score and similarity of the feature words. On a search server that generates a list of companies for the purpose of searching for business partners
The acquisition procedure to acquire the keyword input from the input device, and
A search procedure for searching a company having a characteristic word corresponding to the keyword from a characteristic word database that stores a plurality of characteristic words representing the characteristics of the company together with a score based on the frequency of appearance of the characteristic word for each company.
With reference to a learning model database that stores the distributed expression of the feature word obtained by performing machine learning that predicts one feature word from the plurality of feature words for each feature word and uses it as a learning model, the search procedure is performed. A similarity calculation procedure for calculating the similarity between the distributed expression based on a plurality of characteristic words of the searched company and the distributed expression based on the characteristic word corresponding to the keyword, and
A computer-readable search program for executing a plurality of feature words of the company searched by the search procedure, a list generation procedure for generating a company list including the score and similarity of the feature words.

Images