Data processing method, data processing system, and data processing program

Abstract

Provided is a data processing method for executing, by a computer, processing of target data in which values have been input to a plurality of data items of a database, the method comprising: vectorizing information specifying the plurality of data items and attribute information of the values of the plurality of data items; preparing design information including the vectorized information specifying the plurality of data items and the vectorized attribute information of the values of the plurality of data items; inputting text data representing content of data processing to a large-scale language model to cause a word serving as a key to the processing to be output; inputting the output word to a vectorization model to generate a key vector; comparing the key vector with vectors of information specifying the plurality of data items and vectors of attribute information of the values of the plurality of data items, and specifying a data item of the target data and processing content on the basis of a degree of similarity; and executing the processing of the specified content on the specified data item.

Description

Data processing method, data processing system, and program for data processing

[0001] The present invention relates to a technique for executing processing on data in which values are input to each of a plurality of predetermined data items.

[0002] In various places such as companies and public offices, databases that record customer information and the like are used. As a database, a relational database configured by setting data items of information to be recorded in one or more columns and associating a plurality of tables in which records are input in the row direction is widely used. In a relational database, for example, a user can obtain information of a desired record by specifying a search target table and column by SQL (Structured Query Language) statements, setting search conditions, and executing a search (for example, Patent Document 1).

[0003] Japanese Patent Application Laid-Open No. 5-120100 Japanese Patent Application Laid-Open No. 2022-91986

[0004] "Understand the concept of Embedding (vectorization)!", [online], Class Method Co., Ltd., retrieved on February 20, 2025, Internet <URL: https: / / zenn.dev / peishim / articles / c696ff85a539bd>"IBM Granite", [online], IBM Japan, Ltd., retrieved on February 20, 2025, Internet <URL:https: / / www.ibm.com / jp-ja / topics / large-language-models>

[0005] To extract the desired records using the database search described above, the user must accurately specify the search conditions in the SQL statement, including the table name, the names of the data items set in each column, and the attributes of the values ​​of each data item (numeric, text, etc.). Therefore, a problem arose where it was impossible to extract the desired records from the database without knowing the database structure. Although this explanation uses database searching as an example, similar problems exist in processes such as inputting, updating, and deleting database records, as it is necessary to accurately specify the name of the table and column being processed. Furthermore, similar problems existed when processing data sets that do not follow the so-called database format.

[0006] The problem that this invention aims to solve is to improve the convenience of processing data in which values ​​have been entered into at least one of several data items in a database.

[0007] A first embodiment of one aspect of the present invention, made to solve the above problems, is a data processing method in which a computer performs processing on target data in which values ​​are entered for each of at least one data item among a plurality of data items in a database, comprising: preparing a large-scale language model constructed by predetermined machine learning that receives text data as input and outputs a key word for data processing; preparing a vectorization model constructed by predetermined machine learning that vectorizes and outputs the input word; using the vectorization model, vectorizing information that identifies each of the plurality of data items and attribute information of the values ​​of each of the plurality of data items; storing design information including information that identifies each of the vectorized plurality of data items and attribute information of the values ​​of each of the vectorized plurality of data items in a design information storage unit; having a user input text data representing the content of processing on the target data; inputting the text data representing the content of processing into the large-scale language model to output a key word for processing from the large-scale language model; and generating a key vector by inputting the word output from the large-scale language model into the vectorization model. The method is characterized by comparing the key vector with a vector of information that identifies each of the multiple data items and a vector of attribute information for each of the values ​​of the multiple data items, identifying the data items and processing content of the target data based on the similarity with the key vector, and performing the processing of the identified content with respect to the identified data items.

[0008] Large Language Models (LLMs) are generative AIs specialized for Natural Language Processing (NLP), constructed using a neural network whose parameters are adjusted by machine learning on vectors generated through tokenization (dividing various text data into the smallest units such as words and punctuation) and vectorization (converting the tokenized data into numerical values). By pre-training such a large language model with the attributes of the words to be output, it is possible to output words according to the intended use (in this invention, for example, database search) from text data input as natural language. Vectorization models (also called embedding models) are pre-trained models constructed by deep learning using, for example, a large number of words and vectors. By using such vectorization models, it is possible to output vectors that have components that reflect the meaning of the input words.

[0009] The data processing method according to the present invention is typically used to process data (records) that constitute a database, but it can also be used to process other types of data. In the data processing method according to the present invention, a large-scale language model and a vectorization model are prepared in advance. Furthermore, design information is created by vectorizing information that identifies data items and attribute information of the record values ​​entered for each data item. The attribute information of values ​​referred to here includes, for example, numbers, free text, and choices (those selected in a pull-down format, etc.). Note that the values ​​of the target data may also include Null. In addition, the design information may also include information on the relationships between data items. Information on the relationships between data items includes, for example, information that identifies the common table when multiple data items constitute one table, and information that when a value is entered for one data item, the value of another data item is updated based on that value.

[0010] The user inputs text data that describes the content of the processing to be performed on the data. This input text data may include information that identifies the type of processing to be performed on the data. Examples of processing types include data retrieval, inputting, updating, and deleting data item values. Information about the type of processing to be performed on the data is not limited to what the user directly inputs, but may also include information that is indirectly input (for example, identifying the type of processing based on input through a specific screen, or identifying the type of processing based on the sending of text data to a specific email address). The input text data is fed into a large-scale language model, which outputs key words for data processing. Examples of key words for data processing include words related to data items such as industry, address, and number of employees, and words related to search conditions such as within one year, within one month, and Kyoto Prefecture (words that can be used as search terms). Furthermore, a vectorization model is used to vectorize the words output from the large-scale language model to generate key vectors. Then, by referring to the design information, the data item to be processed and the processing content are identified from the key vectors. This method of vectorizing the meaning of words and sentences and identifying those with similar meanings is called semantic search or vector search. By performing such searches, even if the words output from a large-scale language model, the information identifying the data items, and the attribute information of the values ​​of each data item do not perfectly match, the target and content of data processing can be identified based on the semantic similarity (degree of similarity of vector components) between the two. For example, the target and content of data processing can be identified based on having the vector that is closest to the processing key vector (smallest difference in components, shortest distance between vectors, smallest angle between vectors, etc.).

[0011] Once the data items to be processed and the content of the processing for those data items are determined, the processing content specified is executed for the specified data items.

[0012] A second embodiment of one aspect of the present invention is a data processing system used to perform processing on target data in which values ​​are entered for each of at least one data item among a plurality of data items in a database, comprising: a large-scale language model storage unit that stores a large-scale language model constructed by predetermined machine learning that receives text data as input and outputs a key word for data processing; a vectorization model storage unit that stores a vectorization model constructed by predetermined machine learning that vectorizes and outputs an input word; a design information storage unit that stores design information including information that identifies each of the plurality of data items vectorized using the vectorization model, and attribute information of the values ​​of each of the plurality of data items vectorized using the vectorization model; a processing key acquisition unit that causes a user to input text data representing the content of processing on the target data, and inputs the text data representing the content of processing into the large-scale language model to output a key word for the processing from the large-scale language model; and a key vector generation unit that generates a key vector by inputting the word output from the large-scale language model into the vectorization model. The system is characterized by comprising: a processing content identification unit that compares the key vector with a vector of information that identifies each of the plurality of data items and a vector of attribute information of the values ​​of each of the plurality of data items, and identifies the data items and processing content of the target data based on the similarity with the key vector; and a processing execution unit that performs processing of the identified content with respect to the identified data items.

[0013] A third embodiment of one aspect of the present invention is a data processing program for causing a computer to perform processing on target data in which values ​​are entered for each of at least one data item among a plurality of data items in a database, the program comprising: a large-scale language model storage unit that stores a large-scale language model constructed by predetermined machine learning that receives text data as input and outputs a key word for data processing; a vectorization model storage unit that stores a vectorization model constructed by predetermined machine learning that vectorizes and outputs an input word; and a design information storage unit that stores design information including information that identifies each of the plurality of data items vectorized using the vectorization model, and attribute information of the values ​​of each of the plurality of data items vectorized using the vectorization model, the computer in the system comprising: a processing key acquisition unit that causes the user to input text data representing the content of processing on the target data, inputs the text data representing the content of processing into the large-scale language model, and outputs a key word for processing from the large-scale language model; and a key vector generation unit that generates a key vector by inputting the word output from the large-scale language model into the vectorization model. The system is characterized by having a processing content identification unit that compares the key vector with a vector of information that identifies each of the plurality of data items and a vector of attribute information of the values ​​of each of the plurality of data items, and identifies the data items and processing content of the target data based on the similarity with the key vector, and a processing execution unit that performs the processing of the identified content with respect to the identified data items.

[0014] A first embodiment of another aspect of the present invention is a data processing method in which a computer performs processing on a first database having target data in which values ​​are entered for each of at least one of a plurality of first data items, and a second database having target data in which values ​​are entered for each of a plurality of second data items, characterized in that a vectorization model is prepared which is constructed by predetermined machine learning and outputs input words as vectors, information that identifies each of the plurality of first data items and information that identifies each of the plurality of second data items is vectorized using the vectorization model, the vectors of information that identify each of the plurality of first data items are compared with the vectors of information that identify each of the plurality of second data items, and at least one of the plurality of first data items is associated with one of the plurality of second data items based on their similarity.

[0015] The data processing method described above is used to unify (or integrate into a single database) two different databases. In this data processing method as well, a large-scale language model and a vectorization model are prepared in advance, similar to the data processing method described above.

[0016] This data processing method vectorizes information that identifies multiple first data items in the first database and information that identifies multiple second data items in the second database. Then, the vectors of information that identify the first data items are compared with the vectors of information that identify the second data items, and based on their similarity, at least one of the multiple first data items is associated with one of the multiple second data items. As a result, the multiple first data items contained in the first database are made common with the second data items in the second database, and processing of two different databases can be performed as if it were processing a single database.

[0017] A second embodiment of another aspect of the present invention is a data processing system used to perform processing on a first database having target data in which values ​​are entered for each of at least one of a plurality of first data items, and a second database having target data in which values ​​are entered for each of a plurality of second data items, the system comprising: a vectorization model storage unit that stores a vectorization model constructed by predetermined machine learning that vectorizes and outputs input words; a vectorization processing unit that uses the vectorization model to vectorize information that identifies each of the plurality of first data items and information that identifies each of the plurality of second data items; and a data item comparison processing unit that compares the vectors of the information that identifies the plurality of first data items with the vectors of the information that identifies each of the plurality of second data items, and associates at least one of the plurality of first data items with one of the plurality of second data items based on their similarity.

[0018] A third embodiment of another aspect of the present invention is a data processing program that causes a computer to perform processing on a first database having target data in which values ​​are entered for each of at least one of a plurality of first data items, and a second database having target data in which values ​​are entered for each of a plurality of second data items, characterized in that the computer in a system having a vectorization model storage unit that stores a vectorization model constructed by predetermined machine learning that vectorizes and outputs input words is configured to function as a data item comparison processing unit that vectorizes information that identifies each of the plurality of first data items and information that identifies each of the plurality of second data items using the vectorization model, and associates at least one of the plurality of first data items with one of the plurality of second data items based on their similarity.

[0019] By using the data processing method, data processing system, or data processing program according to the present invention, the convenience of processing data in which values ​​have been entered for at least one data item among multiple items in a database can be improved.

[0020] A diagram illustrating the main components of one embodiment of the data processing system according to the present invention. A diagram illustrating the configuration of design information related to data processing in this embodiment. A diagram illustrating the configuration of design information related to knowledge processing in this embodiment. An example of design information related to column processing groups (information groups) and design information related to join information in this embodiment. An example of design information related to basic information and design information related to item information in this embodiment. An example of design information related to selection information in this embodiment. An example of design information related to sample value information in this embodiment. An example of design information related to trigger information in this embodiment. An example of design information related to update detail information by trigger in this embodiment. An example of design information related to knowledge information and design information related to knowledge information-related documents in this embodiment. An example of processing for a database in this embodiment. An example of processing for data entered into a screen, etc., in this embodiment. An example of knowledge processing in this embodiment. An image flow illustrating the flow of integrating the data processing device of this embodiment into an existing business system to execute data processing. An example of request information (structure) in this embodiment. An example of decomposition result information (structure) in this embodiment. An example of return information (structure) in this embodiment. An example of the input form screen for a customer visit report in this embodiment. An example of the input support screen in this embodiment. An example of another input support screen in this embodiment. An example of return information (structure) in this embodiment. An image flow illustrating the process of unifying two databases using the data processing device of this embodiment.

[0021] Embodiments of the data processing method, data processing system, and data processing program according to the present invention will be described below with reference to the drawings. In this embodiment, an example of processing information stored in a business system (corporate database) established in a company will be described.

[0022] Figure 1 shows the main components of the data processing system 1 of this embodiment. The data processing system 1 includes a business system 10 and a data processing device 20 that a company or other entity has within its private network.

[0023] The business system 10 includes a sales support system 11, a sales management system 12, a payroll system 13, and a human resources management system 14. Furthermore, the sales support system 11 has a customer management database (DB) 11D, the sales management system 12 has a financial management database (DB) 12D, the payroll system 13 has a payroll management database (DB) 13D, and the human resources management system 14 has a human resources management database (DB) 14D. While this example describes a scenario where each system has one database, a single system may have multiple databases.

[0024] The customer management database 11D, the financial management database 12D, the payroll management database 13D, and the human resources management database 14D are all relational databases that link together multiple tables in which data items are arranged in columns and records are arranged in rows.

[0025] The data processing device 20 includes a storage unit 21. The storage unit 21 is provided with a design information storage unit 211, a tuning data storage unit 212, a large-scale language model (LLM) storage unit 213, a vectorized model storage unit 214, and an input form storage unit 215.

[0026] The design information storage unit 211 is one of the characteristic elements of this embodiment, and it stores information regarding the structure of various databases that are the target of processing by the data processing device 20 as design information.

[0027] Design information can be broadly categorized into two types: information related to data processing (including database processing and other natural language processing) and knowledge information (data processing performed based on accumulated data).

[0028] Figure 2 shows the structure of the design information related to data processing. The design information related to data processing includes (1) processing groups to specify the content of the processing to be executed, (2) join information representing the join relationships between tables in the database and between multiple input screens, (3) basic information about the database tables and screens, (4) item information regarding columns in tables and items provided in screens, (5) choice information regarding the choices presented in the database columns and items in screens, (6) value sample information regarding the values ​​entered in the database columns and items in screens, and (11) (12) trigger information.

[0029] Figure 3 shows the structure of the design information related to knowledge processing. The design information related to knowledge processing includes (7) knowledge information that stores documents such as daily reports, and (8) knowledge information-related documents (templates, etc.) for generating new information such as instructions and advice based on the knowledge information.

[0030] Figure 4 shows examples of design information for (1) processing groups (information groups) and (2) join information. (1) Processing groups (information groups) do not directly handle information from database tables or screens, but refer to labels as a collection of information. Design information for processing groups (information groups) includes items related to system identification, function group, group code, target identification, name, meaning, vector value, and access rights, each with a value (character, numeric, vector array). (2) Join information refers to joiner information that allows a processing group to process several basic pieces of information as a single unit, such as which combination of tables in the database to execute processing, or, in the case of screens, which collection of screens or objects to process. Design information for join information includes items related to system identification, function group, group code, integration sequence information, information ID, name, identifier, integration type, and integration source information conditions, each with a value (character, numeric).

[0031] Figure 5 shows examples of design information for (3) basic information and (4) item information. (3) Basic information refers to tables in databases, labels for the entire screen in screens, and labels for the entire process in processes. Design information for basic information includes items related to system identification, function group, target identification, information ID, name, meaning, vector value, and access rights, each with a value (character, numeric, vector array). (4) Item information refers to columns in databases, item objects in screens, and labels for process details in processes. Design information for item information includes items related to system identification, function group, target identification, information ID, item ID, selection reference category, selection reference item ID, name, meaning, vector value, attribute, number of digits, default value, item type, required item, automatic setting target, and access rights, each with a value (character, numeric, vector array).

[0032] Figure 6 shows an example of design information related to (5) choice information. (5) Choice information refers to information for each selection value when the attribute is a choice (such as a selection list or radio buttons). The design information related to choice information includes items related to system identification, function group, target identification, information ID, item ID, choice, name, meaning, vector value, attribute, and number of digits, and a value (character, numerical, or vector array) is set for each of them.

[0033] Figure 7 shows an example of design information related to (6) sample value information. (6) Sample value information refers to information for each sample value that represents the characteristics of an item when the attribute of the item is not a choice. The design information related to sample value information includes items for system identification, function group, target identification, information ID, item ID, attribute, three types of sample values ​​(character, number, date), name, meaning, and vector value, and values ​​(character, number, vector array) are set for each. In addition to item-specific information, it is also possible to register and use public information that is common within the system or within a function group for choice information and sample information.

[0034] Figure 8 shows an example of design information related to (11) trigger information. (11) Trigger information refers to monitoring item information and specification information that links and updates other items when there is input or change in that item. The design information related to (11) trigger information includes items related to system identification, function group, trigger management number, target information ID, target item ID, activation condition, activation specified value (character, number, date), action, and access rights, and each of these is set with a value (character, number, date).

[0035] Figure 9 shows an example of design information for (12) trigger detail information. (12) Trigger detail information refers to the source information and destination information after the trigger is activated. The design information for (12) trigger detail information includes items such as system identification, function group, trigger management number, classification for identifying the source / update destination, branch number (row), target information ID, target item ID, read / update conditions, prompt, specified value (character, number, date), and operation-related items, each of which is set with a value (character, number, date).

[0036] Figure 10 shows examples of design information for (7) knowledge information and (8) knowledge information-related documents. The design information for (7) knowledge information includes items for system identification, function group, information-specific number, related type, record document, vector value, keys 1 to 3, and name, each with a value (character, numeric, vector array). The knowledge information-related documents for (8) knowledge information include items for system identification, function group, information-specific number, related document sub-number, and related document, each with a value (character, numeric).

[0037] The tuning data storage unit 212 stores data for tuning the large-scale language model. Tuning the large-scale language model is performed using learning methods such as prompt tuning and fine tuning. This results in a large-scale language model that has learned industry-specific information (e.g., industry terms and their meanings, abbreviations). As tuning data, for example, in the manufacturing industry, information such as the industry term "yield rate," which means the probability that no defective products will occur, that the yield rate is expressed as a numerical value between 0 and 1 (or a percentage between 0 and 100), and that the closer the yield rate is to 1, the higher the proportion of good products, and the closer it is to 0, the higher the proportion of defective products, is used.

[0038] Furthermore, the memory unit 21 includes a large-scale language model memory unit 213 that stores a large-scale language model, which is a trained model constructed by machine learning that receives text data as input and outputs key words for data processing, and a vectorization model memory unit 214 that stores a vectorization model constructed by machine learning that vectorizes input words and outputs them.

[0039] The Large Language Models (LLMs) used in this embodiment are generative AI specialized for natural language processing (NLP). They are constructed using a neural network whose parameters are adjusted by machine learning on vectors generated through tokenization, which divides various text data into the smallest units such as words and punctuation, and vectorization, which quantifies the tokenized data. In this embodiment, the attributes of the words to be output are pre-trained in such a large language model, and the tokenization and cleansing of text data input in natural language, etc., are performed to output words related to the processing of the data.

[0040] In this embodiment, a large-scale language model stored in a large-scale language model storage unit 213, which is only usable within the company (located within a private network with no external connections), is used. It is difficult for a single company to own a large-scale language model, and it is common to share a large-scale language model provided on the cloud with an unspecified number of people (or a specific number of people). In this invention, it is also possible to use a large-scale language model that is generally widely available (public), but if information related to confidential company matters such as customer information is trained on the large-scale language model, there is a risk that the information may be leaked to the outside.

[0041] Therefore, as in this embodiment, it is preferable to use a large-scale language model that is only available in a closed environment (private). In this embodiment, it is sufficient to extract words related to the processing of information stored in the database (search, data entry, data update, etc.) from the input natural language, and it is not necessary to use a machine learning model that has acquired all kinds of words. As such a large-scale language model, for example, IBM's Granite (Non-Patent Literature 2) can be suitably used. Granite is a large-scale language model designed to be lightweight and have a low computational load, and it can be owned and used by companies even if they are not large corporations. When using such an unofficial large-scale language model, it is preferable to perform tuning using the design information stored in the design information storage unit 211 when performing the above tuning. This makes it possible to output processing keys that can more accurately identify tables, columns, etc. from the large-scale language model.

[0042] The vectorization model used in this embodiment is a trained model constructed by deep learning using a large number of word-vector pairs, and constitutes part of a large-scale language model. In this embodiment, by using such a vectorization model, it is possible to output a vector having components that reflect the meaning of the input word. The vector array (vector value) included in the above design information can be generated using such a vectorization model by the vectorization processing unit 30 described later. The vectorization model is also called an embedding model (for example, Non-Patent Literature 1) and is used to perform semantic search (also called vector search). Suitable software that can be used via a network such as the Internet may be used for the vectorization model, or software owned by a company may be used.

[0043] The storage unit 21 is further equipped with an input form storage unit 215 that stores input forms for various types of data. For example, input forms related to the customer management database 11D include a form for registering a new customer, a form for reporting the results of customer visits, and a form for entering the results of transactions with customers. These input forms are provided with data items corresponding to columns and display fields for entering the values ​​of those columns. Furthermore, for specific data items, processing is associated with the input of a value in a particular data item, triggering changes to the value of another column (this is stored as trigger information; Figures 8 and 9). In addition, when text data such as natural language is entered into specific input / display fields of each input form, processing is performed using a large-scale language model and / or a vectorization model (this is also stored as trigger information).

[0044] The data processing device 20 includes, as functional blocks, a vectorization processing unit 30, a design information creation unit 31, a tuning execution unit 32, a processing key acquisition unit 33, a key vector generation unit 34, a processing content specification unit 35, a processing instruction generation unit 36, a processing execution unit 37, a data item comparison processing unit 38, and a record value conversion unit 39. The data processing device 20 is composed of, for example, a general personal computer or a workstation, and these functional blocks are realized by executing a pre-installed data processing program with a processor. Alternatively, the data processing device 20 may be provided in a cloud server. Further, an input unit 51 composed of a keyboard, a mouse, etc. and a display unit 52 composed of a liquid crystal display, etc. are connected to the data processing device 20.

[0045] FIGS. 11 to 13 show an overview of the data processing performed in the data processing system 1 of the present embodiment. FIG. 11 shows processing for a database, FIG. 12 shows processing for data input to a screen, etc., and FIG. 13 relates to knowledge processing.

[0046] As shown in FIG. 11, there are four types of processing for a database: data search (SELECT), data update (UPDATE), data creation (INSERT), and data deletion (DELETE). When executing these data processes, text data consisting of natural sentences is input in each case, and processing by a large language model (LLM) (language processing 1 by LLM) and vector database (DB) processing (vector DB processing 1) are performed to obtain outputs respectively.

[0047] As shown in FIG. 12, there are three types of processing for data input to a screen or the like: processing for generating a summary from an item group, processing for generating an item group from a sentence, and processing for generating a condition group from a sentence. In the processing for generating a summary from an item group, information on the structure of the item group and values is input, and processing by a vector database (DB) (vector DB processing 1) and processing by a large language model (LLM) (language processing 2 by LLM) are performed to obtain a summary document. In the processing for generating an item group from a sentence and the processing for generating a condition group from a sentence, natural text and structure information of the item group to be set are input, and processing by a large language model (LLM) (language processing 1 by LLM) and processing by a vector database (DB) (vector DB processing 1) are performed to obtain an array of the item group and set values.

[0048] As shown in FIG. 13, there are three types of knowledge processing: accumulation (vector DB), summarization, and proposal. In the accumulation processing, information on the structure of the item group and values that are the source of summarization, related types, and related documents are input, and processing by a large language model (LLM) (language processing 1 by LLM), processing by a vector database (DB) (vector DB processing 1), processing by a large language model (LLM) (language processing 2 by LLM), and processing by a vector database (DB) (vector DB processing 2) are performed to accumulate data (no output is performed in the accumulation processing). In the summarization processing, information on the structure of the item group and values that are the source of summarization are input, and processing by a large language model (LLM), processing by a vector database (DB) (vector DB processing 1), and processing by a large language model (LLM) (language processing 2 by LLM) are performed to obtain a summary document as output. In the proposal processing, the item group and structure information of values that are the source of summarization and the return related type are input, and processing by a large language model (LLM) (language processing 1 by LLM), processing by a vector database (DB) (vector DB processing 1), processing by a large language model (LLM) (language processing 2 by LLM), and processing by a vector database (DB) (vector DB processing 2) are performed to obtain related documents as output.

[0049] Next, the operation of the data processing system of the present embodiment (an embodiment of the data processing method according to the present invention) will be described.

[0050] In the data processing system 1 of this embodiment, for each database of the business system 10 that is the target of processing, the processing described with reference to Figures 11 to 13 is executed using the design information described with reference to Figures 2 to 10.

[0051] The data processing device 20 of this embodiment can be integrated into an existing business system 10. Figure 14 is an image flow illustrating the flow of execution of the above processing in a data processing system 1 in which the data processing device 20 is integrated into an existing business system 10. In Figure 14, the area enclosed by the dashed line shows the processing and operation in the business system 10, and the area enclosed by the dashed line shows the processing and operation by the data processing device 20. The business system 10 and the data processing device 20 are configured to send and receive information via an Application Programming Interface (API).

[0052] Before operating the data processing system 1 of this embodiment, the design information creation unit 31 creates the design information described with reference to Figures 2 to 10 and stores it as a vector database (DB) in the design information storage unit 211 of the data processing device 20. In addition, the tuning execution unit 32 performs learning such as prompt tuning and fine tuning of the large-scale language model using the tuning data stored in the tuning data storage unit 212 and stores the trained large-scale language model in the large-scale language model storage unit 213.

[0053] Generally, a database is associated with a program for processing data within that database, and through this program, users can perform various data operations on that database. In conventional data processing, users input commands into the program in a language that the program can process, thereby performing data processing.

[0054] On the other hand, when the data processing device 20 of this embodiment is incorporated, when a user inputs text data (request information) such as natural language, it is sent to the data processing device 20 via the API. In the data processing device 20, the processing key acquisition unit 33 inputs the text data as request information (structure) into the large-scale language model, and obtains the decomposition result information (structure) output by the large-scale language model as a processing key (language processing 1 by LLM).

[0055] As shown in Figure 15, the request information (structure) may include API basic information, target identification, request action, requested natural language statement, additional prompts, information ID, and input information array. Of these, the API basic information and target identification can be identified, for example, from the screen entered by the user or the email address to which the user sent text data. Also, as shown in Figure 16, the decomposition result information (structure) may include API basic information, target identification, information ID, input information array, and decomposition information array.

[0056] Next, the key vector generation unit 34 vectorizes the processing key using a vectorization model. Then, the processing content identification unit 35 uses the vector to perform a semantic search of specification information, values, and options within the design information (vector DB) to identify the content of the data processing requested by the text data entered by the user (vector DB processing 1).

[0057] Here, if the processing content is processing on the database (Figure 11), the processing instruction generation unit 36 ​​generates an instruction statement (an instruction statement written in a programming language associated with the target database) describing the content of the specified data processing as return information (structure), and sends it to the business system 10 via the API. As shown in Figure 17, the return information (structure) may include result status information and a result information array.

[0058] If the processing content is a process that generates a summary from a group of items (Figure 12) or knowledge processing (Figure 13), the processing execution unit 37 executes processing using the large-scale language model again (such as creating a summary text) (Language processing using LLM 2). If the processing content is summarization processing or suggestion processing (Figure 13), the processing execution unit 37 executes a search of the vector DB using the vectorization model again (Vector DB processing 2). After these processes, the processing instruction generation unit 36 ​​generates an instruction statement (an instruction statement written in a programming language associated with the target database) describing the content of the identified data processing as return information (a structure) and sends it to the business system 10 via the API.

[0059] The data processing device 20 outputs information in a format pre-associated with the content of each data processing operation after the above processing has been performed.

[0060] The information output from the data processing device 20 is input to the business system 10 via an API. In the business system 10, programs associated with each database process this information. This allows for data processing based on request information entered by the user in text data such as natural language.

[0061] In conventional data processing systems, in order for users to perform desired data processing (such as extracting records), they had to accurately specify the table name, the names of the data items set in each column, the attributes of the values ​​of each data item (numeric, text, etc.), and the processing content in an SQL statement. This presented a problem: users could not extract the desired records from the database without knowing the database structure or programming language.

[0062] In contrast, in this embodiment, even if the processing key (the word) does not match the names set in the table information, column information, etc., that constitute the database, the table and column to be searched can be identified based on the semantic similarity determined by comparing vector values, and further, the processing content for the values ​​of those columns can be identified. This type of processing is called semantic search or vector search. By performing a semantic search in this way, the desired information can be extracted even if the user does not have knowledge of the structure of the database being processed. In addition, the user only needs to input the content of the desired data processing in text data such as natural language, and no knowledge of programming languages ​​is required.

[0063] Furthermore, in business systems, it is common for various employees to add records or change or update the values ​​of data items. However, not all employees know the input format for entering data into data items. As a result, even values ​​with the same meaning may be mixed in different formats. Examples of different formats include the mixing of Arabic numerals and Chinese numerals, and the presence or absence of units (entering units causes the data to be recognized as text data and excluded from statistical processing of numerical data). In conventional data processing systems, when values ​​in various formats are entered in this way, the entered values ​​are reflected as they are. For example, when performing statistical processing on numerical data, values ​​entered with Chinese numerals or values ​​entered with units are not processed statistically, which was a problem.

[0064] In contrast, in the data processing system 1 of this embodiment, data processing is performed in a format specified in the design information from text data entered by the user in various formats, so statistical processing of numerical data and data extraction can be performed accurately.

[0065] The data processing system 1 of this embodiment can implement the following functions.

[0066] Support functions: For example, automatically filling in remaining items based on some input data such as request natural language text. Collecting and summarizing individual input values ​​to create a summary document (e.g., the body of a daily report). Analyzing the content of natural language text written in emails, etc., and calling up the processing flow that should be performed afterward.

[0067] Automatic creation function: By attaching audio, memos, photos, etc., to emails from outside the office and sending them to a designated email address that has been pre-associated with API basic information, target identification, and requested actions, daily reports and other documents are automatically created.

[0068] Response function: The system searches for past inquiries with similar meanings to the inquiry received via the inquiry form or the email address set as the contact address, and proposes a response using a pre-prepared template.

[0069] Data extraction (search) function: Automatically sets appropriate search conditions according to the requested text. Storage function: Stores data on successful and unsuccessful patterns and utilizes it as knowledge. For example, it generates summary reports and proposals from stored data that has a similar meaning to the data entered, such as daily reports. Application: Reads text from various image data such as business cards and invoices and automatically extracts customers and personnel.

[0070] Below, we will describe some specific examples of data processing in the data processing system 1 of this embodiment.

[0071] <Database Processing: SELECT> When a user inputs the text data "List of customers who have not been visited recently" from a screen pre-associated with the customer management database 11D, the large-scale language model outputs the words "customer," "list," and "not recently visited." Semantic search of the design information identifies that "customer" refers to the target table, "list" refers to the target column group, and "not recently visited" refers to the search condition. The processing command generation unit 36 ​​then generates a command statement to execute this data search and sends it to the business system 10 as return information.

[0072] <Database Processing: UPDATE> When a user enters the text data "Enable direct mailing to prospective customers in negotiations within a 5km radius" from a screen pre-associated with the customer management database 11D, the large-scale language model outputs the phrases "In negotiations; Customer", "Direct mailing; Enabled", and "Within a 5km radius; Prospective customer". Then, a semantic search of the design information identifies that the phrase "In negotiations; Customer" refers to the target table. The system also searches for column information corresponding to the phrase "Direct mailing; Enabled" and searches for the specification related to "Enabled", which represents the processing for that column. The system then identifies the option information that is closest in meaning. Furthermore, even if a specific numerical value such as "Within a 5km radius" is entered, the system uses a semantic search of the design information to identify the option information that is closest in meaning. As a result, for customers that meet the search criteria "Within a 5km radius; Prospective customer", the option for the data item "Direct mailing" is updated to the value "Enabled".

[0073] <Processing on the database: INSERT & Processing on items on screens, etc.: Conditional group from text> The user accesses the screen to add a "customer visit report" to the customer management database 11D. The processing key acquisition unit 33 reads the input form for the customer visit report stored in the storage unit 21 and displays it on the screen of the display unit 52.

[0074] Figure 18 shows an example of the input form screen for a customer visit report. This screen has fields for customer code, customer name, status report, ABC rank, and probability. Of these, the user enters text data in the customer name and status report fields (fields shown with solid lines in Figure 18). When the user enters the customer name, the customer code is retrieved from the database and displayed in the customer code field.

[0075] The status report field is used to input natural language text data. When the user inputs natural language text data into the status report field, the processing key acquisition unit 33 inputs the input natural language into the large-scale language model. The large-scale language model outputs words related to data processing (processing keys) from the input natural language, such as "order amount," "high value," "promising," "competitor," "slightly unfavorable," and "situation."

[0076] The key vector generation unit 34 inputs the text output from the large-scale language model into the vectorization model and generates key vectors. The processing content identification unit 35 refers to the design information and identifies the content of the data processing corresponding to the generated vector values. In this example, the design information includes trigger information that generates values ​​for priority and accuracy data items when a value is input for the status report data item. Referring to this design information and the attribute information of the values ​​input to priority and accuracy respectively, the processing content identification unit 35 determines the input values ​​for the priority and accuracy data items from the output word group (in the example in Figure 18, the priority value is "A" and the accuracy value is "50%").

[0077] The processing command generation unit 36 ​​generates a command statement to add the processing content (data item and its value pair) identified by the processing content identification unit 35 as a record to the customer management database 11D, and sends it to the business system 10 as return information (structure). The business system 10 executes the processing command statement generated by the processing command generation unit 36 ​​and adds the record to the database specified by the user (in this case, the customer management database 11D).

[0078] In conventional data processing systems, information entered by users in natural language, such as the situation report in the example above, is simply entered as database values, and meaningful statistical processing cannot be performed. In contrast, in the data processing system 1 of this embodiment, predetermined data items (in this case, priority and accuracy) are determined from the natural language entered by the user, making it possible to perform meaningful statistical processing using these values. Furthermore, with natural language input, even if the content is the same, the length and expression of the text can differ depending on the person entering it, making it difficult to correctly grasp the situation without carefully reading the entered natural language text. In contrast, in the data processing system 1 of this embodiment, a large-scale language model can be used to evaluate natural language written by various users using common criteria, and values ​​obtained from this evaluation can be registered in the database.

[0079] Patent Document 2 describes a technique for generating SQL statements from natural language, which involves pre-matching the question information to be processed with the table content of a target table to obtain pre-matching results, extracting the word segment with the highest degree of matching to the question, determining the target attribute value semantically associated with the extracted word segment, and generating an SQL query statement using that target attribute value. According to Patent Document 2, this technique makes it possible to generate an SQL statement even if the question information presented by the user does not perfectly match the column names or attribute values ​​of the table.

[0080] Patent Document 2 only conceptually presents the above technology, and does not show any specific examples of its application. This technology generates SQL statements in response to interactive requests from a human and outputs the data resulting from executing those statements; it cannot use any search methods other than SQL statements. Therefore, it can only perform database searches that conform to the input format of the SQL statement. Furthermore, it is not easy to accurately output the elements necessary for data processing from a large-scale language model and generate SQL statements. In addition, the meaning of the term "attribute" is not clear in Patent Document 2. Moreover, it is unclear how to apply this technology when there are multiple tables in the database (such as in a so-called relational database).

[0081] On the other hand, while SQL statements can be used in the data processing system 1 of this embodiment, it essentially outputs the elements necessary for database searching through semantic search, and can be used for search methods other than database searching using SQL statements. Furthermore, in this embodiment, as explained with reference to Figures 2 to 10, the attributes of each item are defined in the database design information, thereby enabling accurate data processing in line with the user's intentions. Moreover, the data processing system 1 of this embodiment uses the large-scale language model and vectorization model not only for database searching, but also for registering records in the database, updating and deleting data item values, and processing input to items on screens, etc. (Figure 12) and knowledge processing (Figure 13), enabling diverse data processing.

[0082] The above processing may be performed solely within the data processing device 20 after receiving text data input from the user. However, it is preferable that the processing content identification unit 35 identifies the content of the data processing requested based on the text data input by the user and displays that content on the screen of the display unit 52. This allows, for example, the user to partially add or modify the data if the processing result by the data processing device 20 is incomplete.

[0083] Below are some specific examples of such screen displays.

[0084] <Input Support (Database Search)> Figure 19 shows an example of a screen (input support screen 6) that assists the user in inputting data when searching for records in the database. Input support screen 6 is provided with a free word input field 61 where the user enters text data (words, sentences, etc.). The processing content identification unit 35 identifies the processing content (database search) from the entered text data and displays it in the search support area 62. The process of identifying the processing content from the text data is the same as described above, so a detailed explanation is omitted. In this example, the search period for the last transaction date is set to October 1, 2024 to September 30, 2025 (example of screen display when input is made on September 30, 2025), and the search word for industry is set to "manufacturing." Furthermore, in this input support screen 6, additional data items such as "Customer Classification" and "Company Size," which are more detailed classifications (e.g., medium classification, minor classification) associated with the data item "Industry" (e.g., major classification), are displayed. Users can perform more detailed refined searches by selecting from the additional data items in a pull-down format. Of course, users can also perform a search without selecting from the additional data items and with the display as shown in Figure 19.

[0085] <Input Assistance (Record Registration in Database)> Figure 20 shows an example of a screen (input assistance screen 7) that assists the user in inputting data when adding a record to the database. This example is a screen for inputting a report on the results of a sales visit to a customer, and is similar to the example in Figure 18. The input assistance screen 7 has a status report field 71 where the user inputs the results of the customer visit as text data (words, sentences, etc.). The processing content identification unit 35 identifies the processing content (adding a record to the database) from the text data and displays it in the input assistance area 72. In this example, the value A is entered into the ranking data item from the phrase "high potential," the value 50% is selected and entered into the probability data item 1 from the phrase "slightly unfavorable situation," and the value Push is selected and entered into data item 2 from the phrase "I will try pressing it." The user confirms that these automatically entered values ​​are appropriate (and corrects them if they are deemed inappropriate), and adds the record by performing a predetermined input operation such as pressing the OK button.

[0086] The above describes a case where the user inputs text data into the status report field 71, which is then analyzed and values ​​are entered into the rank, accuracy 1, and accuracy 2 data items. However, the reverse process is also possible. That is, as shown in the input support screen 8 in Figure 21, the user can select and input values ​​into the rank, accuracy 1, and accuracy 2 data items in the selection item field 81, and the processing content identification unit 35 can create natural language text based on these selected values ​​and display it in the status report field 82. In this case as well, the user can review the created text, make corrections as necessary, and then add a record by performing a predetermined input operation such as pressing the confirm button.

[0087] Furthermore, the data processing system 1 of this embodiment can also unify two different databases. For example, it can be suitably used when importing a corporate information database registered in a data bank into a company's own customer management database. Generally, public databases such as corporate information databases have a wide range of data items (for example, about 200 data items), but when used as a company's own customer information database, only a portion of them (for example, about 10 data items) are necessary. Of course, the uses are not limited to this, and it can be suitably used when integrating databases in conjunction with corporate mergers and acquisitions, or when integrating old databases into new databases in conjunction with database updates.

[0088] First, the user specifies the database to be integrated (for example, a public corporate information database; hereinafter referred to as the "first database") and the target database for integration (for example, a customer management database 11D; hereinafter referred to as the "second database") through a predetermined input operation via the input unit 51, and instructs the system to unify (integrate) the databases.

[0089] Upon receiving the above instructions, the vectorization processing unit 30 reads information that identifies each of the multiple first data items in the first database (typically the name of the data item) and attribute information of the values ​​of each of the multiple first data items (attributes of the record entered into the data item), and inputs these into the vectorization model stored in the vectorization model storage unit 214 to vectorize them. The vectorization processing unit 30 also reads from the design information storage unit 211 vectors of information that identifies the multiple data items in the second database and vectors of attribute information of the values ​​of each of the multiple data items. In this example, since the design information for the second database is already stored in the design information storage unit 211, the vectors are read from there. However, if the design information for the second database is not stored, the vectorization processing unit 30 vectorizes the information that identifies each of the multiple data items and the attribute information of the values ​​of each of the multiple data items by the same process as for the first database.

[0090] Next, the data item comparison processing unit 38 compares (performs a semantic search) the information vectors that identify each of the multiple data items in the first database (hereinafter referred to as "first data item vectors") with the information vectors that identify each of the multiple data items in the second database (hereinafter referred to as "second data item vectors"). Then, for each of the multiple first data item vectors, it identifies the second data item vector whose similarity satisfies a predetermined criterion (for example, the difference in vector components or the distance between vectors is the closest, and that distance is below a predetermined threshold), and associates the two. If no second data item vector satisfies the criterion, it identifies that no corresponding second data item vector exists.

[0091] Furthermore, the record value conversion unit 39 determines, for each pair of the first data item vector and the second data item vector, whether the attribute information of the value of the first data item is the same as the attribute information of the value of the second data item to which the first data item is associated. If they are different, the record value conversion unit 39 converts the value of the first data item to match the attribute information of the value of the second data item.

[0092] By performing the above process, it becomes possible to use two originally separate databases as a single database, sharing common data items and attribute information for the values ​​of those data items. This allows data retrieval and other processes that previously had to be performed individually for each database to be performed all at once. Furthermore, by incorporating only the records of the data items in the first database that are associated with the second data items as records in the second database, it becomes possible to extract the necessary information from the records contained in the first database and integrate it into the second database.

[0093] Furthermore, the above process does not modify the first database itself, but rather identifies the first data items that are common to the second database and changes the attribute information of their values ​​(i.e., the first database itself can be left as is). Therefore, even after unifying the old and new databases (the old database being the first database and the new database being the second database), the functions unique to the first database (record input into the first database, data retrieval in the first database) can be maintained. This makes it possible to continue using the functions of the old database that you are familiar with.

[0094] Figure 22 is an image flow illustrating the process of performing database standardization (integration). In Figure 22, the area enclosed by the dashed line shows the processing and operation by the data processing device 20.

[0095] When performing database standardization (integration), the vectorization processing unit 30 creates the design information for the first database in advance.

[0096] Next, the data item comparison processing unit 38 compares the vectors of multiple first data items contained in the design information of the first database with the vectors of second data items contained in the design information of the second database. Based on the similarity between the two, it associates each of the multiple first data items with one of the second data items. Furthermore, if the attribute information of the values ​​differs between the associated data items, the record value conversion unit 39 converts the value of the first data item to match the attribute information of the value of the second data item.

[0097] As described above, the data processing method, data processing system, and data processing program of this embodiment combine artificial intelligence (AI) with business systems (including existing, vast business systems). By using the data processing method, data processing system, and data processing program of this embodiment, existing business systems can be improved into flexible and efficient business systems that are compatible with AI with minor modifications, without the need to redesign existing business systems or redevelop the systems. Furthermore, the data processing method, data processing system, and data processing program of this embodiment can accumulate more design information as it is used, thereby improving the AI's understanding of the design information. As a result, it is anticipated that in the future, it will be possible to have the AI ​​generate various data input / output screens and reports themselves according to user needs, and the programs that handle such processing will become unnecessary.

[0098] The above embodiments are examples and can be modified as appropriate in accordance with the spirit of the present invention.

[0099] In the above embodiment, we described a case where the user identifies the database to be processed or inputs the data processing content from an entry point (a specific screen, a specific email address) corresponding to the data processing content. However, the entry point for executing all databases and all types of data processing may be common. In that case, the processing content identification unit 35 refers to the design information of all databases and identifies the database to be processed and the type and content of data processing in that database based on the key vector generated by the key vector generation unit 34 from the text data such as natural language input by the user. However, by separating the entry point, such as a screen for the user to input instructions, for at least one database, the load related to semantic search is reduced, and more accurate data processing becomes easier.

[0100] In some cases, it may not be possible to extract all the elements necessary for the instruction statement to execute data processing from the text data entered by the user. In such cases, the processing content identification unit 35 should notify the user by displaying the missing elements on the screen of the display unit 52, or by outputting an error message.

[0101] In the above embodiment, an example was described in which all the necessary configurations for searching information in the database and inputting data into the database are provided inside the data processing device 20. However, some of these configurations may be provided in other devices (including other devices in a private network, or cloud servers provided on an open network with access restricted by encryption or passwords). For example, in the above embodiment, the large-scale language model storage unit 213 and the vectorization model storage unit 214 are provided inside the data processing device 20. However, large-scale language models and vectorization models provided in other devices (same as above) connected to the data processing device 20 via a network may be used. Furthermore, when using public large-scale language models and vectorization models, those provided in other devices (cloud computers, etc.) connected via an open network (the Internet, etc.) may be used. In addition, functional blocks related to database searching and functional blocks related to data input may be provided in separate devices (same as above), and these devices may be connected to the network inside the data processing system 1.

[0102] In the above embodiment, an example of data processing for a customer management database 11D provided in a company's business system 10 was described. However, a similar configuration can be adopted when processing information in various types of databases provided in public institutions such as government agencies. Furthermore, the specific examples of data processing described in the above embodiment are only a part of what can be executed in the data processing system 1 of this embodiment, and can be appropriately modified depending on the contents of the database and its intended use. In addition, in the above embodiment, a configuration in which the business system 10 and the data processing device 20 are provided within a single private network was described, but they may be connected via an open network (such as the Internet).

[0103] In the above embodiment, we described a case where access to all data in the database is permitted. However, it is also possible to restrict access to certain data (for example, data related to management, financial data, or personnel data) to only those with access privileges. For example, when using the data processing system 1, users can log in using a user ID and password, and the scope of access can be determined according to the authenticated user ID.

[0104] In the above embodiment, an example of processing data in a database was described, but the same configuration can be used when processing data other than in a database. For example, the process of adding a customer visit report described in the above embodiment can also be used for purposes such as submitting the customer visit report in writing (i.e., for purposes that do not involve adding it as a record in a database). Also, in the above embodiment, the values ​​of data items such as priority and probability were generated from natural language text entered by the user, but conversely, the system may be configured so that the user inputs the priority and probability, and natural language text is generated from these values ​​using a large-scale language model.

[0105] Furthermore, multiple trigger information in the above embodiment can be combined to execute flowchart-like processing. For example, if the number of records with priority A in the customer management database 11D exceeds 10, those records can be extracted and output as a list of data. Various processes can be combined and executed in this way.

[0106] In the above embodiment, after associating the first data item with the second data item, a process was performed to convert the value of the first data item so that the attribute information of the values ​​of both items matched. However, this is a preferred process but not mandatory. For example, if there is no problem in processing the database after commonization (integration) even if the attribute information of the data item values ​​is not identical, this process may be omitted. Also, although two databases were commonized (integrated) in the above example, the same process can be applied to three or more databases.

[0107] [Embodiments] It will be apparent to those skilled in the art that the exemplary embodiments described above are specific examples of the following embodiments.

[0108] (Section 1) A first embodiment of one aspect of the present invention is a data processing method in which a computer performs processing on target data in which values ​​are entered for each of at least one data item among a plurality of data items in a database, comprising: preparing a large-scale language model constructed by predetermined machine learning that receives text data as input and outputs a key word for data processing; preparing a vectorization model constructed by predetermined machine learning that vectorizes and outputs the input word; using the vectorization model, vectorizing information that identifies each of the plurality of data items and attribute information of the values ​​of each of the plurality of data items; storing design information including information that identifies each of the vectorized plurality of data items and attribute information of the values ​​of each of the vectorized plurality of data items in a design information storage unit; having a user input text data representing the content of processing on the target data; inputting the text data representing the content of processing into the large-scale language model to output a key word for processing from the large-scale language model; and generating a key vector by inputting the word output from the large-scale language model into the vectorization model. The method is characterized by comparing the key vector with a vector of information that identifies each of the multiple data items and a vector of attribute information for each of the values ​​of the multiple data items, identifying the data items and processing content of the target data based on the similarity with the key vector, and performing the processing of the identified content with respect to the identified data items.

[0109] (Section 7) A second embodiment of one aspect of the present invention is a data processing system used to perform processing on target data in which values ​​are entered for each of at least one data item among a plurality of data items in a database, comprising: a large-scale language model storage unit that stores a large-scale language model constructed by predetermined machine learning that receives text data as input and outputs a key word for data processing; a vectorization model storage unit that stores a vectorization model constructed by predetermined machine learning that vectorizes and outputs an input word; a design information storage unit that stores design information including information that identifies each of the plurality of data items vectorized using the vectorization model, and attribute information of the values ​​of each of the plurality of data items vectorized using the vectorization model; a processing key acquisition unit that causes a user to input text data representing the content of processing on the target data, and inputs the text data representing the content of processing into the large-scale language model to output a key word for processing from the large-scale language model; and a key vector generation unit that generates a key vector by inputting the word output from the large-scale language model into the vectorization model. The system is characterized by comprising: a processing content identification unit that compares the key vector with a vector of information that identifies each of the plurality of data items and a vector of attribute information of the values ​​of each of the plurality of data items, and identifies the data items and processing content of the target data based on the similarity with the key vector; and a processing execution unit that performs processing of the identified content with respect to the identified data items.

[0110] (Section 9) A third embodiment of one aspect of the present invention is a data processing program for causing a computer to perform processing on target data in which values ​​are entered for each of at least one data item among a plurality of data items in a database, the program comprising: a large-scale language model storage unit that stores a large-scale language model constructed by predetermined machine learning that receives text data as input and outputs a key word for data processing; a vectorization model storage unit that stores a vectorization model constructed by predetermined machine learning that vectorizes and outputs an input word; a design information storage unit that stores design information including information that identifies each of the plurality of data items vectorized using the vectorization model, and attribute information of the values ​​of each of the plurality of data items vectorized using the vectorization model, the computer in the system comprising: a processing key acquisition unit that causes a user to input text data representing the content of processing on the target data, inputs the text data representing the content of processing into the large-scale language model, and outputs a key word for processing from the large-scale language model; and a key vector generation unit that generates a key vector by inputting the word output from the large-scale language model into the vectorization model. The system is characterized by having a processing content identification unit that compares the key vector with a vector of information that identifies each of the multiple data items and a vector of attribute information for the values ​​of each of the multiple data items, and identifies the data items and processing content of the target data based on the similarity with the key vector, and a processing execution unit that performs the processing of the identified content on the identified data items.

[0111] The data processing method described in paragraph 1, the data processing system described in paragraph 7, and the data processing program described in paragraph 9 are typically used to process data (records) that constitute a database, but they can also be used to process other types of data. In the data processing method according to the present invention, a large-scale language model and a vectorization model are prepared in advance. Furthermore, information that identifies data items and attribute information of the record values ​​entered for each data item are vectorized, and design information is created together with information on the relationships between data items. The attribute information of values ​​referred to here includes, for example, numerical values, free text, and selection options (those selected in a pull-down format, etc.). Note that the values ​​of the target data may also include Null. Furthermore, information on the relationships between data items includes, for example, information that identifies the common table when multiple data items constitute one table, and information that when a value is entered for one data item, the value of another data item is updated based on that value.

[0112] The user inputs information that identifies the type of processing to be performed on the data, along with text data representing the content of that processing. Examples of data processing types include data retrieval, data item value input, updating, and deletion. The information regarding the type of processing is not limited to direct user input; it may also include indirect input (e.g., identifying the type of processing based on input through a specific screen, or identifying the type of processing based on text data being sent to a specific email address). The input text data is fed into a large-scale language model, which outputs key words for data processing. Examples of key words for data processing include words related to data items such as industry, address, and number of employees, as well as words related to search conditions such as "within one year," "within one month," and "Kyoto Prefecture." Furthermore, a vectorization model is used to vectorize the words output from the large-scale language model, generating key vectors. Then, by referring to design information, the data items to be processed and the processing content are identified from the key vectors. This method of vectorizing the meaning of words and sentences and identifying similar meanings is called semantic search or vector search. By performing such searches, even if the words output from a large-scale language model, the information identifying the data items, and the attribute information of the values ​​of each data item do not perfectly match, the target and content of data processing can be identified based on the semantic similarity (degree of similarity of vector components) between the two. For example, the target and content of data processing can be identified based on having the vector that is closest to the processing key vector (smallest difference in components, shortest distance between vectors, smallest angle between vectors, etc.).

[0113] Once the data items to be processed and the content of the processing for those data items are determined, the system either executes the specified processing on the identified data items or outputs information for executing the processing. This improves the convenience of processing data in which values ​​have been entered for each of multiple predetermined items.

[0114] (Paragraph 2) The data processing method relating to Paragraph 2 is the data processing method relating to Paragraph 1, wherein the large-scale language model has learned predetermined words and their meanings, which are included in a pre-prepared library.

[0115] The data processing method described in paragraph 2 can correctly process abbreviations and colloquialisms commonly used within the industry using a large-scale language model.

[0116] (Paragraph 3) The data processing method relating to Paragraph 3 is the data processing method relating to Paragraph 1 or Paragraph 2, wherein the large-scale language model is located within a private network.

[0117] The data processing method described in paragraph 3 can prevent confidential internal information of companies and other organizations from being leaked to the outside.

[0118] (Paragraph 4) The data processing method relating to Paragraph 4 is the data processing method relating to Paragraph 3, wherein the large-scale language model has further learned the design information.

[0119] The data processing method described in paragraph 4 uses a large-scale language model that has learned the structure and attributes of the database, thereby enabling the output of processing keys that can more accurately identify tables and columns.

[0120] (Paragraph 5) The data processing method relating to Paragraph 5 is a data processing method relating to any of Paragraphs 1 to 4, further comprising trigger information which executes data processing for a predetermined second data item when input is made to a predetermined first data item in the design information.

[0121] The data processing method described in paragraph 5 allows information entered by the user into the first data item to be automatically reflected in the second data item.

[0122] (Paragraph 6) The data processing method relating to Paragraph 6 is a data processing method relating to any of Paragraphs 1 to 5, wherein the design information includes conversion information for converting a value entered for a predetermined first data item into an input value for a predetermined second data item in accordance with a predetermined standard, and when a value is entered for the first data item, the input value obtained by converting the entered value based on the conversion information is entered as the value of the second data item.

[0123] The data processing method described in paragraph 6 allows for the automatic determination of importance rankings and accuracy values ​​from text input by the user in natural language and output from a large-scale language model.

[0124] (Clause 8) The data processing system relating to paragraph 8 is characterized in that, in the data processing system relating to paragraph 7, the processing execution unit outputs an instruction statement written in a predetermined programming language that executes the processing of the specified content.

[0125] The data processing system described in paragraph 8 can be used by incorporating it into an existing database.

[0126] (Clause 10) A first embodiment of another aspect of the present invention is a data processing method in which a computer performs processing on a first database having target data in which values ​​are entered for each of at least one of a plurality of first data items, and a second database having target data in which values ​​are entered for each of at least one of a plurality of second data items, characterized in that a vectorization model is prepared which is constructed by predetermined machine learning and outputs input words as vectors, information that identifies each of the plurality of first data items and information that identifies each of the plurality of second data items is vectorized using the vectorization model, the vectors of information that identify the plurality of first data items are compared with the vectors of information that identify each of the plurality of second data items, and at least one of the plurality of first data items is associated with one of the plurality of second data items based on their similarity.

[0127] (Clause 12) A second embodiment of another aspect of the present invention is a data processing system used to perform processing on a second database having target data in which values ​​are entered for each of at least one of a plurality of first data items, and a second database having target data in which values ​​are entered for each of at least one of a plurality of second data items, the system comprising: a vectorization model storage unit that stores a vectorization model constructed by predetermined machine learning that vectorizes and outputs input words; a vectorization processing unit that uses the vectorization model to vectorize information that identifies each of the plurality of first data items and information that identifies each of the plurality of second data items; and a data item comparison processing unit that compares the vectors of the information that identifies the plurality of first data items with the vectors of the information that identifies each of the plurality of second data items, and associates at least one of the plurality of first data items with one of the plurality of second data items based on their similarity.

[0128] (Clause 13) A third embodiment of another aspect of the present invention is a data processing program that causes a computer to perform processing on a first database having target data in which values ​​are entered for each of at least one of a plurality of first data items, and a second database having target data in which values ​​are entered for each of a plurality of second data items, the data processing program characterized in that the computer in a system having a vectorization model storage unit that stores a vectorization model constructed by predetermined machine learning that vectorizes and outputs input words is configured to function as a vectorization processing unit that vectorizes information identifying each of the plurality of first data items and information identifying each of the plurality of second data items using the vectorization model, and as a data item comparison processing unit that compares the information identifying each of the plurality of first data items with information identifying each of the plurality of second data items and associates at least one of the plurality of first data items with any of the plurality of second data items based on their similarity.

[0129] The data processing method described in paragraph 10, the data processing system described in paragraph 12, and the data processing program described in paragraph 13 are used to unify (or integrate into one) two different databases. In the data processing method described in paragraph 10, the data processing system described in paragraph 12, and the data processing program described in paragraph 13, as in the embodiments described above, a large-scale language model and a vectorization model are prepared in advance.

[0130] In the data processing method described in paragraph 10, the data processing system described in paragraph 12, and the data processing program described in paragraph 13, information identifying multiple first data items in the first database and information identifying multiple second data items in the second database are vectorized. Then, the information identifying multiple first data items is compared with the information identifying multiple second data items, and based on their similarity, at least one of the multiple first data items is associated with one of the multiple second data items. As a result, multiple first data items contained in the first database are made common with second data items in the second database, and processing for two different databases can be performed in the same way as processing for a single database.

[0131] (Clause 11) The data processing method relating to paragraph 11 is characterized in that, in the data processing method relating to paragraph 10, the attribute information of each value of the plurality of first data items and the attribute information of each value of the plurality of second data items are vectorized, and when the attribute information of the value of the first data item differs from the attribute information of the value of the second data item to which the first data item is associated, the value of the first data item is converted to match the attribute information of the value of the second data item.

[0132] Even if a second data item corresponding to a first data item in the first database exists in the second database, the attribute information of the value (the input format of the record value) may differ from database to database. The data processing method described in Section 11 can improve convenience by standardizing the attribute information of the data item values, even in such cases.

[0133] 1...Data processing system 10...Business system 11...Sales support system 11D...Customer management database 12...Sales management system 13...Payroll calculation system 14...Personnel management system 20...Data processing device 21...Storage unit 211...Design information storage unit 212...Tuning data storage unit 213...Large-scale language model storage unit 214...Vectorization model storage unit 215...Input form storage unit 30...Vectorization processing unit 31...Design information creation unit 32...Tuning execution unit 33...Processing key acquisition unit 34...Key vector generation unit 35...Processing content identification unit 36...Processing command generation unit 37...Processing execution unit 38...Data item comparison processing unit 39...Record value conversion unit 51...Input unit 52...Display unit 6, 7, 8...Input support screen 61...Free word input field 62...Search support area 71, 82...Status report field 72...Input support area 81...Selection item field

Claims

1. A data processing method in which a computer performs processing on target data in which values ​​are entered in at least one of multiple data items of a database, comprising: preparing a large-scale language model constructed by predetermined machine learning that receives text data as input and outputs a key word for data processing; preparing a vectorization model constructed by predetermined machine learning that vectorizes and outputs the input word; using the vectorization model, vectorizing information that identifies each of the multiple data items and attribute information of the values ​​of each of the multiple data items; storing design information including information that identifies each of the vectorized multiple data items and attribute information of the values ​​of each of the vectorized multiple data items in a design information storage unit; having a user input text data representing the content of processing on the target data; inputting the text data representing the content of processing into the large-scale language model to output a key word for processing from the large-scale language model; and generating a key vector by inputting the word output from the large-scale language model into the vectorization model. A data processing method characterized by comparing the key vector with a vector of information that identifies each of the plurality of data items and a vector of attribute information of the values ​​of each of the plurality of data items, identifying the data items and processing content of the target data based on the similarity with the key vector, and performing the processing of the identified content with respect to the identified data items.

2. The data processing method according to claim 1, wherein the large-scale language model has learned predetermined words and their meanings, which are included in a pre-prepared library.

3. The data processing method according to claim 1, wherein the large-scale language model is located within a private network.

4. The data processing method according to claim 3, wherein the large-scale language model has further learned the design information.

5. The data processing method according to claim 1, further comprising trigger information which performs data processing on a predetermined second data item when input is made to a predetermined first data item.

6. The data processing method according to claim 1, further comprising conversion information for converting a value entered for a predetermined first data item into an input value for a predetermined second data item in accordance with a predetermined standard, wherein when a value is entered for the first data item, the input value obtained by converting the entered value based on the conversion information is entered as the value for the second data item.

7. A data processing system used to perform processing on target data in which values ​​are entered for at least one of multiple data items in a database, comprising: a large-scale language model storage unit that stores a large-scale language model constructed by predetermined machine learning that receives text data as input and outputs a key word for data processing; a vectorization model storage unit that stores a vectorization model constructed by predetermined machine learning that vectorizes and outputs an input word; a design information storage unit that stores design information including information that identifies each of the multiple data items vectorized using the vectorization model, and attribute information of the values ​​of each of the multiple data items vectorized using the vectorization model; a processing key acquisition unit that causes a user to input text data representing the content of processing on the target data, and inputs the text data representing the content of processing into the large-scale language model to output a key word for the processing from the large-scale language model; and a key vector generation unit that generates a key vector by inputting the word output from the large-scale language model into the vectorization model. A data processing system comprising: a processing content identification unit that compares the key vector with a vector of information that identifies each of the plurality of data items and a vector of attribute information of the values ​​of each of the plurality of data items, and identifies the data items and processing content of the target data based on the similarity with the key vector; and a processing execution unit that performs the processing of the identified content with respect to the identified data items.

8. The data processing system according to claim 7, characterized in that the processing execution unit outputs an instruction statement written in a predetermined programming language for executing the processing of the specified content.

9. A data processing program for causing a computer to perform processing on target data in which values ​​have been entered into at least one of several data items in a database, comprising: a large-scale language model storage unit that stores a large-scale language model constructed by predetermined machine learning that receives text data as input and outputs a key word for data processing; a vectorization model storage unit that stores a vectorization model constructed by predetermined machine learning that vectorizes and outputs an input word; a design information storage unit that stores design information including information that identifies each of the multiple data items vectorized using the vectorization model, and attribute information of the values ​​of each of the multiple data items vectorized using the vectorization model, wherein the computer in the system comprises: a processing key acquisition unit that causes the user to input text data representing the content of processing on the target data, inputs the text data representing the content of processing into the large-scale language model, and outputs a key word for processing from the large-scale language model; and a key vector generation unit that generates a key vector by inputting the word output from the large-scale language model into the vectorization model. A data processing program characterized by having a processing content identification unit that compares the key vector with a vector of information that identifies each of the plurality of data items and a vector of attribute information of the values ​​of each of the plurality of data items, and identifies the data items and processing content of the target data based on the similarity with the key vector, and a processing execution unit that performs the processing of the identified content with respect to the identified data items.

10. A data processing method in which a computer performs processing on a first database having target data in which values ​​are entered for each of at least one of a plurality of first data items, and a second database having target data in which values ​​are entered for each of a plurality of second data items, the method comprising: preparing a vectorization model constructed by predetermined machine learning that vectorizes and outputs input words; using the vectorization model, vectorizing information that identifies each of the plurality of first data items and information that identifies each of the plurality of second data items; comparing the vectors of information that identify the plurality of first data items with the vectors of information that identify each of the plurality of second data items, and associating at least one of the plurality of first data items with one of the plurality of second data items based on their similarity.

11. The data processing method according to claim 10, further characterized by vectorizing the attribute information of each value of the plurality of first data items and the attribute information of each value of the plurality of second data items, and converting the value of the first data item to match the attribute information of the value of the second data item when the attribute information of the value of the first data item differs from the attribute information of the value of the second data item to which the first data item is associated.

12. A data processing system used to perform processing on a first database having target data in which values ​​are entered for each of at least one of a plurality of first data items, and a second database having target data in which values ​​are entered for each of a plurality of second data items, comprising: a vectorization model storage unit that stores a vectorization model constructed by predetermined machine learning that vectorizes and outputs input words; a vectorization processing unit that vectorizes information that identifies each of the plurality of first data items and information that identifies each of the plurality of second data items using the vectorization model; and a data item comparison processing unit that compares the vectors of information that identify the plurality of first data items with the vectors of information that identify each of the plurality of second data items, and associates at least one of the plurality of first data items with one of the plurality of second data items based on their similarity.

13. A data processing program that causes a computer to perform processing on a first database having target data in which values ​​are entered for each of at least one of a plurality of first data items, and a second database having target data in which values ​​are entered for each of a plurality of second data items, wherein the computer in a system having a vectorization model storage unit that stores a vectorization model constructed by predetermined machine learning that vectorizes input words and outputs them, functions as a vectorization processing unit that vectorizes information that identifies each of the plurality of first data items and information that identifies each of the plurality of second data items using the vectorization model, and a data item comparison processing unit that compares the vectors of information that identify the plurality of first data items with the vectors of information that identify each of the plurality of second data items, and associates at least one of the plurality of first data items with one of the plurality of second data items based on their similarity.

Images