Program generation system, program generation method, and program code generation program

Abstract

Even those without programming skills will be made to create conversion programs. [Solution] The program generation system S is a system for generating a data conversion program for converting unnormalized data to normalized data ND, and comprises an acquisition unit and a generation unit. The acquisition unit acquires data correspondence definition data including the item names of the data to be converted, the data positions of the items, and natural language descriptions indicating the conversion rules, as well as sample data for identifying the physical structure of the data to be converted. The generation unit uses a neural network model to interpret the correspondence definition data into specific data extraction commands based on the physical structure of the sample data, and generates program code that becomes a data conversion program.

Description

Technical Field

[0001] The present invention relates to a converter maker for generating a converter that converts data.

Background Art

[0002] In recent years, in the apparel industry and manufacturing retail (SPA: Specialty store retailer of Private label Apparel), in order to improve the efficiency of the supply chain, the introduction of systems such as product lifecycle management (PLM) systems has been progressing. As a result, order information and product specifications (specification documents) have come to be managed as electronic data. However, the data formats exchanged between companies or systems are not unified, and unstructured data or CSV / Excel data with unique layouts are still mainstream. As an invention related to data conversion, for example, there is the invention described in Patent Document 1, but it does not consider customer data having a wide variety of formats.

[0003] Conventionally, in order to convert customer data having such a wide variety of formats into normalized data that can be imported by an order receiving and issuing system (for example, a tag label issuing system), one of the following methods has been adopted, but each has problems.

[0004] First, as the first method, there is manual transcription. In this method, the person in charge performs copy & paste on a spreadsheet. This method has problems such as being a breeding ground for human error and requiring a great deal of labor in the apparel business that deals with a huge number of SKUs (Stock Keeping Unit).

[0005] The second approach involves developing a dedicated conversion program (converter). In this method, a skilled programmer designs the conversion logic for each customer and implements a dedicated program using scripting languages ​​or ETL tools. While this method offers high reliability, it has the drawback of requiring several weeks to several months for development. Furthermore, it suffers from the problem of "personalization," meaning it cannot be handled by field staff (such as sales assistants) who lack programming skills. [Prior art documents] [Patent Documents]

[0006] [Patent Document 1] Patent No. 6776435 [Overview of the project] [Problems that the invention aims to solve]

[0007] Therefore, the problem that the present invention aims to solve is to provide a program generation system that enables even those without programming skills to create a conversion program that converts unnormalized data into normalized data that can be imported by an ordering system. [Means for solving the problem]

[0008] To solve the above problems, the program generation system according to the present invention is a program generation system that generates a data conversion program for converting unnormalized data to normalized data, An acquisition unit that acquires data correspondence definition data including the item names of the data to be converted, the data location of the items, and natural language descriptions indicating the conversion rules, as well as sample data for identifying the physical structure of the data to be converted. The system includes a generation unit that uses a neural network model to interpret correspondence definition data based on the physical structure of sample data into specific data extraction commands, thereby generating program code that becomes a data conversion program.

[0009] The above program generation system preferably, The program further comprises a constraint section that specifies technical constraint information, which is the specific data processing library and data transformation algorithm that the generated program should use. The generation unit generates program code by using a neural network model and applying data processing logic that conforms to the technical constraints specified by the constraint unit.

[0010] The above program generation system preferably, The interpretation process in the generation unit includes matching a relative position specification for a specific string contained in the correspondence definition data or a search specification using a keyword contained in the item name with a physical row number, column index, or string pattern in the sample data, and determining a unique data acquisition function in the generated program code.

[0011] The above program generation system preferably, The technical constraint information includes an instruction to execute a dimensional transformation process that expands data described in matrix format into record data in list format. The generation unit generates a code block that applies a specific function specified in the technical constraint information to multiple attribute item groups specified by the correspondence definition data.

[0012] The above program generation system preferably, The correspondence definition data includes information extraction rules from multi-attribute mixed data where multiple attribute information is mixed within a single data field. The generation unit analyzes the text patterns of the multi-attribute mixed data contained in the sample data, and generates code blocks that decompose and store the multi-attribute mixed data into individual attribute fields using pattern matching with regular expressions, splitting using standard string processing, or semantic analysis using natural language processing, according to the technical constraint information.

[0013] The above program generation system preferably, The correspondence definition data includes information extraction rules from multi-attribute mixed data where multiple attribute information is mixed within a single data field. The generation unit is, The system is configured to analyze the text patterns of the multi-attribute mixed data contained in the sample data, select a method from a predetermined set of different information extraction methods according to the text patterns, and generate a code block that decomposes the multi-attribute mixed data into individual attribute fields and stores them, in accordance with technical constraint information. In analyzing text patterns, if there are repetitions of specific symbols or patterns, pattern matching using regular expressions is selected; if there are clear delimiters, splitting using standard string processing is selected; and for natural language texts without delimiters, semantic analysis using natural language processing is selected.

[0014] The above program generation system preferably, It further includes a verification unit that verifies the consistency between the correspondence definition data and the sample data. The correspondence definition data includes search criteria. The verification unit generates a proposed correction to correct errors in the correspondence definition data based on feature analysis of the sample data, if the search conditions specified by the correspondence definition data do not exist in the sample data, or if the data format is inconsistent.

[0015] The program generation system described above is, for example, a language model in which a neural network model is located within a local environment isolated from the external network.

[0016] The program generation system preferably further comprises a definition data creation unit that generates at least a portion of correspondence definition data for associating specific data sequences contained in sample data with items in normalized data by analyzing both or either the physical structure and data content characteristics of the sample data.

[0017] In order to solve the above problems, a program generation method according to the present invention is a program generation method for generating a data conversion program by a computer to convert non-normalized data into normalized data, by a computer, a step of obtaining correspondence definition data between data including the item name of the data to be converted, the data position of the item, and a description in natural language indicating a conversion rule, and sample data for specifying the physical structure of the data to be converted; a step of using a neural network model to interpret the correspondence definition data based on the physical structure of the sample data into specific data extraction instructions and generating program code to be a data conversion program.

[0018] In order to solve the above problems, a program code generation program according to the present invention causes a computer to execute the above program generation method.

Advantages of the Invention

[0019] According to the program generation system according to the present invention, even a person without programming skills can create a conversion program.

Brief Description of the Drawings

[0020] [Figure 1] It is a schematic overall view according to an embodiment of the program generation system of the present invention. [Figure 2] It is a diagram showing an example of items related to an order system. [Figure 3] It is a block diagram of the converter maker shown in FIG. 1. [Figure 4] It is a diagram showing an example of correspondence definition data. [Figure 5] It is a flowchart showing the flow of the program generation system.

Modes for Carrying Out the Invention

[0021] Hereinafter, an embodiment of the converter maker according to the present invention and a program generation system equipped with the converter maker will be described with reference to the attached figures. In this embodiment, the ordering system 100 is an ordering system 100 for customers to order apparel-related auxiliary materials, but this is merely an example, and the ordering system according to the present invention is not limited to an ordering system for apparel-related auxiliary materials.

[0022] Figure 1 is a schematic diagram showing the overall configuration of one embodiment of the program generation system S of the present invention. As shown in Figure 1, the program generation system S according to this embodiment is configured to generate a conversion program related to a converter 4 for normalizing unnormalized order data OD (hereinafter referred to as "unnormalized data DD") provided from the customer system 200 into a data format that can be read by the order system 100, using a data conversion program. Here, "unnormalized data DD" refers to order data OD created by each unspecified customer in their own unique format (form). In addition, in the present invention, "normalized data ND" refers to data whose format has been standardized so that a specific order system 100 can properly read and process it.

[0023] Figure 2 shows an example of items handled by the ordering system 100. As shown in Figure 2, these items mainly consist of items related to apparel-related auxiliary materials. Items in the ordering system 100 include project and management hierarchy information, which can include company identification information, management hierarchy information, and product identifier information. Product specification information can include product attribute information, manufacturer information, and country of origin information. Logistics and commercial flow information can include logistics base information and transaction settlement information. Furthermore, information on the specifications of the materials (tags and labels) themselves can include auxiliary material specification information and media type information. SKU (stock keeping unit) and quantity information can include color attribute information, size attribute information, dimensional standard information, and order quantity information. Quality label (care label) printing content information can include handling symbol information, maintenance instruction information, material composition information, mixing ratio information, and supplementary quality information. System control and output option information can include notification setting information, optional additional information, and output control parameters. Thus, the number of items related to apparel accessories is very large, and the information entered into these items can be described using a variety of expressions.

[0024] The program generation system S according to this embodiment includes a definition data creation terminal 1 and a converter maker 2 for creating a correspondence definition data CD, which will be described later, in an on-premise environment (internal network). Figure 3 is a block diagram showing the internal configuration of the converter maker 2 shown in Figure 1.

[0025] <Definition data creation terminal> The definition data creation terminal 1 is a terminal having a computer, the computer having a storage device, an arithmetic unit, and memory. The definition data creation terminal 1 is configured to receive sample data SD, which is a sample of denormalized data DD, from the customer system 200. The storage device of the definition data creation terminal 1 stores a definition data creation program that causes the computer to operate as a definition data creation unit 10.

[0026] The definition data creation unit 10 generates at least a portion of the correspondence definition data CD for associating specific data sequences contained in the sample data SD with items in the normalized data ND by analyzing at least one of the physical structure and data content characteristics of the sample data SD. The definition data creation unit 10 may generate the correspondence definition data CD using, for example, a neural network model.

[0027] Figure 4 shows an example of a correspondence definition data CD. The correspondence definition data CD is a dataset expressed in tabular form that defines rules for converting input data, which is unnormalized data (DD), to normalized data (ND), and each cell in the table is described in natural language. The correspondence definition data CD is composed of the following elements in pairs.

[0028] There are three main configuration elements for performing the conversion. First, there is the "item (name)" of the ordering system 100, which defines the data structure of the target data. Second, there is the "data location specification," which identifies where the data to be converted is located within the input data. This includes identification by matching with keywords in the header, relative position from a specific indicator, or by characteristics or ranges of values ​​(size item groups, etc.). Third, there are the "processing rules" applied to the extracted data. These include format unification and partial extraction, as well as structural conversion (Unpivot) which expands matrix data into list format, and attribute decomposition which splits complex information within a single item using regular expressions, etc.

[0029] As described later, the generation unit 30 according to the present invention compares (grounds) the correspondence definition data CD with the sample data SD. This resolves ambiguity in item names and generates executable program code including conditional branching. As a result, semantic instructions independent of physical coordinates become possible, providing high versatility to flexibly absorb differences in data formats for each customer, and enabling the definition of complex transformations such as matrix expansion without specialized knowledge.

[0030] The neural network model used by the definition data creation unit 10 may, for example, be a pre-trained model that has learned the relationships between each item of the ordering system 100 and various denormalized data DDs. Alternatively, the neural network model used by the definition data creation unit 10 may consist of a so-called language model, such as a large-scale language model (LLM) or a small-scale language model (SLM). In this case, the LLM (SLM) is configured as a local LLM (SLM) in an on-premise environment and is configured to prevent customer confidential data from being leaked to the outside.

[0031] The portion of the correspondence definition data CD that is not generated by the definition data creation unit 10 is created by the user. In this case, the user uses the definition data creation terminal 1 to refer to the sample data SD and describes the data location and processing rules in natural language for each item of the ordering system 100. The correspondence definition data CD does not need to be written in a programming language, but in the natural language that the user uses on a daily basis, so it can be handled even by field staff (sales assistants, etc.) who do not have programming skills.

[0032] <Converter Manufacturer> Converter Maker 2 is a terminal having a computer, the computer having a storage device, an arithmetic unit, and memory. The storage device stores a program code generation program that causes the computer to operate as an acquisition unit 22, a prompt construction unit 24, a verification unit 26, a constraint unit 28, and a generation unit 30.

[0033] Next, the functional configuration of Converter Maker 2 will be described. As shown in Figure 3, Converter Maker 2 comprises a storage unit 20, an acquisition unit 22, a prompt construction unit 24, a verification unit 26, a constraint unit 28, and a generation unit 30.

[0034] The memory unit 20 stores the prompt body of the prompts that are input to the neural network model.

[0035] The neural network model is composed of so-called language models, such as large-scale language models (LLMs) or small-scale language models (SLMs). This neural network model may be shared with, for example, the neural network model used by the definition data creation terminal 1. In any case, the neural network model is located in an on-premise environment / corporate network. The neural network model may also be owned by the converter manufacturer 2.

[0036] The prompt body includes prompts for defining professional roles and assigning domain knowledge bias. Defining professional roles and assigning domain knowledge bias refers to prompts that, for example, make a general-purpose language model act as an expert specializing in a specific domain (in this embodiment, apparel data processing), prioritizing the activation of knowledge related to that specific domain, and fixing the quality and expertise of the generated code.

[0037] Furthermore, the memory unit 20 stores prompts for the converter maker 2 to operate as the verification unit 26, constraint unit 28, and generation unit 30.

[0038] The acquisition unit 22 acquires correspondence definition data CD and sample data SD for identifying the physical structure of the data to be converted. The acquired data is sent to the prompt construction unit 24.

[0039] In this embodiment, the prompt construction unit 24 introduces the correspondence definition data CD and sample data SD as context into the prompt body and constructs them as a single prompt. The constructed prompt is then input to the neural network model.

[0040] The verification unit 26 generates a proposed correction to correct errors in the correspondence definition data CD based on the feature analysis of the sample data SD if the "search conditions" specified by the correspondence definition data CD do not exist in the sample data SD, or if the data format is inconsistent.

[0041] "Search criteria" are specific "clues" that a neural network model uses to find and identify target data from among the many data points included in the sample data (SD). Specifically, search criteria include specifying by a particular string, specifying by relative position / range, and specifying by data features / patterns.

[0042] Specifying by a specific string refers to the "item name" or "label" itself (for example, "order date," "product number," "color," etc.) that should be clearly stated within the data.

[0043] Specifying by relative position or range refers to instructions that limit the physical location or the area to be searched. Examples include instructions such as "the cell to the right of the one that says ~", "the column in the header row that contains ~", or "the signature field at the bottom of the file".

[0044] Furthermore, specifying by data characteristics and patterns means that a column is identified not by its header name, but by the characteristics of the values ​​within it. For example, this could be "XS to XL" or "numeric only" (identifying the size column), or "JIS / ISO code" (identifying the pictogram column).

[0045] The case where the "search condition" does not exist in the sample data SD is, for example, when the user instructs the system to "find 'Date:'" (search condition), but the actual sample data SD only contains numerical data such as "2025,039..." and the string "Date:" cannot be found anywhere (Grounding failure). In this case, the verification unit 26 may perform feature analysis of the sample data SD to generate a suggested correction such as "The search condition (Date:) was not found, but a column with a data format similar to 'YYYY / MM / DD' was detected," or it may modify the corresponding part of the correspondence definition data CD.

[0046] The generated revised draft and the revised correspondence definition data CD may be sent to the definition data creation terminal 1. The user can review the revised draft and modify the correspondence definition data CD, or review the revised correspondence definition data CD. The revised correspondence definition data CD is then acquired again by the acquisition unit 22.

[0047] The constraint section 28 specifies technical constraint information, which includes specific data processing libraries (e.g., pandas, polars, numpy, re, json, etc.) and data transformation algorithms (e.g., structure transformation algorithms such as melt and stack, as well as string and semantic analysis algorithms) that the generated program should use. In other words, the constraint section 28 restricts the writing rules of the generated code. By thus restricting the data processing libraries used by the generated program to libraries that can efficiently handle tabular data, avoid row-by-row loop processing, and perform CSV reading and writing, as well as data extraction and processing at high speed, even language models with limited computing resources, such as local LLMs (SLMs), can output standardized code that is highly maintainable and has few errors.

[0048] The technical constraint information provided by the constraint unit 28 may include an execution instruction for a dimensional transformation process that expands data described in matrix format into record data in list format. Specifically, this dimensional transformation process converts data arranged in two dimensions (or possibly multiple dimensions) of "rows x columns," such as a cross-tabulation table, into a vertically oriented list table where one record is placed in each row. This transforms the structure into a format that is easier to use for analysis and aggregation.

[0049] The generation unit 30 uses a neural network model to interpret the correspondence definition data CD based on the physical structure of the sample data SD into specific data extraction commands, thereby generating program code that will become a data conversion program. Furthermore, while using the neural network model, the generation unit 30 applies data processing logic according to the technical constraint information specified by the constraint unit 28 to generate program code.

[0050] The interpretation process in the generation unit 30 includes comparing a relative position specification for a specific string contained in the correspondence definition data CD or a search specification using a keyword contained in the item name with the physical row number, column index, or string pattern in the sample data SD, and determining a unique data acquisition function in the program code to be generated.

[0051] A keyword-based search specification is, for example, a specification that states, "a column containing either 'Style,' 'Item,' or 'Product Number'," and matching is the process of searching for (matching) the string pattern within the sample data SD according to this instruction. In other words, the generation unit 30 is configured to bridge the gap between (human) instructions and (real) actual data by interpreting the correspondence definition data CD and matching the sample data SD according to its instructions. The process of determining a unique data acquisition function in the generated program code is to generate a strict code that produces the same result regardless of who executes it. This process may include selecting which function to use to always uniquely acquire the same element.

[0052] Furthermore, the generation unit 30 may analyze the text pattern of the multi-attribute mixed data contained in the sample data SD, and, in accordance with the technical constraint information, select the optimal method from a predetermined set of different information extraction methods that corresponds to the text pattern, without depending on a specific technology (such as regular expressions only), to generate a code block that decomposes and stores the multi-attribute mixed data into individual attribute fields. Specifically, even in cases where multiple meanings (items) are packed into a single cell, such as "Quality: 100% cotton / polyester" or "Pictogram: Hand wash, medium iron," the generation unit 30 analyzes the character pattern by interpreting it using a neural network model. Then, from among the multiple methods presented by the constraint unit 28, it selects regular expressions (e.g., re.search) if there is repetition of a specific symbol or pattern, standard string processing (e.g., pandas.Series.str.split) if there is clear delimiter such as a comma or slash, and semantic analysis (e.g., NLP libraries such as spaCy or JSON format extraction from LLM) if it is a natural language sentence without delimiters. This allows for the accurate and effortless organization of diverse customer data into the format required by the company's internal systems, without requiring any prior knowledge.

[0053] A data conversion program is generated by the program code generated by the generation unit 30. The created data conversion program functions as a data converter 4, so as shown in Figure 1, when order data OD (denormalized data DD) is input from the customer system 200, it can convert the order data OD into normalized data ND. This allows customers (ordering parties) to place orders using data they manage themselves without having to process the data for the order system 100. In addition, users (contractors) can reduce the work of creating auxiliary material order forms to almost zero by the program generation system S. In this way, the program generation system S realizes close data linkage to eliminate transcription work.

[0054] Next, we will explain the operation of the program generation system S again, referring to the flowchart in Figure 5.

[0055] (1) The program generation system S generates at least a portion of the correspondence definition data CD by analyzing both or either the physical structure and data content characteristics of the sample data SD using the definition data creation unit 10. The remaining correspondence definition data CD not generated by the definition data creation unit 10 is created by the user (see S1 in Figure 5).

[0056] (2) Next, the program generation system S acquires the correspondence definition data CD and sample data SD using the acquisition unit 22 (see S2 in Figure 5), and verifies the consistency between the correspondence definition data CD and the sample data SD using the verification unit 26 (see S3 in Figure 5).

[0057] (3) Next, if the program generation system S finds that the correspondence definition data CD and the sample data SD are inconsistent or contradictory as a result of the verification (No. in S4 of Figure 5), it generates a revised version (see S5 in Figure 5). On the other hand, if the program generation system S finds that the correspondence definition data CD and the sample data SD are consistent (Yes in S4 of Figure 5), the prompt construction unit 24 constructs a prompt that includes the correspondence definition data CD and the sample data SD as context (see S6 in Figure 5).

[0058] (4) Next, the program generation system S takes the constructed prompt as input to the neural network model, interprets the correspondence definition data CD into specific data extraction commands (see S7 in Figure 5), and generates program code by applying data processing logic according to the technical constraint information specified by the constraint unit 28 (see S8 in Figure 5).

[0059] As a result, the program generation system S allows even those without programming skills to create conversion programs. Furthermore, by pre-verifying the consistency between the correspondence definition data CD and the sample data SD, it can generate an appropriate conversion program. Moreover, if such consistency is not found, the system generates a proposed revision of the correspondence definition data CD, allowing the user to easily modify the CD. In addition, the program generation system S can also assist in the creation of the correspondence definition data CD by analyzing both or either the physical structure and data content characteristics of the sample data SD, thereby generating at least a portion of the correspondence definition data CD.

[0060] Furthermore, because the program generation system S is configured in a local environment, it prevents the leakage of confidential customer data to external parties. Moreover, by instructing the program generation system S on technical constraint information, which consists of data processing libraries and data transformation algorithms, it can output highly maintainable, error-free, standardized code even with language models that have limited computing resources, such as local LLM (SLM).

[0061] Although one embodiment of the program generation system S, program generation method, and program code generation program according to the present invention has been described above, the program generation system S, program generation method, and program code generation program according to the present invention are not limited thereto and may be implemented by the following modifications.

[0062] <Variation> The neural network model related to the program generation system S of the present invention does not necessarily have to be configured locally.

[0063] • Converter maker 2 does not necessarily have to be configured separately from definition data creation terminal 1, and may also serve as definition data creation terminal 1.

[0064] The program generation system S does not necessarily have to introduce the correspondence definition data CD and sample data SD at the prompt. In this case, the correspondence definition data CD and sample data SD only need to be input to the neural network model as sources. [Explanation of symbols]

[0065] S Program Generation System OD Order Data ND Normalized Data DD (Denormalized Data) SD Sample Data CD correspondence relationship definition data 1. Definition data creation terminal 10 Definition Data Creation Unit 2 Converter Manufacturers 20 Memory section 22 Acquisition Department 24 Prompt Construction Unit 26 Verification Department 28 Constraint part 30 Generation part 4 Converters 100 Ordering System 200 Customer Systems

Claims

1. A program generation system that generates a data conversion program for converting denormalized data into normalized data, An acquisition unit that acquires data correspondence definition data including the item names of the data to be converted, the data location of the items, and a natural language description indicating the conversion rules, as well as sample data for identifying the physical structure of the data to be converted. A program generation system comprising: a generation unit that uses a neural network model to interpret the correspondence definition data into specific data extraction commands based on the physical structure of the sample data, and generates program code that becomes the data conversion program.

2. The program code generated further comprises a constraint section that specifies technical constraint information, which is a specific data processing library and data transformation algorithm to be used. The program generation system according to claim 1, wherein the generation unit generates the program code by using the neural network model and further applying data processing logic in accordance with the technical constraint information specified by the constraint unit.

3. The program generation system according to claim 1, wherein the interpretation processing in the generation unit includes comparing a relative position specification for a specific string included in the correspondence definition data or a search specification using a keyword included in the item name with a physical row number, column index, or string pattern in the sample data, and determining a unique data acquisition function in the program code to be generated.

4. The aforementioned technical constraint information includes an execution instruction for a dimensional transformation process that expands data described in matrix format into record data in list format. The program generation system according to claim 2, wherein the generation unit generates a code block that applies a specific function specified in the technical constraint information to a plurality of attribute item groups specified by the correspondence relationship definition data.

5. The aforementioned correspondence definition data includes information extraction rules from multi-attribute mixed data in which multiple attribute information is mixed within a single data field. The program generation system according to claim 2, wherein the generation unit analyzes the text pattern of the multi-attribute mixed data contained in the sample data, and generates code blocks that decompose and store the multi-attribute mixed data into individual attribute fields by pattern matching using regular expressions, splitting using standard string processing, or semantic analysis using natural language processing, in accordance with the technical constraint information.

6. The aforementioned correspondence definition data includes information extraction rules from multi-attribute mixed data in which multiple attribute information is mixed within a single data field. The generating unit is The system is configured to analyze the text pattern of the multi-attribute mixed data contained in the sample data, select a method from a predetermined set of different information extraction methods according to the text pattern in accordance with the technical constraint information, and generate a code block that decomposes the multi-attribute mixed data into individual attribute fields and stores them. The program generation system according to claim 2, wherein, in the analysis of the text pattern, pattern matching using regular expressions is selected when there is repetition of specific symbols or patterns, splitting by standard string processing is selected when there is a clear delimiter, and semantic analysis using natural language processing is selected when there is a natural language sentence without a delimiter.

7. The system further includes a verification unit that verifies the consistency between the correspondence definition data and the sample data. The aforementioned correspondence definition data includes search conditions. The program generation system according to claim 1, wherein the verification unit generates a proposed correction for correcting errors in the correspondence definition data based on a feature analysis of the sample data when the search conditions specified by the correspondence definition data do not exist in the sample data or the data format is inconsistent.

8. The program generation system according to claim 1, wherein the neural network model is a language model located in a local environment isolated from an external network.

9. The program generation system according to claim 1, further comprising a definition data creation unit that generates at least a portion of the correspondence relationship definition data for associating a specific data sequence contained in the sample data with an item in the normalized data by analyzing both or either the physical structure and the characteristics of the data content of the sample data.

10. A method for generating a program that generates a data conversion program for converting denormalized data into normalized data using a computer, According to the aforementioned computer, The steps include obtaining data that defines the correspondence between data, including the item names of the data to be converted, the data location of the item, and a natural language description indicating the conversion rule, and sample data for identifying the physical structure of the data to be converted. A program generation method comprising the steps of: using a neural network model to interpret the correspondence definition data into specific data extraction commands based on the physical structure of the sample data, thereby generating program code that constitutes the data conversion program.

11. A program code generation program that causes a computer to execute the program generation method described in claim 10.

Images