Information processing device, information processing method, and information processing program

Abstract

[Problem] To address the problem that intended information cannot be extracted from a large amount of file data groups including various kinds of information and various formats in a conventional technique. [Solution] In order to solve this problem, provided is an information processing device comprising: a first reception unit that receives, from a user, a setting of a main prompt for indicating one or more extraction targets from a plurality of pieces of file data including a plurality of pieces of information, the main prompt being input to a large-scale language model; a second reception unit that receives, from the user, a setting of a detailed prompt having one or a plurality of items, the detailed prompt being for setting in detail the indication for the extraction targets by the main prompt; and an extraction unit that inputs, to the large-scale language model, the main prompt according to the setting received by the first reception unit and the detailed prompt according to the setting received by the second reception unit, and extracts the extraction targets from the file data.

Description

Information processing device, information processing method, and information processing program

[0001] The present invention relates to an information processing apparatus that extracts desired information from a large amount of file data that includes a wide variety of information and has a variety of formats.

[0002] For example, file data such as documents and drawings used in the architecture, design, civil engineering, and plant industries contain a wide variety of information and come in a variety of formats. Examples include architectural drawings, regulation documents, construction management histories, design summaries, elevations, and detailed drawings. For example, a single piece of architectural design drawing data may contain hundreds of data items. The information contained in these file data is not neatly organized like "columns and rows" or "JSON format." This makes it unstructured, meaning it is not suitable for use in search, statistical processing, or as AI (artificial intelligence) learning data. Therefore, there is a demand for structuring this unstructured file data like a spreadsheet.

[0003] In order to structure the unstructured file data described above, it is necessary to extract individual pieces of information contained in the unstructured file data for each data item. A technology for extracting such information is described, for example, in Patent Document 1. Specifically, predetermined information is extracted from image data generated by scanning a paper document using OCR (optical character recognition) based on positional information on the paper surface defined for each item.

[0004] Meanwhile, in recent years, there has been progress in the development of technologies that utilize AI (artificial intelligence) to extract information. For example, there is a technology that reads invoices, converts them into text data, and automatically extracts information necessary for accounting journal processing (such as "invoice amount" and "invoice originator") (Non-Patent Document 1).

[0005] Japanese Patent Application Publication No. 2011-60219

[0006] https: / / cloud.watch.impress.co.jp / docs / news / 1349596.html

[0007] In the case of a technology such as that disclosed in Patent Document 1 that extracts characters present in an area predefined by position information, it works well for extracting characters from file data in a standard format. However, when the target file data has a wide variety of formats and data items included, it is necessary to customize the area settings and definitions for each piece of information included in the file data. Performing this customization for all of the wide variety of file data requires a huge amount of work for the user and is extremely difficult.

[0008] Furthermore, in the technology of Non-Patent Document 1, the information that is automatically extracted is information related to accounting journal processing, and is limited to information that has been registered in advance by the application provider (development company, etc.), so it is not possible to appropriately extract the information that the user desires from the wide variety of information contained in file data of various formats.

[0009] Therefore, in order to solve the above problem, the present invention provides the following information processing device etc. That is, the information processing device includes: a first reception unit that receives from a user a setting of a main prompt that is a prompt to be input into a large-scale language model and that indicates one or more extraction targets from a plurality of file data containing a plurality of pieces of information; a second reception unit that receives from the user a setting of a detail prompt that is a prompt for setting in detail the extraction target indicated by the main prompt and that has one or more items; a storage unit that stores the main prompt based on the setting received by the first reception unit and the detail prompt based on the setting received by the second reception unit; and an extraction unit that inputs the main prompt based on the setting received by the first reception unit and the detail prompt based on the setting received by the second reception unit and / or the main prompt and the detail prompt stored in the storage unit into the large-scale language model, and extracts extraction targets from the plurality of file data.

[0010] In addition to the above features, the present invention also provides an information processing device further having a conversion unit that converts the format of the file data in accordance with the characteristics of the file data, and the extraction unit extracts from the file data converted by the conversion unit.

[0011] In addition to the above features, the present invention provides an information processing device further comprising a verification unit that verifies the extraction result extracted by the extraction unit.

[0012] In addition to the above features, the present invention also provides an information processing device in which the verification unit performs verification using the received detailed prompt settings.

[0013] In addition to the above features, the information processing device provides an information processing device in which the extraction unit performs re-extraction while reflecting the verification result by the verification unit.

[0014] In addition to the above features, the present invention provides an information processing device further comprising a basis display unit that displays, for the extraction result extracted by the extraction unit, the basis for the extraction of the extraction result.

[0015] Also provided is an information processing method executed by a computer in an information processing device, the information processing method including: a first receiving step of receiving from a user a setting of a main prompt to be input to a large-scale language model, the main prompt indicating one or more extraction targets from a plurality of file data containing a plurality of pieces of information; a second receiving step of receiving from the user a setting of a detail prompt having one or more items, the detail prompt being a prompt for setting in detail the extraction target indicated by the main prompt; a storage step of storing the main prompt based on the setting received by the first receiving unit and the detail prompt based on the setting received by the second receiving unit; and an extraction step of inputting the main prompt based on the setting received in the first receiving step, the detail prompt based on the setting received in the second receiving step, and / or the main prompt and the detail prompt stored in the storage unit into the large-scale language model, and extracting the extraction targets from the plurality of file data.

[0016] The present invention also provides an information processing program that causes a computer to execute the following steps in an information processing device: a first receiving step of receiving from a user a setting for a main prompt that is a prompt to be input into a large-scale language model and that indicates one or more extraction targets from a plurality of file data containing a plurality of pieces of information; a second receiving step of receiving from the user a setting for a detail prompt that is a prompt for setting in detail the extraction target indicated by the main prompt and has one or more items; a storage step of storing the main prompt based on the setting received by the first receiving unit and the detail prompt based on the setting received by the second receiving unit; and an extraction step of inputting the main prompt based on the setting received in the first receiving step, the detail prompt based on the setting received in the second receiving step, and / or the main prompt and the detail prompt stored in the storage unit into a large-scale language model, and extracting the extraction targets from the plurality of file data.

[0017] According to the present invention, desired information can be extracted from a large amount of file data that includes a wide variety of information and has a variety of formats.

[0018] 1. Block diagram showing an example of the functional configuration of an information processing device of Example 1. 2. Conceptual diagram showing an example of converting PDF format file data into a text file with location information added. 3. Conceptual diagram showing another example of converting the format of file data. 4. Diagram showing an example of a detailed prompt. 5. Diagram showing how each setting item of the detailed prompt is applied. 6. Conceptual diagram showing an example of an input screen when setting a detailed prompt. 7. Conceptual diagram showing an example of the configuration of hardware that realizes the information processing device of Example 1. 8. Flow diagram showing an example of the processing flow of the information processing device of Example 1. 9. Block diagram showing an example of the functional configuration of an information processing device of Example 2. 10. Diagrams illustrating various settings when performing verification. 11. Block diagram showing an example of the functional configuration of an information processing device of Example 1. 12. Conceptual diagram showing an example of displaying a position on an image as evidence. 13. Conceptual diagram showing another aspect of data reduction.

[0019] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, the present invention should not be limited to these embodiments and can be embodied in various forms without departing from the spirit and scope of the present invention.

[0020] <Example 1> <Overview> The present invention is configured to allow a user to set prompts for generating output in a large-scale language model (LLM; hereinafter, the term may be abbreviated as "LLM"), and by inputting the user-set prompts into the LLM, desired information is extracted from a large volume of file data containing a wide variety of information and in various formats. Furthermore, the present invention aims to improve the accuracy of information extraction by the LLM, improve robustness (the degree to which extraction can be performed accurately from a variety of formats), and improve simplification, operability, and processing efficiency.

[0021] The following describes the functions and processing flow of the information processing device, as well as the hardware. The functional blocks of the system described below can be implemented as a combination of hardware and software. Specifically, if a computer is used, these include hardware components such as a CPU (central processing unit), main memory, bus, or secondary storage device (such as a hard disk drive, nonvolatile memory, or storage media such as CDs or DVDs, and their reader drives), input devices used for information input, printers, display devices, and other external peripheral devices, as well as interfaces for those external peripheral devices, communication interfaces, driver programs and other application programs for controlling the hardware, and user interface applications. The CPU processes data input from input devices and other interfaces and stored in memory or on a hard disk in accordance with a program deployed in main memory, processing and storing the data, and generating commands to control the hardware and software. Alternatively, the functional blocks of the system can be implemented using dedicated hardware.

[0022] Furthermore, this invention can be realized not only as a system but also as a method. Furthermore, a part of such an invention can be configured as software. Furthermore, a program used to cause a computer to execute such software, and a recording medium on which the program is fixed, are naturally included within the technical scope of this invention (the same applies throughout this specification).

[0023] <Functional Configuration> Fig. 1 is a block diagram showing an example of the functional configuration of the information processing device of this embodiment. As shown in Fig. 1, the information processing device 100 has a first receiving unit 101, an extracting unit 102, a second receiving unit 103, a converting unit 104, and a storing unit 105.

[0024] <First Reception Unit> The first reception unit 101 has a function of receiving from a user the setting of a main prompt, which is a prompt to be input into a large-scale language model and indicates one or more extraction targets from multiple file data containing multiple pieces of information.

[0025] The main prompt can contain not only instructions to the LLM, but also conditions for leading to an appropriate answer and examples to refer to. For example, when setting up a main prompt to instruct the LLM to extract the "drawing number," which is the ID of file data, the main prompt can be set to a sentence such as "Please extract the value of 'drawing number'. 'Drawing number' is the file ID and consists only of alphanumeric characters and hyphens. It also conforms to the following regular expression: [AZ]-\d{5}."

[0026] Furthermore, the main prompt is not limited to one, and multiple prompts can be set. For example, in addition to the instruction to extract the "drawing number" mentioned above, instructions to extract the "drawing title" and the "type and number of elevators" can also be set. To extract the "drawing title," the prompt can be set using a sentence such as, "Extract the value of 'Drawing Title'. The 'Drawing Title' summarizes the contents of the file and provides an overview of the file. The value is located below the words 'Drawing Title'." To extract the "type and number of elevators," the prompt can be set using a sentence such as, "Extract the type and number of elevators. If none apply, leave blank."

[0027] The extraction unit 102 has a function of inputting a main prompt based on the settings received by the first reception unit into a large-scale language model and extracting an extraction target from a plurality of file data. The plurality of file data may be stored in a storage device provided in the information processing device, or may be stored in a storage device connected to the information processing device via a communication line or the like. The file data may be in a variety of file formats, including, but not limited to, Word files, Portable Document Format (PDF) files, Excel files, Comma Separated Values ​​(CSV) files, text files, and Building Information Modeling (BIM) files.

[0028] The file data is then input to the LLM, and the extraction target is extracted in response to the instruction of the main prompt. Also, detailed prompts received by the second receiving unit (described later) can be further input and extracted. Also, the main prompt and detailed prompts stored in the storage unit (described later) can be input and extracted.

[0029] When the user enters the main prompt that instructs the extraction of the "drawing number" described above, for example, "Z-01301" is extracted as the "drawing number" and can be output along with the file name and a thumbnail image of the file data. The extracted information can then be tagged and entered and saved in an Excel file, or registered in a database, turning the unstructured file data into structured data suitable for use in search, statistical processing, AI learning data, etc.

[0030] In addition to importing the extracted information into an Excel file or the like and creating a database, the extracted information can also be processed according to the format of the file data from which it was extracted. For example, in the case of a CSV file, the extracted information can be added to a new column (row) within the CSV file and saved. In the case of an Excel file, the extracted information can be added to an added column or sheet within the Excel file and saved. In the case of a BIM file, the extracted information can be added to the BIM file and saved.

[0031] <Conversion Unit> The conversion unit 104 converts the format of file data according to the characteristics of the file data. If the information processing device has a conversion unit, the extraction unit has a function of extracting from the file data converted by the conversion unit. For example, if the file data is in PDF format, it is converted to a text file format. In addition, in the case of image data, character recognition is performed using OCR to convert it to a text file format. Performing such conversion contributes to improving robustness.

[0032] Furthermore, when converting to a text file, it is also preferable to add location information to the text. Figure 2 is a conceptual diagram showing an example of converting PDF file data into a text file with location information added. Figure 2(a) shows a portion of a PDF file showing a building area table for a house. Figure 2(b) shows an example of converting the text into a JSON format, in which the text for "shape," "formula," and "area" is written along with their location information. Figure 2(c) shows an example of converting the text into an array format, in which the text is also written along with location information. By adding location information in this way, extraction takes into account the relationships between text, contributing to improved robustness. Note that in the above example, the array format has the advantage of reducing the number of tokens input into the LLM. Figure 2(d) shows an example of converting text that, in the original file, shows a table showing the "area" and "tatami mats" for "Western-style Room 1," "Western-style Room 2," and "Western-style Room 3," into text with table information rather than location information (within the dotted-line frame in the figure). By converting in this way, it is possible to input data into the LLM while maintaining the table structure.

[0033] Figure 3 is a conceptual diagram showing another example of converting the file data format. As shown in Figure 3, in the case of a file containing graphics (a kitchen floor plan), converting it to SVG format and converting it into text with graphic information makes it easier to extract using LLM.

[0034] <Second Receiving Unit> The second receiving unit 103 has a function of receiving from the user the setting of a detailed prompt, which is a prompt for setting details of an extraction target instruction by a main prompt and has one or more items.

[0035] 4 is a diagram showing an example of a detailed prompt. The first line shows the setting items "No.", "Name.", "Description.", "Necessity.", and "Example." In this example, the main prompt is "Please extract the value of 'Drawing Number.'" and the instructions of this main prompt are set in detail.

[0036] For example, setting item 1 is called "Assigning Multiple Tag Names," and is indicated as a setting item for "dealing with variations in character notation." The explanation for this setting is "explicitly assign multiple data item names to be extracted," with an example being "["Drawing Number," "Drawing ID"]." By setting this detailed prompt, an instruction is added to the instruction "Please extract the value of 'Drawing Number,'" to include not only "Drawing Number," but also "Drawing ID" as a target for extraction.

[0037] Other advanced prompt settings include "Tag Description," which adds a description of the data item name to be extracted to address inconsistencies in character notation; "Absolute Position Specification," which specifies the range on the image where the data item exists to narrow the extraction range; "Target Page Specification," which specifies the page number of a multi-page PDF file; "Sample Specification," which specifies samples close to or far from the extracted value to prevent incorrect extraction; "Format Specification," which allows users to specify the format they want to extract to prevent incorrect extraction and standardize the format of output results; "Relative Position Specification," which allows users to specify the relative position of some content to support table formats; and "Condition Settings," which allows users to change the extracted content based on specific conditions if the desired content is not available in all files. Of course, users can also configure various settings themselves to configure advanced prompt settings.

[0038] Here, the application methods for each of the settings mentioned above can be broadly divided into three: Method 1, "narrowing down file contents," Method 2, "connecting prompts," and Method 3, "checking and reshaping detection results." Method 1 involves narrowing down the file data to be input before inputting it into LLM, Method 2 involves connecting the set detailed prompt to the main prompt, and Method 3 involves reshaping the extraction results according to the specified sample, format, etc.

[0039] FIG. 5 shows how each setting item in the detailed prompt is applied. As shown in the figure, "Connecting prompts" of Method 2 is applied to the items "Assigning multiple tag names," "Assigning tag descriptions," and "Setting conditions." Furthermore, "Specifying absolute position" and "Specifying target page" are applied to "Narrowing file contents" of Method 1 and Method 2. Furthermore, "Specifying sample" is applied to "Confirming extracted results and reformatting" of Method 2 and Method 3. Furthermore, "Specifying format" and "Specifying relative position" are applied to all of Methods 1 to 3. Examples of application are described below.

[0040] FIG. 6 is a conceptual diagram showing an example of an input screen when setting a detailed prompt. As shown in the figure, a main prompt 602 that has already been set is displayed in the upper left part of the screen 601. Input fields 603 for setting items such as "tag name," "absolute position specification," "sample specification," and "relative position specification" are displayed on the left side of the screen. Furthermore, input fields for other setting items are displayed in response to a user operation. A detailed prompt is set by accepting an input operation by the user into each of these input fields. The extraction unit then applies the accepted detailed prompt settings to perform extraction. By setting and extracting detailed prompts in this manner, extraction accuracy can be improved.

[0041] Here's an example of applying the above detailed prompt. For example, if you set the "Sample" field to "Similar content to D-00033, F-00102" and the "Relative Position" field to "This item is located in the bottom left corner of the page," this detailed prompt will be connected to the main prompt and entered. Then, by applying "Relative Position," only the bottom left corner of the page is entered into the LLM (Method 1: Narrowing the file contents). Then, by applying "Sample," the extracted "Drawing Number" is extracted, with drawing numbers similar to "D-00033, F-00102" (Method 3: Check the extracted results and reformat). This reformat can be performed using the LLM or by entering a separate script to instruct the reformat.

[0042] Furthermore, the second reception unit may be configured to automatically set each setting item according to the set main prompt when accepting the detailed prompt settings. Alternatively, the second reception unit may be configured to accept the user's settings and recommend some setting for items that have not been set. Accepting the detailed prompt settings from the user allows for flexible and appropriate extraction based on the user's intentions, but it may be difficult for the user to set the optimal prompt. Therefore, functions such as automatic entry and suggestion of detailed prompts are effective. These functions can simplify the process and improve operability.

[0043] For simplification and ease of use, the extraction unit is preferably configured to perform a test extraction using the set main prompt or the set main prompt and detailed prompts before extracting all of the target file data. Test extraction refers to extracting a portion of the target file data. Based on the results of this test extraction, the user can determine whether the set main prompt and detailed prompts will produce the desired extraction results, and can reset the prompts to obtain more desirable extraction results.

[0044] As shown in Figure 6, a "Test" button 605 is displayed on the left side of a screen 601, below a detailed prompt input field 603 and a main prompt display field 604, and test extraction is performed by clicking this button. The results of the test extraction are displayed in a dotted-line frame 606, which is shown for explanatory purposes in the figure. Based on the extraction results, the user can reset the prompt or perform actual extraction using the prompt for which the test extraction results were obtained.

[0045] <Storage Unit> The storage unit 105 has a function of storing a main prompt based on the settings received by the first reception unit and a detail prompt based on the settings received by the second reception unit. The extraction results and file data used for the extraction may be stored in association with the main prompt and detail prompt used for the extraction. The main prompt and detail prompt stored in the storage unit are input to the LLM for extraction by the extraction unit in place of or together with the main prompt and detail prompt based on the settings received by the first and second reception units.

[0046] For example, when an extraction process for a certain file data group is first performed, the main prompt and detail prompts set by the first and second reception units are input into the LLM and extracted. Then, if a change (such as an addition or deletion) occurs in the files contained in the file data group and a new extraction is performed, the main prompt and detail prompts used in the previous extraction process stored in the storage unit are read, and if necessary, the settings are changed and input into the LLM for extraction. Alternatively, the user may select the main prompt and detail prompts stored in the storage unit and use them for the extraction process. Alternatively, when a file data group to be extracted is selected, recommended main prompts and detail prompts from the stored prompts are presented. This configuration eliminates the need to set prompts each time an extraction is performed, and extraction is performed using prompts appropriate for each file data group, resulting in, for example, maintaining an up-to-date database.

[0047] <Hardware Configuration> Fig. 7 is a conceptual diagram showing an example of the configuration of hardware that realizes an information processing device according to the first embodiment. As shown in the figure, the information processing device 700 includes a CPU 701 that performs various arithmetic processing, a RAM 702 that is a volatile storage medium, a storage 703 such as a flash memory or HDD that is a non-volatile storage medium, a communication interface 704, and an input / output interface 705. The RAM 702 reads programs that perform various arithmetic processing to be executed by the CPU 701, and provides a work area (working area) for the programs. In addition, multiple addresses are assigned to the RAM 702, and programs executed by the CPU 701 can exchange data and perform processing by identifying and accessing the addresses (the same applies throughout this specification).

[0048] Here, the functions of the first reception unit 101, the second reception unit 103, the extraction unit 102, and the conversion unit 104 of the information processing device 100 in Fig. 1 are mainly realized by the CPU 701 and RAM 702 in Fig. 7. The function of the accumulation unit 105 is mainly realized by the storage 703 in Fig. 7. In addition, when an external database for accumulating file data exists or when the extraction unit uses an external large-scale language model, each function is realized by mutually exchanging signals and information via a communication interface 704 and an input / output interface 705.

[0049] <Processing Flow> FIG. 8 is a flow diagram illustrating a simplified example of the processing flow of an information processing device. First, a main prompt setting, which is a prompt to be input to a large-scale language model and indicates one or more extraction targets from multiple file data containing multiple pieces of information, is received from the user (S801: first receiving step). Then, a detail prompt setting, which is a prompt for specifying in detail the extraction target indicated by the main prompt and has one or more items, is received from the user (S802: second receiving step). Then, the format of the file data is converted according to the characteristics of the file data (S803: conversion step). Then, the main prompt based on the setting received in the first receiving step and the detail prompt based on the setting received in the second receiving step are accumulated (S804: accumulation step). Then, the main prompt based on the setting received in the first receiving step and the detail prompt based on the setting received in the second receiving step and / or the main prompt and the detail prompt accumulated in the accumulation step are input to the large-scale language model, and extraction targets are extracted from the multiple file data (S805: extraction step). The conversion step may be performed before the first reception step, or may be performed after the first reception step and before the second reception step.

[0050] <Effects> According to the information processing apparatus of this embodiment, it is possible to extract desired information from a large amount of file data that includes a wide variety of information and has a variety of formats.

[0051] Second Embodiment Overview This embodiment is an information processing device based on the first embodiment, and further includes a configuration for verifying the extraction results and a configuration for filtering file data to be input to the LLM.

[0052] <Functional Configuration> Fig. 9 is a block diagram showing an example of the functional configuration of an information processing device of this embodiment. As shown in Fig. 9, the information processing device 900 has a first reception unit 901, an extraction unit 902, a second reception unit 903, a conversion unit 904, and a storage unit 905, similar to the first embodiment, and further has a verification unit 906 and a filtering unit 907. A description of the same configuration as the first embodiment will be omitted, and the verification unit 906 and the filtering unit 907 will be described below.

[0053] <Verification Unit> The verification unit 906 has a function of verifying the extraction results extracted by the extraction unit. There are various specific modes of verification by the verification unit. FIG. 10 is a diagram illustrating various settings when performing verification. As shown in the figure, four types of settings can be made, namely, "Use detailed prompt," "Verification prompt," "Verification script," and "Compare between files."

[0054] "Use advanced prompt" uses the extraction settings set in the advanced prompt for verification. For example, if a negative sample is specified in the "Specify sample" setting of the advanced prompt, the specified negative sample is reused to verify whether the extraction result is similar to the negative sample. Also, the setting in "Specify relative position" is used to verify whether the word "roof" is to the left of the extracted value.

[0055] The "verification prompt" is used to individually set a prompt for verification, for example, by setting and inputting a prompt such as "Please check whether this value indicates the material of the elevator door." The "verification script" is used to individually set a script for verification, for example, as shown in the figure, by inputting a command to return "Error" if the detection result is less than "1" and "OK" otherwise.

[0056] Furthermore, when extracting from multiple files, "file comparison" compares the extracted results with other files to confirm their validity. As shown in the example, when comparing extracted files A to D, files A, C, and D all contain only numbers and symbols, while file B is primarily written in kanji. In such a case, the verification result is that "file B is not valid compared to the others."

[0057] Here, when the verification unit verifies the extraction result, the extraction unit can be configured to reflect the verification result and perform re-extraction. For example, if the value of the extraction result is an error as described above, an instruction such as "The previous extraction extracted the value XXX, but the verification result was an error. Please extract the value again taking this error into consideration" can be input and the extraction will be performed again.

[0058] <Filtering Unit> The filtering unit 907 has the function of removing predetermined file data from among multiple file data items from the extraction unit. There are various filtering methods. For example, filtering is performed based on file attributes (extension, size, file name, data items associated with the file, etc.). Also, individual keywords are set and filtering is performed based on these keywords. Specifically, filtering is performed by performing semantic search, clustering, character search, etc. Also, filtering is performed by giving instructions to the LLM. For example, by inputting an instruction such as "Does this file contain elevator content?", filtering can be performed based on whether or not the file contains elevator content.

[0059] This filtering function can meet the user's needs, for example, to extract the area from the architectural summary of 500 target file data, but skip the elevation drawings.

[0060] <Hardware Configuration> The information processing device of this embodiment can be realized by the hardware configuration shown in Fig. 7 according to the first embodiment. The functions of the verification unit 906 and the filtering unit 907, which are components specific to the information processing device of the second embodiment shown in Fig. 9, are mainly realized by the CPU 701 and the RAM 702 in Fig. 7.

[0061] <Processing Flow> The processing flow of the information processing device of this embodiment is basically the same as the processing flow of the information processing device of Example 1. The information processing device of this embodiment further includes a verification step of verifying the extraction result extracted in the extraction step, and a filtering step of removing predetermined file data from among the plurality of file data to be extracted in the extraction step.

[0062] <Effects> According to the information processing device of this embodiment, it is possible to provide an information processing device with improved accuracy and robustness.

[0063] <Example 3> <Overview> This example is an information processing device based on Example 1 or Example 2, and further includes a configuration for displaying the basis for extracting the extraction result in order to improve simplicity and operability, or a configuration for reducing the data provided for extraction processing in order to improve processing efficiency.

[0064] <Functional Configuration> Fig. 11 is a block diagram showing an example of the functional configuration of the information processing device of this embodiment. As shown in Fig. 11, the information processing device 1100 includes, similarly to Example 1, a first reception unit 1101, an extraction unit 1102, a second reception unit 1103, a conversion unit 1104, a storage unit 1105, a verification unit 1106, and a filtering unit 1107, and further includes a basis display unit 1108 and a data reduction unit 1109. A description of the same configuration as Example 1 or 2 will be omitted, and the basis display unit 1108 and the data reduction unit 1109 will be described below.

[0065] <Basis Display Unit> The basis display unit 1108 has a function of displaying the basis for the extraction result extracted by the extraction unit. Note that, although the basis display unit 1108 passes through the verification unit 1106 in Fig. 11, it is also possible to display the basis for the extraction result by the extraction unit 1102 without passing through the verification unit 1106.

[0066] This information processing method allows users to set prompts and instruct the LLM to perform flexible extraction in line with their intentions. However, it can be difficult to set optimal prompts that meet the user's intentions. Therefore, by displaying the reasons (intermediate processes, etc.) that led to the extraction results, the user can be shown "how to improve the extraction results."

[0067] 12 is a conceptual diagram showing an example of displaying a position on an image as a basis. The figure shows a PDF file showing a building area table of a house, and the position of a value (character) extracted from this file is highlighted on the image by enclosing it in a thick dotted frame 1201. This allows the user to understand where the extraction result came from.

[0068] The reason for the extraction can also be displayed. For example, before extraction, a preprocessing instruction can be entered into the LLM: "Please output the detailed process-by-process reason for outputting this value as the extraction result." After extraction, the extraction results will be displayed along with the rationale, such as, "Z-01301 is located 140px to the right of the text 'Drawing Number' and matches the specified regular expression, so Z-01301 was extracted as the drawing number." In this way, by being shown the effect of the prompts you set on the extraction results, users can reflect on the prompts you set and optimize them.

[0069] <Data Reduction Unit> The data reduction unit 1109 has the function of reducing the file data to be provided to the extraction unit. Data reduction can be performed in various ways. For example, data reduction can be performed by narrowing down the target file data to specific pages, sections, paragraphs, tables, etc. It is also possible to filter by file attributes (page number, position, etc.) or by individual keywords (semantic search, clustering, character search). It is also possible to filter by inputting instructions to the LLM ("Does this file contain elevator contents?").

[0070] FIG. 13 is a conceptual diagram showing another aspect of data reduction. As shown in FIG. 13( a), a specific portion 1302 is cut out from one file data 1301, and a check is made to see if it contains the target data. If the target data is found to be included, that portion is subjected to extraction processing by the extraction unit. If the target data is not included, a portion 1303 shifted from the previous portion is cut out, and a similar check is performed. In this way, processing efficiency can be improved by excluding from the extraction target any portion that does not contain the target data in the pre-extraction check. Note that this check may be performed using LLM or other methods.

[0071] 13B, the above check is performed while shifting the specific location, and the location containing the target data is temporarily stored in memory 1304. Then, it is checked whether processing memory 1304 contains all the elements necessary to extract the target data, and if it does, the location held in memory is subjected to extraction processing by the extraction unit. On the other hand, if it does not contain all the elements, the specific location is shifted further and the series of processes starting from the above check are repeated.

[0072] <Hardware Configuration> The information processing apparatus of this embodiment can be realized by the hardware configuration shown in Fig. 7 according to the first or second embodiment. The function of the basis display unit 1108, which is a configuration unique to the information processing apparatus of the third embodiment shown in Fig. 11, is realized mainly by the CPU 701, RAM 702, and input / output interface 705 in Fig. 7. Furthermore, the function of the data reduction unit 1109 is realized mainly by the CPU 701 and RAM 702 in Fig. 7.

[0073] <Processing Flow> The processing flow of the information processing device of this embodiment is basically the same as the processing flow of the information processing device of Embodiment 1 or 2. The information processing device of this embodiment further includes a basis display step of displaying the basis for the extraction result extracted in the extraction step, and a data reduction step of reducing the data of file data to be used in the extraction step.

[0074] <Effects> According to the information processing apparatus of this embodiment, it is possible to provide an information processing apparatus that is simpler, easier to operate, and more efficient in processing.

[0075] 100: Information processing device 101: First reception unit 102: Extraction unit 103: Second reception unit 104: Conversion unit 105: Storage unit 906: Verification unit 907: Filtering unit 1108: Basis display unit 1109: Data reduction unit 701: CPU 702: RAM 703: Storage 704: Communication interface 705: Input / output interface

Claims

1. An information processing device having: a first reception unit that receives from a user settings for a main prompt to be input into a large-scale language model, the main prompt indicating one or more extraction targets from a plurality of file data containing a plurality of pieces of information; a second reception unit that receives from the user settings for a detailed prompt having one or more items for setting in detail the extraction target indicated by the main prompt; a storage unit that stores the main prompt based on the settings received by the first reception unit and the detailed prompt based on the settings received by the second reception unit; and an extraction unit that inputs the main prompt based on the settings received by the first reception unit and the detailed prompt based on the settings received by the second reception unit and / or the main prompt and the detailed prompts stored in the storage unit into a large-scale language model and extracts extraction targets from the plurality of file data.

2. The information processing device according to claim 1, further comprising a conversion unit that converts the format of the file data in accordance with characteristics of the file data, and the extraction unit extracts from the file data converted by the conversion unit.

3. The information processing device according to claim 1 or 2, further comprising a verification unit that verifies the extraction result extracted by said extraction unit.

4. The information processing apparatus according to claim 3, wherein the verification unit performs verification using the received detailed prompt settings.

5. The information processing device according to claim 3 or 4, wherein the extraction section performs re-extraction by reflecting the result of verification by the verification section.

6. The information processing device according to claim 1 or 2, further comprising a reason display unit that displays a reason why the extraction result extracted by said extraction unit is extracted.

7. An information processing method executed by a computer in an information processing device, comprising: a first receiving step of receiving from a user a setting of a main prompt to be input into a large-scale language model, the main prompt indicating one or more extraction targets from a plurality of file data containing a plurality of information; a second receiving step of receiving from the user a setting of a detailed prompt having one or more items, the detailed prompt being a prompt for setting in detail the extraction target indicated by the main prompt; an accumulation step of accumulating the main prompt based on the setting received in the first receiving step and the detailed prompt based on the setting received in the second receiving step; and an extraction step of inputting the main prompt based on the setting received in the first receiving step and the detailed prompt based on the setting received in the second receiving step and / or the main prompt and the detailed prompt accumulated in the accumulation step into a large-scale language model and extracting the extraction targets from the plurality of file data.

8. An information processing program that causes a computer to execute the following steps in an information processing device: a first reception step of receiving from a user a setting of a main prompt to be input into a large-scale language model, the main prompt indicating one or more extraction targets from a plurality of file data containing a plurality of information; a second reception step of receiving from the user a setting of a detailed prompt having one or more items for setting in detail the extraction target indicated by the main prompt; an accumulation step of accumulating the main prompt based on the setting received in the first reception step and the detailed prompt based on the setting received in the second reception step; and an extraction step of inputting the main prompt based on the setting received in the first reception step and the detailed prompt based on the setting received in the second reception step and / or the main prompt and the detailed prompt accumulated in the accumulation step into a large-scale language model and extracting the extraction targets from the plurality of file data.

Images