JP7551203B1 - Information processing device, information processing method, and information processing program - Google Patents

Abstract

In the conventional technology, the LLM must be called up every time information collection processing is performed, which is time-consuming and costly.
[Solution] In order to solve the above problem, an information processing device is provided that has a prompt setting unit that sets a prompt for specifying desired information, which is information about a desired item, from multiple file data including multiple target items, a sample setting unit that sets sample file data to be input into a trained language model together with the set prompt, a script generation unit that inputs the set prompt and the set sample file data into the trained language model to generate a script for acquiring the specified desired information, and a desired information acquisition unit that acquires the specified desired information from the target file data using the generated script.

Description

The present invention relates to an information processing device that obtains desired information from a file data group containing various information.

In recent years, the use of large-scale language models (LLMs, hereafter sometimes abbreviated as LLMs) to obtain specific information from file data groups has become widespread. For example, Patent Literature 1 discloses a technology that uses LLMs to collect character information related to specific items from a variety of printed materials with no set format, such as advertising flyers.

Patent No. 7430437

The technology of Patent Document 1 collects one or more pieces of character information relating to one or more specific items from a group of character strings printed on a flyer read by OCR by providing the LLM with prompts including instructions for specifying the items to be collected and various other instructions related to the collection. With this type of technology, the LLM must be called up each time information collection is processed, which creates the problem of taking time and cost. In addition, because the LLM must be called up each time processing is performed, there is the problem of inconsistent extraction accuracy.

Therefore, in order to solve the above problem, the present invention provides the following information processing device etc. That is, an information processing device is provided that has a prompt setting unit that sets a prompt for specifying desired information, which is information related to a desired item, from a plurality of file data including a plurality of target items, a sample setting unit that sets sample file data to be input to a trained language model together with the set prompt, a script generation unit that inputs the set prompt and the set sample file data to the trained language model to generate a script for acquiring the specified desired information, and a desired information acquisition unit that acquires the specified desired information from the target file data by the generated script.

In addition to the above features, an information processing device is provided that further has a script storage unit that stores the generated script.

In addition to the above features, an information processing device is provided in which the script generation unit further inputs correct answer information, which is information on the correct answer of the specified desired information for the sample file data, to generate the script.

In addition to the above features, the information processing device further includes a script evaluation unit that evaluates the result of obtaining the specified desired information from the sample file data using the generated script in comparison with the correct answer information.

In addition to the above features, an information processing device is provided in which the script generation unit further inputs the evaluation results by the script evaluation unit to generate the script.

In addition to the above features, the present invention also provides an information processing device further having a script correction unit that inputs the generated script and some or all of the target file data into the trained language model and corrects the generated script to adapt it to the target file data.

In addition to the above features, the script correction unit also provides an information processing device that inputs the prompt that is the basis of the generated script into the trained language model to correct it.

In addition to the above features, an information processing device is provided in which the script correction unit also corrects the scripts stored in the script storage unit.

In addition to the above features, an information processing device is provided that further has a detailed prompt setting unit that sets a detailed prompt, which is a detailed prompt for obtaining desired information specified by the set prompt.

In addition to the above features, an information processing device is provided that further has a verification unit that verifies the acquisition results acquired by the desired information acquisition unit.

In addition to the above features, an information processing device is provided in which the script correction unit corrects the script by reflecting the verification results by the verification unit.

Also provided is an information processing method executed by an information processing device, comprising: a prompt setting step for setting a prompt for specifying desired information, which is information relating to a desired item, from a plurality of file data including a plurality of target items; a sample setting step for setting sample file data to be input to a trained language model together with the set prompt; a script generation step for inputting the set prompt and the sample file data into the trained language model to generate a script for acquiring the desired information; and a desired information acquisition step for acquiring the desired information from the plurality of target file data using the generated script.

Also provided is an information processing program that causes an information processing device to execute the following steps: a prompt setting step for setting a prompt to specify desired information, which is information relating to a desired item, from a plurality of file data including a plurality of target items; a sample setting step for setting sample file data to be input to a trained language model together with the set prompt; a script generation step for inputting the set prompt and sample file data into the trained language model to generate a script for acquiring the desired information; and a desired information acquisition step for acquiring the desired information from the plurality of target file data using the generated script.

Also provided is an information processing device having a script holding unit that holds a script for acquiring desired information, which is information relating to a desired item, from a plurality of file data including a plurality of target items, a held script modification unit that inputs a part or all of the target file data into the trained language model and modifies the held script to adapt it to the target file data, and a second desired information acquisition unit that acquires the desired information from the plurality of target file data using the modified script.

In addition to the above features, an information processing device is provided in which the retained script correction unit further inputs a prompt for correcting the retained script into the pre-trained language model to correct it.

Also provided is an information processing method executed by an information processing device, comprising: a script retaining step of retaining a script for acquiring desired information, which is information relating to a desired item, from a plurality of file data including a plurality of target items; a retained script modification step of inputting a part or all of the target file data into the trained language model and modifying the retained script to adapt it to the target file data; and a second desired information acquisition step of acquiring the desired information from the plurality of target file data using the modified script.

Also provided is an information processing program that causes an information processing device to execute a script retaining step of retaining a script for acquiring desired information, which is information relating to a desired item, from a plurality of file data including a plurality of target items; a retained script modification step of inputting a part or all of the target file data into the trained language model and modifying the retained script to adapt it to the target file data; and a second desired information acquisition step of acquiring the desired information from the plurality of target file data using the modified script.

The present invention makes it possible to obtain desired information from a group of file data containing various information while reducing the time and cost required, and also enables reproducible extraction of information with a certain degree of extraction accuracy.

FIG. 1 is a block diagram showing an example of a functional configuration of an information processing apparatus according to a first embodiment; Conceptual diagram explaining the functions of the script generation part FIG. 1 is a conceptual diagram showing an example of a hardware configuration for implementing an information processing apparatus according to a first embodiment of the present invention; FIG. 1 is a flow chart showing an example of a process flow of an information processing apparatus according to a first embodiment of the present invention; FIG. 11 is a block diagram showing an example of a functional configuration of an information processing apparatus according to a second embodiment of the present invention; A diagram showing the detailed prompt settings and examples A diagram showing how each setting in the Advanced prompt is applied Conceptual diagram showing an example of an input screen when setting detailed prompts A diagram showing various settings for performing verification. A conceptual diagram showing an example of converting PDF file data into a text file with location information added. Conceptual diagram showing another example of converting file data formats FIG. 13 is a conceptual diagram showing another aspect of data reduction. Conceptual diagram showing an example of displaying the location on the image as evidence FIG. 11 is a block diagram showing an example of a functional configuration of an information processing apparatus according to a third embodiment. FIG. 11 is a flow diagram showing an example of a process flow of an information processing apparatus according to a third embodiment.

Hereinafter, the embodiments of the present invention will be described with reference to the accompanying drawings. Note that the present invention should not be limited to these embodiments, and may be implemented in various forms without departing from the spirit of the present invention.

Example 1
＜Overview＞
The present invention is directed to obtaining desired information by executing a script (simple program) from various file data groups, and generates the script by inputting sample file data and a prompt into a learned language model (LLM, hereinafter sometimes abbreviated as LLM), which is typically a large-scale language model. Here, a static script can be generated that ensures robustness (the degree to which extraction can be performed accurately from a variety of formats) within the range of the input samples. In addition, the LLM is called only once when generating the script, which reduces the time and cost required by the conventional technology. In addition, since processing is performed using a static script, the extraction result is uniquely determined.

The following describes the functions and processing flow of the information processing device, as well as the contents of the hardware. The functional blocks of the system described below can be realized as a combination of hardware and software. Specifically, if a computer is used, the hardware components include a CPU (central processing unit), main memory, bus, or secondary storage device (hard disk drive, non-volatile memory, storage media such as CDs and DVDs, and read drives for these media, etc.), input devices used for inputting information, printing devices, display devices, and other external peripheral devices, as well as interfaces for these external peripheral devices, communication interfaces, driver programs and other application programs for controlling these hardware, and user interface applications. Then, by the arithmetic processing of the CPU according to the program deployed on the main memory, data input from input devices and other interfaces and stored in memory or hard disk is processed and stored, and commands for controlling the above hardware and software are generated. Alternatively, the functional blocks of the system may be realized by dedicated hardware.

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, the program used to cause a computer to execute such software, and the recording medium on which the program is fixed, are naturally included in the technical scope of this invention (the same applies throughout this specification).

<Functional configuration>
1 is a block diagram showing an example of the functional configuration of an information processing apparatus according to the present embodiment. As shown in FIG. 1, the information processing apparatus 100 includes a prompt setting unit 101, a sample setting unit 102, a script generating unit 103, a script storage unit 104, and a desired information acquiring unit 105.

<Prompt setting section>
The prompt setting unit 101 has a function of setting a prompt for specifying desired information, which is information on a desired item, from a plurality of file data including a plurality of target items. The target file data may be stored in a storage provided in the information processing device, or may be stored in a storage connected to the information processing device via a communication line or the like. In addition, there are many types of file formats for the file data, and the file formats are not limited to these, but include, for example, Word files, Portable Document Format (PDF) files, Excel files, image files, Comma Separated Values (CSV) files, text files, Building Information Modeling (BIM) files, Computer Aided Design (2D CAD) files, and 3D model files.

In addition, "desired information" refers to information on a desired item or information obtained by performing some kind of calculation or processing on the information on that item (total number, average information, maximum information, spatial relationships such as distance and collision, three-dimensional model information, three-dimensional graphic information, images, text information, etc.). For example, if the "total number of elevators" is desired from a group of file data of CSV files related to the facilities of multiple buildings, the number of elevators in the buildings included in each file is extracted and the total number of these numbers becomes the desired information. In addition to calculating numerical information, the results of processing such as three-dimensional space processing and three-dimensional graphic processing may also be set as the desired information. For example, if the "collision points between columns and beams" are desired for three-dimensional model files of multiple buildings, the three-dimensional space information obtained by extracting all the models of columns and beams included in each file, calculating whether they collide, and extracting the collision points becomes the desired information.

Furthermore, for example, if the target is 3D drawing files of multiple buildings and the "inconsistencies between the elevations and floor plans" are desired, all architectural design information for the elevations and floor plans contained in each file is extracted, and the presence or absence of consistency when reconstructed into a 3D solid is checked, and the text that documents the inconsistencies becomes the desired information. In such 3D spaces or processing of 3D spaces, the present invention, which performs a certain amount of processing statically, is often able to obtain the desired information more accurately. A prompt is set to specify such desired information. This prompt may be set by the user, or may be set in advance by the person who provides the device to the user.

Note that a script is generated to obtain the desired information specified by a prompt set in the script generation unit described below, but a prompt that also instructs the generation of the script may be set in the prompt setting unit. For example, if the desired information is "total number of elevators" as in the above example, a prompt may be set with instructions specifying the desired information, such as "Please find the total number of elevators from multiple given CSV files," or a prompt may be set with instructions to create a script to obtain the specified desired information, such as "Please create a script that finds the total number of elevators from multiple given CSV files."

In addition, when the script processing becomes complicated, a prompt for generating the script may be set in multiple steps. For example, in the example of extracting "collision points between pillars and beams" from the aforementioned 3D model, the prompt may be divided into (1) a prompt for generating a script for extracting the models of pillars and beams contained in each file, (2) a prompt for generating a script for checking whether there is a collision between them, and (3) a prompt for generating a script for acquiring information on the 3D space where the collision points have been extracted, and each prompt for generating a script may be specified.

<Sample setting section>
The sample setting unit 102 has a function of setting sample file data to be input to the trained language model together with the set prompt. The sample file data is a model file data when desired information is acquired from a plurality of target file data. For example, when "total number of elevators" is set as the desired information as in the above example, a CSV file for a plurality of "building facilities" including information on the number of elevators is set as the sample file data. When "collision points between columns and beams" is set as the desired information, a three-dimensional model file for a plurality of building structures including information on columns and beams is set as the sample file data. When "consistency between elevations and floor plans" is set as the desired information, a two-dimensional drawing file for a plurality of "building designs" including information on elevations and floor plans is set as the sample file data.

It is also preferable to set multiple file data containing the desired information, each of which has a different format, as sample file data. This is because it contributes to the generation of a script that can generically obtain the desired information from file data of various formats.

Note that in many cases, this sample file data will naturally be a small amount compared to the multiple file data to be acquired, but the opposite is also possible. Also, the sample file data may be a part of the multiple file data to be acquired, or it may be separate file data. The sample file data may be set by the user, or may be set in advance by the person who provides the device to the user.

For convenience, the prompt setting and the sample file data setting have been described as separate configurations, but the prompt and sample file data may be configured to be set together. In other words, by setting a prompt that specifies the desired information and also includes link information to the sample file data, the functions of the prompt setting unit and sample setting unit described above can be realized in the same way.

<Script generation part>
The script generation unit 103 has a function of inputting the set prompt and the set sample file data into the trained language model, and generating a script for acquiring the specified desired information.

Figure 2 is a conceptual diagram explaining the function of the script generation unit. The examples shown are sample file data 201 and 202. The sample file data is a list showing the equipment (elevators, fire extinguishing equipment, sanitary equipment) and their numbers for each building, with sample file data 201 being written in Japanese and sample file data 202 being written mainly in English.

These sample file data are input into the LLM along with the prompts. In the following, we will use the LLM, which is a typical example of a trained language model, for explanation. Note that the trained language model may be anything other than the LLM, such as a small language model or a multimodal language model that can handle images and other formats, as long as it can perform natural language processing tasks.

And the prompt 203 shown is:
"I will now provide several CSV files of building equipment lists as samples. These lines show the types and quantities of equipment in several buildings. Please create a script that extracts the number of elevators from these CSV files and calculates the total number."
This instructs the generation of a script that calculates the total number of elevators, which is the desired information.

The LLM 204 generates a script in response to the input of such a prompt 203 and sample file data 201, 202. By using a list of English characters as well as a list of Japanese characters as sample file data, the generated script 205 executes a process of adding up the total number of pieces of information in the "Number" column in the row where the "Equipment" column is "Elevator" and the total number of pieces of information in the "Number" column in the row where the "Elevator" column is "Equipment". By inputting various samples of notation, it is possible to create a script that can handle so-called "variations in notation."

In addition, if the script processing is complicated as mentioned above, a script may be generated for each of the set prompts in multiple steps. It is also possible for the user to make their own modifications to the generated script.

In addition, by inputting sample file data in various formats, it is possible to generate a script that can appropriately extract information from file data in each of those formats. In the example of extracting "collision points between columns and beams" from the 3D model mentioned above, by inputting multiple 3D models in different formats as sample data, a general-purpose script that can handle a variety of model formats can be generated.

The script generation unit can also be configured to generate a script by further inputting correct answer information, which is information on the correct answer of the specified desired information for the sample file data. For example, in the above example, if the total number of elevators extracted from the set sample file data is "241," this "241" becomes the correct answer information. Inputting this correct answer information into the LLM also contributes to the generation of a script that can more accurately obtain desired information.

The correct answer information specified here may be in any format, such as not only numerical information but also a three-dimensional space model resulting from the processing of the script, a two-dimensional figure, an image, or text. In addition, the correct answer information may be not only information on the final result of the script, but also information on the correct answer in the processing process. For example, in the example of extracting "collision points between pillars and beams" from the above-mentioned three-dimensional model, (1) the pillar and beam models included in each file, (2) the presence or absence of collisions between them, and (3) information on the three-dimensional space from which the collision points are extracted may be given as correct answer information in the processing process. Furthermore, since it requires expertise in operation for the user to specify the three-dimensional model that is the correct answer information, it is also desirable to install a three-dimensional viewer or display unit that allows the user to easily specify the three-dimensional model or three-dimensional space that is the correct answer information.

<Script storage section>
The script storage unit 104 has a function of storing the generated script. By having the script storage unit, it is possible to reuse a script that has been set once, and to obtain values with reproducibility from a large amount of target file data. In addition, the generated script may be used to store the results of acquisition obtained by a desired information acquisition unit (described later) in association with the original prompt and sample file data that were input together.

<Desired Information Acquisition Unit>
The desired information acquisition unit 105 has a function of acquiring the specified desired information from the target file data by the generated script. The generated script is executed on the target file data to perform processing to acquire the specified desired information. It is preferable to save the acquired results. This is to provide the acquired results for post-processing, such as verifying and reformatting the acquired results, as will be described later.

The information acquired by the desired information acquisition unit may be passed to another script as new target file data. In other words, the generation of the script and the acquisition of desired information from the target file in this invention can be linked, making it possible to configure more complex processing.

As described above, the process of obtaining the desired information is performed by executing a script, which reduces the time and cost required compared to calling the LLM each time to obtain the desired information, and also enables reproducible information to be extracted with a certain degree of extraction accuracy.

Furthermore, there was a problem in that "even if we tried to prepare a static script to solve this problem, only engineers with specialized knowledge could create the script themselves." However, by generating scripts with LLM, the effect is that anyone can create the script themselves.

<Hardware configuration>
3 is a conceptual diagram showing an example of the configuration of hardware that realizes the information processing device of the first embodiment. As shown in the figure, the information processing device 300 has a CPU 301 that performs various arithmetic processing, a RAM 302 that is a volatile recording medium, a storage 303 such as a flash memory or HDD that is a non-volatile storage medium, a communication interface 304, and an input/output interface 305. The RAM 302 reads out a program that performs various arithmetic processing to cause the CPU 301 to execute it, and provides a work area (working area) for the program. In addition, a plurality of addresses are assigned to the RAM 302, and the program executed by the CPU 301 can exchange data with each other and perform processing by identifying and accessing the address (the same applies throughout this specification).

Here, the functions of the prompt setting unit 101, sample setting unit 102, script generation unit 103 and desired information acquisition unit 105 of the information processing device 100 in Figure 1 are mainly realized by the CPU 301 and RAM 302 in Figure 3. Also, the function of the script accumulation unit 105 is mainly realized by the storage 303 in Figure 3. Also, when a database that accumulates file data is external, or when a learned language model such as an LLM in which the extraction unit is external is used, each function is realized by sending and receiving signals and information to and from each other via the communication interface 304 and the input/output interface 305.

<Processing flow>
FIG. 4 is a flow diagram showing an example of a process flow of the information processing apparatus according to the first embodiment. First, a prompt is set for specifying desired information, which is information on a desired item, from a plurality of file data including a plurality of target items (S401: prompt setting step). Then, sample file data to be input to the trained language model together with the set prompt is set (S402: sample setting step). Then, the set prompt and the set sample file data are input to the trained language model to generate a script for acquiring the specified desired information (S403: script generation step). Then, the generated script is accumulated (S404: script accumulation step). Then, the specified desired information is acquired from the target file data by the generated script (S405: desired information acquisition step). The sample setting step may be performed before the prompt setting step, or both steps may be performed integrally. Also, the script accumulation step may be performed after the desired information acquisition step.

＜Effects＞
According to the information processing device of this embodiment, it is possible to obtain desired information from a group of file data containing various information while reducing the time and cost required, and it is also possible to extract reproducible information with a certain degree of extraction accuracy.

Example 2
＜Overview＞
The present embodiment is an information processing apparatus based on the first embodiment, and further includes a configuration for evaluating and modifying a generated script, a configuration for setting more detailed prompts, and a configuration for verifying acquired results.

<Functional configuration>
Fig. 5 is a block diagram showing an example of the functional configuration of the information processing apparatus of this embodiment. As shown in Fig. 5, the information processing apparatus 500 has a prompt setting unit 501, a sample setting unit 502, a script generating unit 503, a script storing unit 504, and a desired information acquiring unit 505, similar to the first embodiment, and further has a script evaluating unit 506, a script correcting unit 507, a detailed prompt setting unit 508, and a verifying unit 509. A description of the same configuration as the first embodiment will be omitted, and the script evaluating unit 506, the script correcting unit 507, the detailed prompt setting unit 508, and the verifying unit 509 will be described below.

<Script Evaluation Section>
The script evaluation unit 506 has a function of evaluating the result of acquiring the specified desired information from the sample file data by the generated script, by comparing it with the correct answer information. This correct answer information is as described in the script generation unit of the first embodiment. By actually acquiring the desired information from the sample file data by the generated script and comparing the result with the correct answer information, it is possible to evaluate how accurate the script is. In other words, if the acquired result matches the correct answer information, it can be evaluated as accurate.

In addition, by configuring the script generation unit to further input the evaluation results into the LLM, a script can be generated based on the evaluation results, which can be useful for generating more accurate scripts. For example, it can be configured to continue regenerating the script until the acquired information matches the correct answer information. Note that the script evaluation may be performed on the scripts stored in the script storage unit.

<Script Modification Section>
The script correction unit 507 has a function of inputting the generated script and a part or all of the target file data into the trained language model, and correcting the generated script so as to adapt it to the target file data. Note that the trained language model used for the correction may be the same as the trained language model used in the script generation unit described above, or may be a different one.

For example, if the target file data contains file data in a file format that was not included in the sample file data entered when creating the script, or contains item names to be obtained, the desired information may not be obtained accurately. To prevent such situations, the generated script is input into the trained language model along with the target file data.

Then modify the script by providing a prompt like the example below:
This script was written with the following objective in mind: "Select the row that shows elevators from the list of facilities in the building from the CSV file, extract the information from the column showing the total number, and find the total number."
However, it may not work properly with the file data you are processing. Please find out what is not working properly and fix the script."
For example, if a script created to obtain the "total number of elevators" executes a process of adding up the total number of pieces of information in the "Number" column in a row where the "Equipment" column is "Elevator" and the total number of pieces of information in the "Number" column in a row where the "Equipment" column is "Elevator", if the target file data includes file data in a row where the "Equipment" column is "Elevator" and has a "Number" column, the information of this file data will not be obtained, and accurate desired information cannot be obtained. Therefore, by inputting the created script and the target file data into the trained language model along with the above-mentioned prompt, a script adapted to the target file data can be generated.

The modifications to accommodate the file data may be done by inputting the file data into the LLM one by one from the target file data, or by inputting all the file data into the LLM. Also, the prompt that was the basis for creating the script to be modified may be input further.

The scripts stored by the script storage unit described above can also be configured to be modified in the same way as generated scripts. Since the stored scripts may need to be modified if the target file data increases, decreases, or changes, it is preferable to configure it so that this can be modified.

<Detailed prompt settings section>
The detailed prompt setting unit 508 has a function of setting a detailed prompt, which is a detailed prompt for acquiring desired information designated by the set prompt. The setting items of the detailed prompt described below include items to be applied in script generation, items to be applied to target file data, items to be applied when acquiring desired information from target file data by executing the generated script, and items to be applied in post-processing using LLM after the execution of the script. The manner in which the detailed prompt is applied can be set in various ways.

Figure 6 is a diagram illustrating the detailed prompt setting items (name, description, necessity) and a setting example (example). In this example, the prompt set to specify the desired information is "Please output the collision points between the columns and beams."

For example, setting item 1 is called "Assign multiple desired information names," and is indicated as a setting item for "accurate understanding of character definitions." The explanation for this is "explicitly assign multiple desired pieces of information to be extracted," with "["pillar,""Pillar"]" given as an example. By setting this detailed prompt, an instruction is added to the instruction "Output the collision points between the pillar and beam," to include not only "pillar," but also anything marked as "Pillar" as a target for extraction.

Other detailed prompt settings include "Add description of desired information," which adds an explanation of the desired information to be extracted to ensure an accurate understanding of the character definition, "Format specification," which allows the user to specify the format they want to extract in order to prevent erroneous extractions and standardize the format of output results, "Condition setting," which changes the extracted content according to specific conditions in order to improve accuracy and handle exceptions, and "Show sample script," which sets in advance a script that LLM will use as a reference when generating scripts in order to improve the accuracy of script generation in LLM. Of course, users can set various settings themselves to configure detailed prompts.

Methods for applying each of the settings items mentioned above can be broadly divided into three: Method 1, "narrowing down file contents," Method 2, "connecting prompts," and Method 3, "checking and reformatting detection results." Method 1 involves narrowing down the file data to be input before inputting it into the LLM, and Method 2 involves connecting the configured detailed prompt to the main prompt. Method 3 involves reformatting the extraction results according to a specified format, etc.

Figure 7 shows how each setting item of the detailed prompt is applied. As shown in the figure, the items "Assigning multiple desired information names," "Assigning description of desired information," "Condition setting," and "Show sample script" are applied using method 2 "Connecting prompts." Additionally, for "Format specification," all of methods 1 to 3 are applied. Examples of application will be described later.

Figure 8 is a conceptual diagram showing an example of an input screen when setting a detailed prompt. As shown in the figure, the already set prompt "Please output the collision point between the column and the beam" is displayed in the upper left part of the screen 801. Then, input fields 802 for each setting item of "desired information name", "format specification", "condition setting", and "sample script" are displayed on the left side of the screen. In addition, input fields for further setting items are displayed by the user's operation. The detailed prompt is set by accepting the user's input operation in each of these input fields. Then, the script generation unit applies the set detailed prompt to generate a script, or the desired information acquisition unit applies the set detailed prompt to acquire the desired information. By setting the detailed prompt in this way, the accuracy of acquiring the desired information can be improved.

Here is an example of applying the above detailed prompt. For example, if you set "'Pillar' also refers to 'column'" in the "Desired information name" field and set "If pillar cannot be extracted, output blank" in the "Condition setting" field, this detailed prompt will be connected to the main prompt and input. Then, by applying the "Desired information name", the 3D model marked "Pillar" will also be extracted as a pillar, and the collision point between the pillar and beam will be output. Also, if a pillar cannot be extracted from the target file data, a blank will be output.

In addition, when setting the detailed prompt, the detailed prompt setting unit can be configured to automatically set each setting item according to the set prompt. Alternatively, it may be configured to accept the settings made by the user while recommending some settings for items that have not been set. By accepting the detailed prompt settings from the user, extraction can be performed flexibly and appropriately in accordance with the user's wishes, 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.

In addition, for simplification and improved operability, it is also preferable to configure the desired information acquisition unit to perform a test applying the set prompt and detailed prompt before executing the script on all of the target file data. This test involves executing the script on a portion of the target file data. Based on the results of this test, the user can determine whether the set prompt and detailed prompt will bring about the results he or she desires, and can reconfigure the prompts, etc., to obtain more desirable acquisition results.

As shown in FIG. 8, a "Test" button 804 is displayed on the left side of screen 801 below the detailed prompt input field 802 and the prompt display field 803, and a test is performed by pressing this button. The test results are then displayed in a dotted frame 805, which is shown for explanatory purposes in the figure. Based on the results extracted from this test, the user can reset the prompt or perform actual retrieval processing using the script that obtained the test results.

In this example, the detailed prompt is set separately from the prompt that specifies the desired information, but the prompt and the detailed prompt may be configured to be set continuously and integrally.

<Verification Department>
The verification unit 509 has a function of verifying the results obtained by the desired information acquisition unit. There are various specific aspects of the verification by the verification unit. The verification described below may be performed by inputting a prompt to the LLM, or by executing a coded script.

Figure 9 is a diagram illustrating various settings when performing verification. As shown in the figure, four types of settings can be made: "Use detailed prompts," "Verification prompts," "Verification script," and "Compare files."

"Use detailed prompt" uses the extraction settings set in the detailed prompt for validation. For example, if a format is specified in the "Format specification" setting item of the detailed prompt, the specified format is compared with the extraction results to validate whether "the extraction results are within a specified similarity range with the specified format." In addition, the settings in "Condition settings" are used to validate whether "if extraction is not possible, are blanks output?"

The "verification prompt" allows you to set your own verification prompt; for example, you can set and input a prompt such as "Please check whether this information indicates the material the elevator doors are made of." The "verification script" allows you to set your own verification script; for example, as shown in the figure, you can input a command to return "Error" if the detection result is less than "1," and "OK" otherwise.

In addition, when extracting from multiple files, "file comparison" compares the extracted results with other files to confirm validity. As an example, when comparing extracted files A to D, file A, file C, and file D are all written using numbers and symbols only, while file B is written mainly in kanji. In such a case, the verification result is that "file B is not valid compared to the others."

Here, when the verification unit verifies the obtained results, the script modification unit described above can be configured to modify the script to reflect the verification results. For example, if the information in the extraction result is an error as described above, an instruction such as "In the previous extraction, information XXX was extracted, but the verification result was an error. Please modify the script taking this error result into consideration" is input to the LLM to modify the script.

<Hardware configuration>
The information processing apparatus of this embodiment can be realized by the hardware configuration shown in Fig. 3 according to the embodiment 1. The functions of a script evaluation unit 506, a script correction unit 507, a detailed prompt setting unit 508, and a verification unit 509, which are components specific to the information processing apparatus of the embodiment 2 shown in Fig. 5, are mainly realized by the CPU 301 and the RAM 302 in Fig. 3.

<Processing flow>
The process flow of the information processing device of this embodiment is basically the same as the process flow of the information processing device of embodiment 1. The information processing device of this embodiment further includes a script evaluation step of evaluating a result of obtaining desired information from sample file data by the generated script in comparison with correct answer information, a script correction step of inputting the generated script and a part or all of the target file data into a trained language model and correcting the generated script to adapt it to the target file data, a detailed prompt setting step of setting a detailed prompt that is a detailed prompt for obtaining desired information by the set prompt, and a verification step of verifying the acquisition result obtained in the desired information acquisition step.

＜Effects＞
According to the information processing apparatus of this embodiment, it is possible to provide an information processing apparatus which aims to improve the accuracy and robustness of desired information acquisition.

<Additional features>
Based on the first or second embodiment, the information processing device may be configured with the following functional configuration added thereto.

<Filtering section>
The filtering unit has a function of removing predetermined file data from the target file data to be acquired by the desired information acquisition 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 generating a script for filtering by performing semantic search, clustering, character search, etc., and executing the script. Filtering may also be performed by giving instructions to the LLM. For example, by inputting an instruction such as "Does this file contain elevator contents?", filtering can be performed based on whether or not the file contains elevator contents.

This filtering function can meet the needs of users who, for example, want to obtain the area from the architectural summary of 500 target file data, but want to skip the elevation drawings.

<Conversion section>
The conversion unit has a function of converting the format of the target file data according to the characteristics of the file data. If the information processing device has a conversion unit, the desired information acquisition unit acquires the desired information from the file data converted by the conversion unit. For example, if the file data is in PDF format, it is converted into a text file format. In addition, in the case of image data, character recognition is performed by OCR to convert the data into a text file format. In addition, in the case of three-dimensional data, it may be converted into two-dimensional data or an image. Such conversion contributes to improving robustness.

Furthermore, when converting to a text file, it is also preferable to convert so that the location information of the text is added. FIG. 10 is a conceptual diagram showing an example of converting PDF format file data into a text file with location information added. FIG. 10(a) shows a part of a PDF format file showing a building area table of a house. FIG. 10(b) shows an example of converting each text of "shape", "formula", and "area" into a JSON format in which the text is described together with their location information. FIG. 10(c) shows an example of converting text into an Array format in which the text is similarly described together with location information. In this way, by adding location information, extraction is performed taking into account the relationship between the texts, which contributes to improving robustness. FIG. 10(d) shows an example of converting text showing a table showing the "area" and "tatami" of each of "Western-style room 1", "Western-style room 2", and "Western-style room 3" in the original file into text with table information instead of location information (within the dotted line frame in the figure). By converting in this way, it is possible to obtain the target while maintaining the table structure.

Figure 11 is a conceptual diagram showing another example of converting the format of file data. As shown in Figure 11, in the case of a file containing graphics (a kitchen floor plan), converting it to SVG format and making it text with graphic information makes it easier to retrieve using a script. Conversely, two-dimensional vector information may be converted into a raster image, or data may be converted from three-dimensional to two-dimensional.

<Data Reduction Department>
The data reduction unit has a function to reduce the target file data. Data reduction can be performed in various ways. For example, data reduction can be performed by narrowing down the target file data to a specific space, property, area, specific page or section, paragraph, table, etc. In addition, filtering can be performed by file attributes (page number, position, etc.), individual keywords (semantic search, clustering, character search), or spatial position (x, y, z, coordinates). Filtering can also be performed by inputting instructions to the LLM ("Does this file contain elevator contents?").

Figure 12 is a conceptual diagram showing another aspect of data reduction. As shown in Figure 12 (a), a specific portion 1202 is cut out from one file data 1201, and it is checked whether the target data is included therein. If the target data is included as a result, that portion is used for the acquisition process to acquire the desired information. If the target data is not included, a portion 1203 shifted from the previous portion is cut out and checked in the same manner. In this way, by excluding from the acquisition target those that do not contain the target data in the check before acquisition, it is possible to improve processing efficiency. Note that this check may be performed by executing a script, using LLM, or by other methods. Also, such processing may be performed in a three-dimensional space.

As shown in FIG. 12(b), the above-mentioned check is performed while shifting the specific location, and the location containing the desired data is temporarily stored in memory 1204. Then, it is checked whether processing memory 1204 contains all the elements necessary to acquire the desired data, and if it does, the location held in the memory is provided for acquisition processing by the desired information acquisition 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-mentioned check are repeated.

<Basis display section>
The reason display unit has a function of displaying the reason why the result obtained by the desired information obtaining unit was obtained.

This information processing method allows the user to set a prompt to specify the desired information, input it into the LLM along with sample file data, and execute the generated script to obtain the desired information. However, it can be difficult to set a prompt to generate an optimal script to obtain the desired information. Therefore, by displaying the reasons (intermediate processes, etc.) that led to the obtained results, the user can be shown "what can be done to improve the obtained results."

Figure 13 is a conceptual diagram showing an example of displaying a position on an image as evidence. The figure shows a PDF file showing a building area table of a house, and the position of the information (text) obtained from this file is highlighted on the image by being surrounded by a thick dotted frame 1301. This allows the user to understand where the results were obtained from. Also, for scripts that process 3D models, displaying the process of the script and highlighting the information used at that time are useful for confirming the evidence.

It is also possible to display an explanation of the reason for the acquisition. For example, as a pre-processing step, an instruction such as "Please output the reason why this information was output as the acquisition result for each detailed process" can be input to the LLM before acquisition. Then, after acquisition, the acquisition result is displayed, and the reason is displayed, for example, "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, the user can reflect on various settings by being shown the effect of the prompt that he or she set and the script generated by it on the acquisition result, which can contribute to optimizing the acquisition of desired information.

Example 3
＜Overview＞
The present embodiment is an information processing apparatus similar to the information processing apparatus of the first or second embodiment.

<Functional configuration>
14 is a block diagram showing an example of the functional configuration of an information processing apparatus according to the present embodiment. As shown in FIG. 14, the information processing apparatus 1400 includes a script storage unit 1401, a stored script correction unit 1402, and a second desired information acquisition unit 1403.

<Script storage section>
The script storage unit 1401 has a function of storing a script for acquiring desired information, which is information on a desired item, from a plurality of file data including a plurality of target items. In the first embodiment, the script generated by the script generation unit was the subject of modification by the script modification unit, but in this embodiment, a script created and set in advance is stored in the script storage unit. The contents of the stored script are the same as those in the first and second embodiments, so a repeated description will be omitted.

<Retention script modification part>
The retained script correction unit 1402 has a function of inputting part or all of the target file data into the trained language model, and correcting the retained script to adapt it to the target file data.

The function of the stored script correction unit is the same as that of the script correction unit in the first and second embodiments, except that the scripts stored in the script storage unit are the targets of correction. Therefore, a redundant explanation will be omitted here.

It can also be configured to provide prompts to modify the stored scripts by inputting them into the trained language model. The prompts instruct the script on how to modify it, such as specifying the file type or format that should be applied to the acquisition target, or specifying data that should be removed.

<Second desired information acquisition unit>
The second desired information acquisition unit 1403 has a function of acquiring the desired information from the target multiple file data by the modified script. Other than being a script modified by the modifying unit, the function of the desired information acquiring unit is the same as that of the desired information acquiring unit in the first and second embodiments, so a redundant description will be omitted here.

<Hardware configuration>
The information processing device of this embodiment can be realized by the hardware configuration shown in Fig. 3 according to the embodiment 1 or 2. The functions of the stored script correction unit 1402 and the second desired information acquisition unit 1403 of the information processing device of the embodiment 3 shown in Fig. 14 are mainly realized by the CPU 301 and the RAM 302 in Fig. 3. Also, the function of the script storage unit 1401 is mainly realized by the storage 303 in Fig. 3.

<Processing flow>
15 is a flow diagram showing an example of the process flow of the information processing device of this embodiment. First, a script for acquiring desired information, which is information on a desired item, from a plurality of file data including a plurality of target items is stored (S1501: script storage step). Then, a part or all of the target file data is input to a trained language model, and the stored script is modified to adapt to the target file data (S1502: stored script modification step). Then, the modified script is used to acquire the desired information from the plurality of target file data (S1503: conversion step).

＜Effects＞
According to the information processing apparatus of this embodiment, by making it possible to modify a script that has been set up and stored in advance, it is possible to obtain desired information using a script that is adapted to the target file data.

100, 500: Information processing device 101, 501: Prompt setting unit 102, 502: Sample setting unit 103, 503: Script generating unit 104, 504: Script storing unit 105, 505: Desired information acquiring unit 506: Script evaluating unit 507: Script correcting unit 508: Detailed prompt setting unit 509: Verification unit 301: CPU
302: RAM
303: Storage 304: Communication interface 305: Input/output interface 1400: Information processing device 1401: Script holding unit 1402: Hold script correction unit 1403: Second desired information acquisition unit

Claims (13)

a prompt setting unit for setting a prompt for designating desired information from file data including the desired information;
a sample setting unit that sets sample file data to be input to a trained language model together with the set prompt;
a script generation unit that inputs the set prompt and the set sample file data into the trained language model to generate a script for acquiring specified desired information;
The script generation unit is an information processing device that further inputs correct answer information that is correct answer information of the specified desired information for the sample file data to generate the script. The information processing device according to claim 1, further comprising a script storage unit that stores the generated script. The information processing device according to claim 1 or 2, further comprising a desired information acquisition unit that acquires the specified desired information from the file data using the generated script. The information processing device according to claim 3, further comprising a script evaluation unit that evaluates the result of obtaining the specified desired information from the sample file data using the generated script in comparison with the correct answer information. The information processing device according to claim 4, wherein the script generation unit further inputs the evaluation result by the script evaluation unit to generate the script. The information processing device according to claim 3 , further comprising a script correction unit that inputs the generated script and a part or all of the file data into the trained language model and corrects the generated script to adapt it to the file data. The information processing device according to claim 6, wherein the script correction unit further inputs a prompt that is the basis of the generated script into the trained language model to correct it. 8. The information processing apparatus according to claim 6 , which refers to claim 2 , or claim 7 , which refers to claim 2, wherein the script correction unit also corrects the script stored in the script storage unit. The information processing device according to claim 1 or 2, further comprising a detailed prompt setting unit that sets a detailed prompt, which is a detailed prompt for acquiring desired information specified by the set prompt. The information processing apparatus according to claim 6 , further comprising a verification unit that verifies the acquisition result acquired by the desired information acquisition unit. The information processing apparatus according to claim 10 , wherein the script correction unit corrects the script by reflecting a result of the verification by the verification unit. An information processing method executed by an information processing device,
a prompt setting step of setting a prompt for designating desired information from file data including the desired information;
a sample setting step of setting sample file data to be input to the trained language model together with the set prompt;
and a script generating step of inputting the set prompt and the set sample file data into the trained language model to generate a script for acquiring specified desired information,
The script generating step is an information processing method in which answer information that is an answer to the specified desired information for the sample file data is further input to generate the script. a prompt setting step of setting a prompt for designating desired information from file data including the desired information;
a sample setting step of setting sample file data to be input to the trained language model together with the set prompt;
a script generation step of inputting the set prompt and the set sample file data into the trained language model to generate a script for acquiring specified desired information;
An information processing program for causing an information processing device to execute the
The script generating step is an information processing program that generates the script by further inputting correct answer information that is correct answer information of the specified desired information for the sample file data.