Document output systems, programs, and document output methods.

Abstract

This system provides stable output when automating the task of creating a specific document from a given document using generation AI. [Solution] The document output system is a system that includes a processor and generates an output document from an input document using a learning model. The processor inputs a class obtained from a class definition that defines the structure of the output document, an explanatory text about the information manipulation of the input document, and the input document to the learning model. The learning model outputs a structured object, which is an object in which information corresponding to the elements of the class has been extracted from the input document. The structured object is converted into an output document using a predetermined function, and the output document is output.

Description

Technical Field

[0001] The present invention relates to a document output system, a program, and a document output method. Background Art

[0002] Patent Document 1 describes an explanatory sentence generation method in which a learning document dividing unit divides each learning document, which is each document in a learning document set, into predetermined units, thereby constructing a learning document component sequence which is a sequence of learning document components that are constituent elements of the divided learning documents. Prior Art Literature Patent Literature

[0003] Patent Document 1 Japanese Unexamined Patent Application Publication No. 2018-28866 Summary of the Invention Problems to be Solved by the Invention

[0004] In recent years, efforts have been made to introduce generative AI into business processes and automate part or all of the work conventionally performed by humans. On the other hand, for work that has been conventionally performed by humans, the conditions necessary for the work are often utilized as the experience of the person in charge, and the conditions are often not documented as so-called tacit knowledge. Therefore, even when an attempt is made to automate work using generative AI, unstable output may result because the conditions for inference by the generative AI are unclear. The present invention provides a system capable of performing stable output when automating the task of creating a predetermined document from a specific document using generative AI. Means for Solving the Problems

[0005] The invention according to claim 1 is a system that includes a processor and generates an output document from an input document using a learning model, wherein the processor: The learning model is given reference materials for the input document and reference materials for the output document, and outputs a class definition that defines the structure of the output document, a predetermined function, and an explanatory text regarding the manipulation of the information in the input document. , The aforementionedClasses obtained from class definitions The aforementioned Descriptive text, The aforementioned The learning model inputs an input document, and outputs a structured object from the learning model, which is an object that extracts information corresponding to the elements of the class from the input document. The aforementioned This is a document output system that converts data into an output document using a predetermined function and then outputs the output document. Claim 2 The invention described herein is a claim in which the class definition is information that classifies the attributes of the information items included in the output document, which is the structure of the output document, according to each attribute. 1 Document output system That is . Claim 3 The invention described herein is a claim that the predetermined function is a function that converts the structured object into the format of an output document. 1 This is the document output system described. Claim 4 The invention described above is The chapter is The claim is a document that includes rules for extracting information from an input document to associate the information from the input document with elements of a class. 1 This is the document output system described. Claim 5 The invention described herein is a program that enables a computer to perform a function of generating an output document from an input document using a learning model, wherein the computer, A function to input reference materials for the input document and reference materials for the output document into the learning model, and a function to output a class definition that defines the structure of the output document, a predetermined function, and an explanatory text regarding the manipulation of information in the input document from the learning model. , The aforementioned Classes obtained from class definitions The aforementioned Descriptive text, The aforementioned A function for inputting an input document to the learning model, a function for outputting a structured object from the learning model which is an object obtained by extracting information that corresponds to the elements of the class from the input document, and the structured object The aforementioned This program implements the functions of converting a document into an output document using a predetermined function, and outputting the said output document. Claim 6 The invention described above is a method for generating an output document from an input document using a terminal and a server having a learning model, The terminal transmits reference materials for the input document and reference materials for the output document to the server; the server inputs the reference materials for the input document and reference materials for the output document to the learning model; and the server outputs a class definition defining the structure of the output document, a predetermined function, and an explanatory text regarding the manipulation of the information in the input document from the learning model to the terminal. , The terminal transmits the class obtained from the class definition, the explanatory text, and the input document to the server. The aforementioned server, Obtained from the aforementioned class definition The class, the descriptive text, and the input document are input to the learning model, and the server causes the learning model to output a structured object, which is an object obtained by extracting information corresponding to the elements of the class from the input document, to the terminal, and the terminal outputs the structured object The aforementioned A document output method that converts to the output document using a predetermined function and outputs the output document. That is . [Effects of the Invention]

[0006] This invention provides a system that can produce stable output when automating the task of creating a predetermined document from a specific document using generation AI. [Brief explanation of the drawing]

[0007] [Figure 1] This figure shows an example configuration of the document output system 1 according to this embodiment. [Figure 2] This figure shows an example of the hardware configuration of a document generation server to which this embodiment is applied. [Figure 3] This figure shows an example of the hardware configuration of a terminal to which this embodiment is applied. [Figure 4] This diagram illustrates the process of outputting structured information. [Figure 5] This diagram illustrates the process from initial creation to outputting the final document. [Figure 6] This flowchart shows the processing flow of the document generation server when using meeting minutes and transcript data. [Figure 7] This diagram illustrates the processing flow of the terminal after the document generation server outputs an object to the terminal. [Figure 8] (a) and (b) are diagrams showing examples of class definitions and structured objects when using meeting minutes and transcript data. [Figure 9]Figs. 1A and 1B are diagrams illustrating an example of a conventional workflow for press release operations and a workflow for press release operations using a learning model according to the present embodiment, respectively. MODE FOR CARRYING OUT THE INVENTION

[0008] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. FIG. 1 is a diagram showing a configuration example of a document output system 1 according to the present embodiment. The document output system 1 includes a document generation server 100. The document output system 1 includes a terminal 200. The document generation server 100 includes a learning model 150. The document generation server 100 and the terminal 200 are connected via a network 90.

[0009] The document generation server 100 is a server that generates an output document based on an input document. The input document is document data input to the learning model 150. The output document is document data obtained by formatting data output from the learning model 150 when the input document is input to the learning model 150. Examples of the input document include transcription data of meetings and hearing sheets that describe content to be pressed before issuing a press release for a new product or the like. Examples of the output document include meeting minutes and draft press releases created from hearing sheets.

[0010] The document generation server 100 converts tacit knowledge inherent in operations into explicit knowledge and extracts it from the correspondence between input documents and output documents. Then, the document generation server 100 creates an output document from an input document in accordance with a certain rule that includes some tacit knowledge. Tacit knowledge is knowledge based on an individual's experience and intuition. Tacit knowledge inherent in operations refers to conditions for operations that have not been clarified among the operation conditions conventionally handled by humans. Examples of tacit knowledge inherent in operations include operation conditions that have not been clarified when creating an output document from an input document.

[0011] The learning model 150 is a pre-trained model trained on a vast amount of text data, known as an LLM, and is a pre-trained model specialized for natural language inference processing. The learning model 150 is an example of a learning model. Note that the learning model 150 is not limited to a model specialized for natural language inference processing, but may also be a general-purpose model. The learning model 150 is not limited to being provided on the document generation server 100, but may also be provided on an external server, such as a server that provides the learning model 150. In this case, the document generation server 100 can use the learning model 150, for example, via API communication.

[0012] Terminal 200 is a terminal operated by a user of the document output system 1. By operating terminal 200, the user inputs and outputs information to the document generation server 100. For example, by operating terminal 200, the user can output an input document to the document generation server 100 or retrieve an output document from the document generation server 100.

[0013] <Document generation server hardware configuration> Figure 2 shows an example of the hardware configuration of the document generation server 100 to which this embodiment is applied. As shown in the figure, the document generation server 100 includes a processor 101, which is a calculation means, and a main memory 102 and an auxiliary memory 103, which are storage means. Various calculation circuits such as a CPU (Central Processing Unit), GPU (Graphics Processing Unit), ASIC (Application Specific Integrated Circuit), and FPGA (Field-Programmable Gate Array) can be used as the processor 101. The processor 101 reads the program stored in the auxiliary memory 103 into the main memory 102 and executes it. For example, RAM (Random Access Memory) can be used as the main memory 102. For example, a magnetic disk drive or SSD (Solid State Drive) can be used as the auxiliary memory 103. The document generation server 100 also includes a communication interface 104 for sending and receiving information to and from external devices via a network.

[0014] <Device Hardware Configuration> Figure 3 shows an example of the hardware configuration of terminal 200 to which this embodiment is applied. As shown in the figure, the terminal 200 includes a processor 51, which is an arithmetic means, and a main memory 52 and an auxiliary memory 53, which are storage means. Various arithmetic circuits such as a CPU, GPU, ASIC, FPGA, etc. can be used as the processor 51. The processor 51 reads the program stored in the auxiliary memory 53 into the main memory 52 and executes it. RAM can be used as the main memory 52, for example. A magnetic disk drive or SSD can be used as the auxiliary memory 53, for example.

[0015] The terminal 200 also includes a display device (display) 54, an input device 55 for operator input, and a communication interface 56 for sending and receiving information to and from an external device via a network. For example, the input device 55 may be a touch panel integrated with the display device 54. Alternatively, a keyboard or mouse may be used as the input device 55. Note that the hardware configurations shown in Figures 2 and 3 are merely examples and are not limited thereto. For example, a configuration that includes non-volatile memory such as flash memory or ROM (Read Only Memory) as storage devices is also possible.

[0016] In this embodiment, each process is executed on any computer. Any computer may be implemented as a processor as hardware, a program as software, or a combination thereof. Any computer may be a general-purpose computer, a computer designed for a specific purpose, a workstation, or any other system capable of performing each of these processes.

[0017] The processor is configured to perform various processes in cooperation with the program. The processor can function as each unit or each means in this embodiment. The execution order of the processes performed by the processor is not limited to the order described in this embodiment and can be changed as needed.

[0018] A processor can be configured with one or more hardware components. The types of hardware that make up a processor are not limited to any particular type. For example, a processor may be a CPU (=Central Processing Unit), an MPU (=Micro Processing Unit), a programmable logic device such as an FPGA (=Field Programmable Gate Array), a dedicated circuit for performing specific processing such as an ASIC (=Application Specific Integrated Circuit), a GPU (=Graphic Processing Unit), or hardware such as an NPU (=Neural Processing Unit).

[0019] A processor can be configured not only with a combination of multiple hardware components of the same type, but also with a combination of multiple hardware components of different types. When multiple hardware components are configured to perform one or more processes of a given processor, these components may reside in physically separate devices or in the same device. Hardware is composed of electrical circuits and other components, such as semiconductor elements. In any embodiment, the execution order of each process by the processor is not limited to the order described in each embodiment, and can be changed as necessary.

[0020] The program can be firmware, or it can be software such as microcode. The program may be, for example, a group of program modules. Each function constituting the group of program modules may be implemented by a processor configured to execute each function. The program in each embodiment may be program code or multiple code segments stored in one or more non-temporary computer-readable media (e.g., semiconductor memory, magnetic or optical storage media, or other storage).

[0021] A program may be divided and stored on multiple non-temporary computer-readable media located on devices that are physically separated from each other. Program code or multiple code segments may be represented by any combination of procedures, functions, subprograms, routines, subroutines, modules, software packages, classes, instructions, data structures, and program statements. Program code or multiple code segments may be connected to other code segments or hardware circuits by sending and receiving information, data, arguments, parameters, or memory contents. The present invention is also applicable to programs and program products.

[0022] Figures 4 and 5 are illustrative diagrams of the processing using the learning model 150 of the document output system 1. Figure 4 is an illustrative diagram of the flow up to the output of each structured piece of information. Figure 5 is an illustrative diagram of the flow up to the output of the final output document. In Figures 4 and 5, A, B, and C are symbols indicating that the processing flow is continuous. In Figures 4 and 5, meeting transcript data is used as an example of an input document, and meeting minutes data is used as an example of an output document.

[0023] First, as shown in Figure 4, the document generation server 100 obtains a model input document and a model output document from the terminal 200. The model input document and model output document are exemplary documents for the input document and output document obtained when the input document to be verified is input to the learning model 150 and an output document is obtained. The model input document is, for example, an input document used in past business operations. The model output document is an output document created in past business operations using the input document.

[0024] The document generation server 100 then inputs a model input document and a model output document to the learning model 150. The document generation server 100 then instructs the learning model 150 to output three elements: a class definition, a rendering function, and an information manipulation guide. These three elements are used to output a structured object from the input document. The process of outputting these three elements clarifies the conditions for the tasks that a human would perform when creating an output document from an input document. The class definition and rendering function are information used in programming languages ​​such as Python.

[0025] A class definition is a string that defines the structure of an output document. A class definition is a code snippet. A class definition can be instantiated and used as a class in the Python environment. The structure of an output document refers to the attributes of the information items contained in the output document. The attributes of the information items refer to the information that indicates the attributes of the characters used to make up the output document. A class is a data structure that classifies and holds the information of an output document by attribute, and is a category of attributes. A class definition clarifies the conditions of the business, for example, the conditions of the characters used to make up the output document. Note that the class definition is output as a string that can be used with Pydantic, a library that guarantees consistency.

[0026] A class definition contains multiple strings that represent the class. It also contains a string defining the name of each class. Furthermore, it contains strings defining elements used to identify each class. These elements include information about attributes that are essential for identifying a class, and information about attributes that are not essential but may be used. The information used to identify a class is sometimes referred to as the class's elements.

[0027] A rendering function is a function used to convert elements of a defined class into a format such as text, which is the format of the output document. The format of the output document refers to information about the layout of the output document. For example, the format of the output document may include information about the arrangement of characters in the output document. Another example is information about the font of the characters in the output document. A rendering function is a code snippet. A rendering function can be used in the Python environment by instantiating it. A rendering function clarifies the conditions of a business, such as the arrangement and size of characters used in the output document.

[0028] An information manipulation guide is supplementary information used to generate an output document using input documents, class definitions, and rendering functions. For example, it is a descriptive document explaining how to manipulate information in an input document when filling in the elements of a class. For instance, the manipulation guide contains rules for extracting information from the input document and rules for transforming information in order to map the information from the input document to the elements of a class. The information manipulation guide is used in the process shown in Figure 5. The information manipulation guide clarifies the conditions of the process, such as the tasks of extracting characters from the input document and transforming characters in order to create an output document.

[0029] The document generation server 100 then outputs the class definition and rendering function, which are code fragments, to the terminal 200. The terminal 200 then instantiates the class definition and rendering function. Instantiation is the process of dynamically defining the class definition and rendering function, which are code fragments, by executing them with the exec function. By instantiating them, the class definition and rendering function become available for use in the Python environment. The class obtained through instantiation and the instantiated rendering function are then used in the process shown in Figure 5.

[0030] Then, as shown in Figure 5, the document generation server 100 obtains the materialized class, information operation guide, and verification input document from the terminal 200. The materialized class, information operation guide, and verification input document are then input to the learning model 150. The verification input document is used to verify whether or not an output document in the same format as the example output document can be produced from the input document.

[0031] The document generation server 100 then outputs a structured object from the learning model 150. A structured object is a class object, which is an object that extracts information that corresponds to the elements of the class from the input document used for validation. For example, a structured object is information that associates the elements of a class with the information in the input document that corresponds to those elements of the class, and this information is organized and written for each element of the class. Furthermore, the structured object is an object that has undergone automatic data validation by Pydantic. In other words, the structured object is an object that has been checked to see whether the information extracted from the input document used for validation conforms to the format of the elements of the class.

[0032] The document generation server 100 then outputs the structured object to the terminal 200. The terminal 200 then transforms the structured object into the final document format using an instantiated function. For example, it converts it to a text format, which is the format of the output document. The final output document is then generated.

[0033] As shown in Figure 4, the information manipulation guide, rendering function, and class definition can be input into the learning model 150 to generate a know-how document for humans as an output document. A know-how document for humans is a document that shows the procedure for when a human performs the task of creating an output document using the input documents.

[0034] Figure 6 is a flowchart showing the processing flow of the document generation server 100 when using meeting minutes and transcript data. First, the document generation server 100 obtains reference meeting transcript data and reference meeting minutes data as the model input document and model output document (step 601). The reference meeting minutes data is, for example, created by a human from the reference meeting transcript data. The reference meeting transcript data and reference meeting minutes data are obtained from terminal 200. Next, the document generation server 100 outputs, as an example of three elements, a class that defines the components of the meeting minutes, a function that converts the class into a text format like meeting minutes, and a document that shows how to manipulate the information (step 602). The components of the meeting minutes are an example of a class. Examples of components of the meeting minutes include a participant class that shows the attributes of the participants, a date and time class that shows the date and time attribute of the meeting minutes, and a to-do class that shows the attribute of to-dos, which are future tasks decided in the meeting minutes. The document describing the method of information manipulation is an example of an explanatory document, and it explains the method of information manipulation from transcribed data to meeting minutes. Step 602 extracts the conditions for the work involved in creating meeting minutes from transcribed data.

[0035] Then, the document generation server 100 obtains the class instantiated by the exec function from the terminal 200 (step 603). Then, the document generation server 100 inputs the text describing how to manipulate information, the instantiated class, and the transcription data for verification to the learning model 150 based on instructions from the terminal 200 (step 604). Then, the document generation server 100 obtains the object of the defined class output by the learning model 150 (step 605). The object is an example of a structured object; for example, if the participant's attribute information is an element of the class, it contains information that the participant's attribute information is an element of the class, and information extracted from the participant's name corresponding to the participant's information. Then, the document generation server 100 outputs the object to the terminal 200 (step 606).

[0036] Figure 7 illustrates the flow after the document generation server 100 outputs an object to the terminal 200. The terminal 200 obtains the object from the output of the document generation server 100 (step 701). Then, the terminal 200 converts the object using a function that converts it into a materialized text format (step 702). As a result, the output object retains its attribute information, while the format of the text layout and fonts are converted into the format of meeting minutes. As a result, the terminal 200 obtains meeting minutes for the transcription data used for verification.

[0037] Figures 8(a) and 8(b) show examples of class definitions and structured objects when using meeting minutes and transcript data. Figure 8(a) shows the class definition, and Figure 8(b) shows the structured object.

[0038] Figures 8(a) and 8(b) show the "class Person" which indicates the class of the meeting participants, and the "class Date" which indicates the class of the meeting date and time. Figures 8(a) and 8(b) also include "name" which indicates the participant's name and "position_title" which indicates the participant's job title as elements of the participant class. Additionally, the "date" element is included as an element of the date and time class.

[0039] Figure 8(a) shows "Field," which describes the elements of the class, listed for each class element. Also in Figure 8(a), "str" ​​is listed as a class element that is required to define the class. Furthermore, in Figure 8(a), "Optional" is listed as a class element that is not necessarily required to define the class. Figure 8(b) shows "Ichiro Sato," "Director," and "2026 / 01 / 27," which were extracted from the transcribed data as information corresponding to the class elements.

[0040] Figures 9(a) and 9(b) illustrate the workflow of a conventional press release and an example of the workflow using the learning model 150 in this embodiment. Figures 9(a) and 9(b) show Department A, which is a designated department within an organization, and the press department, which is responsible for reporting within the organization. Department A is the department that has the content for which a press release is to be issued. The press department is the department that checks the content of the press release draft.

[0041] Traditionally, Department A would present the press release content to the media department. If the media department found no issues with the content, they would request Department A to create a hearing sheet. Department A would then create a draft press release based on past press releases. After several rounds of exchanges with the media department, the final press release would be issued. In this traditional process, Department A would spend a considerable amount of time and effort creating the draft press release and making subsequent revisions. Furthermore, the media department would spend a significant amount of time scrutinizing and providing feedback on the draft press release.

[0042] On the other hand, in this embodiment, the effort required to create and revise press releases can be reduced by using the learning model 150. Specifically, by inputting past interview sheets as example input documents and past press releases as example output documents into the learning model 150, class definitions, rendering functions, and information manipulation guides related to press releases can be obtained. Then, by inputting the interview sheets containing the class definitions, rendering functions, information manipulation guides related to press releases, and the content to be pressed into the learning model 150, a structured press release can be obtained. As a result, the press department only needs to check whether there is any missing information in the obtained press release, significantly reducing the time spent scrutinizing the content of the press release.

[0043] The above describes a document output system 1 that uses class definitions, rendering functions, and information manipulation guides to output output documents from a learning model 150. Conventional methods of obtaining output documents by inputting input documents and instruction sentences to the learning model 150 are greatly affected by the non-deterministic behavior of the learning model 150. On the other hand, this embodiment allows the behavior of the learning model 150 to be controlled by deterministic conditioning by using class definitions, rendering functions, and information manipulation guides. This improves the stability and reliability of business automation by the learning model 150. Furthermore, this allows, for example, the tacit knowledge of in-house experts to be formalized through the learning model 150, and can be reproduced even by users without specialized knowledge. This can lead to the standardization of business quality, the promotion of business succession, and the improvement of business productivity. [Explanation of Symbols]

[0044] 1…Document output system, 100…Document generation server, 150…Learning model, 200…Terminal

Claims

1. A system equipped with a processor that generates output documents from input documents using a learning model, The aforementioned processor, The reference materials for the aforementioned input document and the reference materials for the aforementioned output document are input to the learning model. The learning model outputs a class definition that defines the structure of the output document, a predetermined function, and an explanatory text regarding the manipulation of information in the input document. The class obtained from the class definition, the explanatory text, and the input document are input to the learning model. The learning model outputs a structured object, which is an object obtained by extracting information that corresponds to the elements of the class from the input document. The structured object is converted into an output document using the predetermined function. A document output system that outputs the aforementioned output document.

2. The document output system according to claim 1, wherein the class definition is information that classifies the attributes of the information items included in the output document, which is the structure of the output document, according to each attribute.

3. The document output system according to claim 1, wherein the predetermined function is a function that converts the structured object into the format of an output document.

4. The document output system according to claim 1, wherein the explanatory text is a text containing rules for extracting information from an input document to associate the information of the input document with elements of a class.

5. A program that enables a computer to generate an output document from an input document using a learning model, To the aforementioned computer, A function to input the reference materials for the aforementioned input document and the reference materials for the aforementioned output document into the learning model. A class definition that defines the structure of the output document, a predetermined function, and a function that causes the learning model to output explanatory text regarding the manipulation of information in the input document, A function for inputting the class obtained from the class definition, the explanatory text, and the input document to the learning model, The function includes outputting a structured object from the learning model, which is an object obtained by extracting information that corresponds to the elements of the class from the input document, A function to convert the structured object into an output document using the predetermined function, A program that implements the function of outputting the aforementioned output document.

6. A method for generating an output document from an input document using a terminal and a server having a learning model, The terminal transmits the reference materials for the input document and the reference materials for the output document to the server. The server inputs the reference materials for the input document and the reference materials for the output document into the learning model. The server outputs a class definition defining the structure of the output document, a predetermined function, and an explanatory text regarding the manipulation of the input document from the learning model to the terminal. The terminal transmits the class obtained from the class definition, the explanatory text, and the input document to the server. The server inputs the class obtained from the class definition, the explanatory text, and the input document to the learning model. The server causes the learning model to output a structured object, which is an object obtained by extracting information that corresponds to the elements of the class from the input document, to the terminal. The terminal converts the structured object into the output document using the predetermined function, and outputs the output document.

Images