Sentence analysis generating device, sentence analysis generating method and sentence analysis generating program

Abstract

To provide a sentence analysis generating device, a sentence analysis generating method and a sentence analysis generating program that specifically provide reply details and the basis thereof, such as a logic like the reason of the reply details and the basis, and further a document and a quotation of a part in such a document.SOLUTION: A sentence analysis generating device includes: an obtaining unit that obtains an initial sentence; a sentence analysis extracting unit that extracts, from multiple documents registered in advance, a description relating to the details described in the initial sentence, and the document; a reply sentence generating unit which generates a reply based on the initial sentence and on the description, and which generates a reply sentence by quoting the description or a part of the description; and a display unit that displays, on the same screen, the initial sentence, the reply sentence, the description, and the document.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to a text analysis generation device, a text analysis generation method, and a text analysis generation program. [Background technology]

[0002] Large-scale language models are learning algorithms capable of language recognition, summarization, translation, prediction, and generation, and are attracting attention for their ability to instantly answer a wide range of questions (see, for example, Patent Documents 1 to 3), and are beginning to spread to business and industry. This trend is expected to have an impact on all sectors in Japan, including corporate law.

[0003] Patent Document 1 discloses a technology for generating answers to questions and outputting evidence information. Patent Document 2 discloses a technology for acquiring questions, presenting answers, and presenting evidence. Patent Document 3 discloses a technology for extracting answer candidates to a question and presenting them to the user while effectively using background knowledge. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 7247878 [Patent Document 2] JP 2017-37372 A [Patent Document 3] JP 2022-67234 A Summary of the Invention [Problem to be solved by the invention]

[0005] However, in the legal field, where accuracy of answers is required, not only the content of the answer but also the logic leading to the answer and the presentation and citation of documents that serve as the basis are important elements. In addition, when applying large-scale language models in the legal field, if the basis of the answer is unclear to the user, it is difficult to determine whether the answer is a hallucination, or a lie that may be convincing, and the user may mistakenly recognize an answer that is not legally correct as correct.

[0006] In view of the above problems, the present invention aims to provide a text analysis generation device, a text analysis generation method, and a text analysis generation program that specifically provide the content of an answer and the basis for that content, such as the logic leading to the answer, including the reasons and basis for the answer, and also citations of documents and relevant parts of those documents. [Means for solving the problem]

[0007] A first aspect of the present invention is a text analysis and generation device comprising an acquisition unit that acquires an initial text, a text analysis and extraction unit that extracts a description related to the content written in the initial text and a document from a plurality of documents registered in advance, a response text generation unit that generates an answer based on the initial text and the description and generates a response text by quoting the description or a part of the description, and a display unit that simultaneously displays the initial text, the response text, the description, and the document.

[0008] In the first aspect of the present invention, the text analysis and extraction unit may extract meta-information in addition to descriptions and documents.

[0009] In the first aspect of the present invention, the response sentence generation unit may generate a response sentence based on the description or a part of the description, as well as a document and meta information.

[0010] In the first aspect of the present invention, the display unit may display meta information on the same screen in addition to the initial sentence, the response sentence, the description, and the document.

[0011] In the first aspect of the present invention, the plurality of documents registered in advance may be documents that satisfy a predetermined criterion.

[0012] In the first aspect of the present invention, a plurality of pre-registered documents may be classified according to a numerical value representing accuracy.

[0013] In a first aspect of the present invention, the pre-registered documents may include legal information.

[0014] In the first aspect of the present invention, the plurality of pre-registered documents may further include business information.

[0015] In a first aspect of the invention, the initial sentence may be a legal question and the response sentence may be a response to the initial sentence.

[0016] In a first aspect of the invention, the initial sentence may be a legal business instruction and the response sentence may be a response to the business instruction.

[0017] In the first aspect of the present invention, when the mouse pointer is placed on a desired portion of the response sentence, the display unit may highlight and display the description that was the basis for generating the desired portion.

[0018] In the first aspect of the present invention, the display unit may display, on the same screen, an entry field for inputting additional text, in addition to the initial text, the response text, the description, and the document.

[0019] In the first aspect of the present invention, a calculation unit may further be provided that calculates a numerical value representing legal accuracy and a numerical value representing practical information accuracy for each document or for each sentence in the document.

[0020] The first aspect of the present invention may further comprise a correction unit that corrects the response text so as to improve a numerical value representing the accuracy of legal or practical information of the response text.

[0021] In the first aspect of the present invention, the system may further comprise a feedback unit that feeds back information including the content corrected by the correction unit to the business information.

[0022] In the first aspect of the present invention, the response sentence generation unit may generate a response based on the extracted document in addition to the initial sentence and the description.

[0023] In a first aspect of the present invention, the answer may be in accordance with the law and / or rules within the organization based on the law.

[0024] In a first aspect of the present invention, when generating a response sentence, the response sentence generation unit may perform at least one of generating a file related to the response sentence, generating a URL that is a link destination for accessing the file related to the response sentence, or specifying a document stored in the document memory unit.

[0025] In a first aspect of the present invention, the system further includes an input unit, and after the acquisition unit acquires an initial sentence and the response sentence generation unit generates a response sentence as a first response sentence, the response sentence generation unit selects each time by the user between an operation of generating a response sentence again based on the initial sentence and an operation of generating another response sentence based on an initial sentence different from the initial sentence, and the operation accepted by the input unit may be repeated as many times as desired by the user.

[0026] In a first aspect of the present invention, the display unit may display all or a portion selected by the user of the descriptions extracted by the sentence analysis extraction unit and the documents from which the descriptions were extracted, and the response sentences generated by the response sentence generation unit, by repeating the operation selected by the user each time and received by the input unit as many times as the user desires.

[0027] A second aspect of the present invention is a text analysis and generation method comprising: an acquisition step of acquiring an initial text; a text analysis and extraction step of extracting a description, document and meta-information relating to the content described in the initial text from a plurality of documents registered in advance; a step of generating an answer based on the initial text and the description, the response text generation step of generating a response text by quoting the description and a portion of the description; and a display step of simultaneously displaying the initial text, the response text, the description and the document.

[0028] A third aspect of the present invention is a text analysis and generation program which enables a computer to realize an acquisition function for acquiring an initial text, a text analysis and extraction function for extracting descriptions, documents and meta-information relating to the contents of the initial text from a plurality of documents registered in advance, a response text generation function for generating an answer based on the initial text and the description, which generates a response text by quoting the description and a portion of the description, and a display function for simultaneously displaying the initial text, the response text, the description and the document.

[0029] According to the present invention, it is possible to provide a text analysis generation device, a text analysis generation method, and a text analysis generation program that specifically provide the content of an answer and the basis for that content, such as the logic leading to the answer, including the reasons and basis for the answer, and also citations of documents and relevant parts of those documents. [Brief description of the drawings]

[0030] [Figure 1] 1 is a block diagram showing an example of a text analysis and generation device according to a first embodiment of the present invention. [Diagram 2] FIG. 1 is a schematic diagram showing an example in which a document is structured and divided into a plurality of sections. [Diagram 3] FIG. 2 is a diagram showing an example of a screen displayed by a display unit of the writing analysis generation device according to the first embodiment. [Figure 4] 4 is a diagram showing a state in which a part of the description is enlarged on the screen shown in FIG. 3. FIG. [Diagram 5]4 is a flowchart illustrating the flow of a text analysis and generation method according to the first embodiment. [Figure 6] FIG. 11 is a block diagram showing an example of a text analysis and generation device according to a second embodiment of the present invention. [Figure 7] FIG. 11 is a diagram for explaining details of scoring by the text analysis generation device according to the second embodiment. [Figure 8] FIG. 13 is a diagram for explaining the generation of a file related to a response sentence by a sentence analysis generation device according to a fourth embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0031] Next, an embodiment of the present invention will be described with reference to the drawings. In the description of the drawings relating to the embodiment, the same or similar parts are denoted by the same or similar reference numerals. However, the drawings are schematic.

[0032] Furthermore, the embodiments are merely examples of devices and methods for embodying the technical idea of ​​the present invention, and the technical idea of ​​the present invention does not specify the configuration, arrangement, layout, etc. of each component to those described below. The technical idea of ​​the present invention can be modified in various ways within the technical scope defined by the claims.

[0033] (First embodiment) The text analysis and generation device of this embodiment acquires an initial text, extracts descriptions, documents, and meta-information related to the content of the initial text from multiple documents registered in advance, and generates a response text that conforms to the law and rules within an organization based on the law in response to the content of the initial text based on the extracted descriptions, documents, and meta-information.

[0034] The text analysis and generation device according to this embodiment will be described with reference to Fig. 1. As shown in Fig. 1, the text analysis and generation device 10 according to this embodiment is composed of a CPU 11 for executing various calculations, a storage unit 12 for storing processing programs, data, etc., an I / O (input-output interface) 13, a display unit 14, and an input unit 15.

[0035] The text analysis and generation device 10 may be various electronic computers (computational resources) such as a personal computer (PC), a mainframe, a workstation, a cloud computing system, or the like.

[0036] The I / O 13 is an interface, buffer, etc. for communication (transmission and reception).

[0037] The display unit 14 is a display device such as a display, and displays the results of calculations performed by the CPU 11, etc.

[0038] The input unit 15 is an input device such as a keyboard, a mouse, etc., and is realized by a device that receives input from a user and transmits it to the CPU 11.

[0039] The functional units within the CPU 11 are shown in the block diagram of Fig. 1. When each functional unit of the CPU 11 is realized by software, the CPU 11 realizes it by executing instructions of a program, which is software that realizes each function. In detail, the CPU 11 includes an initial sentence acquisition unit 111, a sentence analysis and extraction unit 112, a response sentence generation unit 113, etc. The storage unit 12 also includes a document storage unit 121.

[0040] The initial sentence may be, for example, data in text format. The initial sentence is a sentence for which a user desires to generate a response sentence in accordance with the law and rules within an organization based on the law, and may be, for example, a question related to the law. Alternatively, the initial sentence may be, for example, a business instruction related to the law. Alternatively, the initial sentence may be a conversation between one or more users. The initial sentence may be transmitted to the CPU 11.

[0041] The response sentence may be, for example, a response to a question. Or, the response sentence may be a sentence that summarizes a legal business instruction or a conversation between one or more users, and is generated in accordance with legal or practical information. Or, the response sentence may be a legal document required to execute the legal business instruction. In this embodiment, a case will be described in which the initial sentence is a legal question, and the response sentence is a response to the question.

[0042] The document storage unit 121 stores a number of pre-registered documents that are referenced by the text analysis and extraction unit 112 described later when extracting descriptions and documents related to the contents written in the initial text, as well as meta-information associated with the documents and text within the documents.

[0043] The pre-registered documents are legal information, practical information, etc. Legal information includes public information such as "treaties, constitutions and laws," "judgments and other court documents," "official gazettes, guidelines, public comments, minutes of councils, etc., notices, notifications, circulars, and other government documents," "decisions and rulings, etc.," "legal information such as the minutes of the Diet and parliamentary documents," "local government ordinances, rules, and parliamentary documents," as well as "explanatory books, explanatory articles, and explanatory papers on the above-mentioned public information," "statistical information on the above-mentioned public information," etc. Practical information includes, for example, practical documents that describe the content of legal practical processing within an organization, contract documents, internal information, internal regulations and guidelines, preparatory documents, opinions, records, email and chat information with inside and outside the organization, disclosure materials, and other internal documents, etc., that are created based on legal information.

[0044] The plurality of pre-registered documents may be written in any format. The plurality of pre-registered documents may be written, for example, as text data, or may be, for example, audio data, image data, etc. The plurality of pre-registered documents may be in a format that can be created by an application such as word processing software, spreadsheet software, presentation software, etc.

[0045] Meta information is information associated with a document or a sentence in a document. In this embodiment, the document may be structured, and when structured, the document may be composed of one or more sections. FIG. 2 shows an example of a document structured and divided into multiple sections. FIG. 2 shows "Explanation of the Premiums and Representations Act (10th Edition)," which is a commentary on the Premiums and Representations Act, as an example of a document. When the document 20 shown in FIG. 2 is structured, the document 20 can be divided into Chapter 1 and Chapter 2, and the first chapter can be further divided into Section 1, etc. Section 1 can be further divided into, for example, phrases. When the document 20 is structured and the contents are divided into one or more sections, the document 20 can be divided into Section 201, which is the entire document 20, Section 202, which is the first chapter, Section 203, which is the first section, Section 204, which is a phrase, Section 205, etc., which is the second chapter. Furthermore, meta information 206 to 210 is assigned to each section.

[0046] Meta information may be attached to each of one or more sections. As an example, meta information may include at least one of the following information: time information on when the document or the text in the document was written based on legal or practical information at that time, bibliographic information and source information such as the title, author, publisher, and date of writing, information on the user's viewing and use of the document or the text in the document, evaluation from the user by a feedback function described later, location information of the section in the document, etc. The location information of the section in the document includes the document's heading title, chapter information, page number, link information that allows direct transition to the response text, etc., which will be described later. Here, meta information may differ for each section even within the same document. For example, time information may differ for each section.

[0047] Current laws and regulations themselves are clearly legally true and legally accurate. Similarly, public information that is up to date at the present time is legally accurate. Public information that is not up to date at the present time, such as "explanatory books, articles, and papers on public information" and practical information, is not necessarily legally true and may therefore be legally inaccurate.

[0048] The initial sentence acquisition unit 111 acquires an initial sentence via the I / O 13 or the input unit 15 , and transmits it to the sentence analysis and extraction unit 112 .

[0049] The text analysis extraction unit 112 extracts, from among a plurality of documents registered in advance, a description related to the content of the acquired initial text and a document from which the description was extracted. When extracting the description and the document, the text analysis extraction unit 112 may also extract meta-information related to the description and the document. When extracting the description and the document, the text analysis extraction unit 112 may extract only documents that satisfy a predetermined criterion. The predetermined criterion is a criterion of legal accuracy, which is a guide to the extent to which the content described in the document to be extracted conforms to the law, and may be classified, for example, by a numerical value indicating accuracy. The legal accuracy of the document can be recursively obtained from the meta-information of the document, the legal accuracy of the document cited in the document, and the legal accuracy of the sections that constitute the document, using a calculation unit described later. The legal accuracy of the response text is recursively determined from the legal accuracy of the document cited.

[0050] The response sentence generating unit 113 generates a response sentence based on the initial sentence, the description, and the document, which is an answer conforming to the law and rules within an organization based on the law, by quoting the description and a part of the description. The response sentence generating unit 113 may generate the response sentence by using a generation AI such as a deep learning AI model capable of generating sentences based on trained data. The response sentence generating unit 113 may generate the response sentence by using, for example, a so-called large language model (LLM), and may generate the response sentence by filling in the initial sentence, the description, and the document in a predetermined format for requesting the generation of a response sentence and inputting the format into the large language model.

[0051] The response sentence generating unit 113 may also generate a response sentence by adding meta information to the description or a part of the description and quoting the description. When generating a response sentence by quoting a description or a part of the description, the response sentence generating unit 113 may generate a response sentence after weighting the description or the part of the description in terms of importance, priority, etc.

[0052] The display unit 14 simultaneously displays the initial sentence, the response sentence, the extracted description, and the document from which the extracted description was extracted. The display unit 14 may further display meta-information related to the description and the document.

[0053] A specific example of this embodiment will be described below with reference to Fig. 3. Fig. 3 shows an example of an initial sentence 31, a response sentence 32, a description 33, and a document 34 displayed by the display unit 14.

[0054] As an example, the initial text 31 is a question about law, and is shown in Fig. 3 as "Should a 'special controlling shareholder' be a company?". The initial text 31 is assumed to have been transmitted to the CPU 11 via the input unit 15.

[0055] The text analysis and extraction unit 112 extracts, from among a plurality of documents registered in advance, a description 33 relating to the content of the acquired initial text 31 and a document 34 from which the description relating to the content was extracted. In FIG. 3, the document from which the description relating to the content was extracted is, for example, a commentary on laws and regulations, and is called "X Law Firm Series Amended Company Law Q&A Co., Ltd." The description 33 relating to the recognized content of the acquired initial text 31 is written in "X Law Firm Series Amended Company Law Q&A Co., Ltd.", and in FIG. 3, the contents are shown with highlights at the bottom of the document 34 from which the description relating to the content was extracted. The description 33 is a direct copy of a portion of the relevant section written in "X Law Firm Series Amended Company Law Q&A Co., Ltd."

[0056] The response sentence generating unit 113 generates a response sentence 32 based on the initial sentence 31 and the description 33 by quoting the description and a part of the description.

[0057] The display unit 14 displays the initial sentence 31, the response sentence 32, the extracted description 33, and the document 34 from which the extracted description was extracted on the same screen. In Fig. 3, the initial sentence 31 is arranged at the top, the response sentence 32 is arranged below the initial sentence 31 on the left side of the screen, and the extracted description 33 and the document 34 from which the extracted description was extracted are arranged below the initial sentence 31 on the right side of the screen.

[0058] When a user inputs an initial sentence 31 from the input unit 15, the sentence analysis and extraction unit 112 extracts a description 33 related to the content of the initial sentence 31 and a document 34, and then the response sentence generation unit 113 generates a response sentence 32. As shown in Fig. 3, when the initial sentence 31, the response sentence 32, the extracted description 33, and the document 34 are displayed by the display unit 14, the user can refer to the extracted description 33 and the document 34 from which the extracted description 33 was extracted, which are displayed on the same screen, when checking the response sentence 32.

[0059] 3, the correspondence between the recognized content of the initial sentence 31 and the corresponding description 33 is clearly displayed on the screen by highlighting, underlining, linking, etc. In this way, the user can easily confirm the legal content on which the response sentence 32 is generated.

[0060] Furthermore, the response text 32 is a text generated by quoting the description 33, and may be, for example, a summary of the description 33.

[0061] When generating a response sentence 32 for an initial sentence 31 in accordance with legal or practical information, if one wishes to determine whether the initial sentence 31 and the response sentence 32 are legally correct and to gain a deeper legal understanding, it is necessary to check the response sentence 32 and the description 33 and document 34 that serve as the basis for generating the response sentence 32 from time to time or at the same time. In general, when using AI, including generation AI, the description and document literature that serve as the basis for generation are not displayed together with the initial sentence and the response sentence, or if they are displayed, the document name and the location of the information are displayed to the extent that they are displayed. Therefore, even if the document name and the link to the document are displayed on the same screen as the initial sentence and the response sentence, in order to check the document name and its content one by one, which is thought to contain a large amount of information, it is necessary to transition to a different screen.

[0062] In the text analysis and generation device 10 according to this embodiment, in order to promote efficient, comprehensive, and thorough legal understanding of the initial text 31 and the response text 32, the description 33, the document 34, and the meta-information added to these descriptions and documents are arranged on the same screen so that they can be easily confirmed.

[0063] In Fig. 3, as an example, the response text 32, the description 33, and the document 34 may be arranged side by side so that they are on the left and right side of each other. In Fig. 3, the entry field 36 of the description 33 and the document 34 may be wider than the entry field 35 of the response text 32, so that more information about the document 34 can be shown.

[0064] As an example, document 34 may be displayed with a book image 38 and meta information such as the book title, author, and publication date to the right of the image. The chapter headings and page numbers of description 33 may be displayed in selectable colors as headings, and a part of description 33 may be displayed below the chapters. When the mouse cursor is placed over a part of description 33, the whole of description 33 is displayed, and further, the whole of description 33 or document 34 can be read using scroll bar 37, so that description 33 and document 34 can be viewed with fewer steps.

[0065] For example, when a user places the mouse cursor over a desired portion of response sentence 32, the details of description 33 and document 34 that serve as the basis for generating that portion can be easily found by highlighting description 33 and document 34, or by displaying description 33 and document 34 in an emphasized manner, such that the user can easily find details of description 33 and document 34 that serve as the basis for generating the desired portion of response sentence 32, by scrolling the corresponding portions of description 33 and document 34 to a display position on the screen and highlighting them, or by displaying them at the top of description 33 and document 34.

[0066] Furthermore, for example, if one wishes to check desired portions of statement 33 and document 34 in more detail, one can click on the portion to display a screen for viewing document 34, with the portion of the original text corresponding to statement 33 highlighted on the screen, and the entire original text can be read by scrolling.

[0067] Fig. 4 shows an example of a case where the corresponding description is enlarged when the mouse cursor is placed over a desired portion of the response document shown in Fig. 3. As shown in Fig. 4, when the mouse pointer 41 is moved to a desired portion of the response document, the description area of ​​the corresponding description 33 is enlarged, and while only a portion of the description was displayed in Fig. 3, the scroll bar 37 is not displayed in Fig. 4, and the entire description is displayed. This operation of the display unit 14 makes it possible to easily check the description that is the basis for generating the response document in a short time.

[0068] 4 shows an example in which the area of ​​the corresponding description is expanded by the mouse pointer and the entire description is displayed, but in this embodiment, as described above, other configurations may be adopted to easily check the description in a short time, such as highlighting the corresponding description by the mouse pointer. Also, the user may be able to select a configuration to easily check the description in a short time.

[0069] In the text analysis generation device 10 according to this embodiment, the text analysis extraction unit 112 extracts the description 33 and the document 34, and then the response text generation unit 113 generates the response text 32. In Fig. 3, the description 33 and the document 34 may be displayed at the time when they are extracted, and then the response text 32 may be displayed.

[0070] The text analysis and generation device 10 according to the present embodiment may be capable of acquiring a plurality of texts as the initial text 31. For example, a question text may be input as the initial text 31 in the entry field 36, and after the description 33, the document 34, and the response text 32 are extracted and generated, an additional text may be input as the initial text 31. This may be configured to allow more detailed information to be obtained. When inputting the additional text, for example, the additional text may be input in the entry field 36 following the initial text 31 input in the entry field 36. Alternatively, a text entry field 35 for inputting the additional text may be provided independently of the entry field 36. As shown in FIG. 3, for example, the entry field 35 may be disposed below the response text 32 and the description 33, and the description, document, and response text extracted and generated by inputting the additional text may be displayed below the entry field 35. By providing a field 35 for inputting additional sentences, the additional sentences can be input while checking the description 33, document 34, and response sentence 32 in response to the initial sentence 31. Furthermore, the description, document, and response sentence extracted and generated by inputting the additional sentences can be checked together with the description 33, document 34, and response sentence 32 and compared, examined, etc.

[0071] When there are multiple recognized contents of the initial sentence 31 and the contents are different from each other, there may be multiple descriptions 33 related to the recognized contents of the initial sentence 31. For example, this includes cases where the contents of the initial sentence 31 are in accordance with multiple articles. In this case, the sentence analysis and extraction unit 112 extracts one or more descriptions 33 related to each of the multiple contents of the initial sentence 31 and one or more documents 34 from which one or more descriptions related to each of the multiple contents are extracted, and then the response sentence generation unit 113 generates the response sentence 32. In this case, the display unit 14 displays the initial sentence 31, the response sentence 32, the one or more descriptions 33, and the one or more documents 34 on the same screen.

[0072] Even if the recognized content of the initial sentence 31 is one, there may be a plurality of patterns of the description 33 and the document 34 that can be extracted by the sentence analysis extraction unit 112. In this case, the display unit 14 may display the description 33 and the document 34 that are close to the recognized content of the initial sentence 31 in addition to the initial sentence 31 and the response sentence 32. Alternatively, the display unit 14 may display all of the descriptions 33 and the documents 34 extracted by the sentence analysis extraction unit 112 in order of closeness, for example, in addition to the initial sentence 31 and the response sentence 32. Alternatively, the display unit 14 may display the descriptions 33 and the documents 34 extracted by the sentence analysis extraction unit 112 in addition to the initial sentence 31 and the response sentence 32, for example, by quantifying the closeness of each of the descriptions 33 and the documents 34, and display the descriptions 33 and the documents 34 whose quantified closeness to the truth exceeds a predetermined value.

[0073] The writing analysis generation device 10 according to the present embodiment may be configured so that when the writing analysis extraction unit 112 extracts a description and a document from a plurality of documents registered in advance, the user can select a document to be used for extraction from the plurality of documents registered in advance. For example, the writing analysis generation device 10 may be configured so that the user can restrict the document based on meta-information or the like associated with the document, such as excluding a specific document by specifying the title, excluding all old documents that have been published or made public for a long time, prioritizing authoritative documents, specifying a legal field, extracting only precedents, or excluding a specific type of document from the extraction target. In addition, as described later in the second embodiment, a case where practical information is added to a plurality of documents registered in advance is described, but the writing analysis generation device 10 may be configured so that the user can select whether or not to add practical information to the document to be used for extraction.

[0074] The text analysis and generation device 10 according to this embodiment may be configured so that the user can select the generation AI to be used when the response text generation unit 113 generates the response text 32 from among a plurality of generation AIs.

[0075] The text analysis and generation device 10 according to this embodiment may be provided with a feedback function for evaluating the response text 32, the description 33, and the document 34, providing feedback, and updating the meta-information stored in the document storage unit 121.

[0076] When the user inputs the initial sentence 31 in the input unit 15, for example, the user inputs the initial sentence 31 while predicting the response sentence 32 to be generated to some extent, but when the generated response sentence 32 is significantly different from what the user predicted, the user may wish to generate the response sentence 32 again. In this case, the user may wish to generate the response sentence 32 again based on an initial sentence having the same content as the initial sentence 31 already input, or may wish to generate the response sentence 32 again based on an initial sentence having a different content from the initial sentence 31 already input. In response to such a user request, the sentence analysis generation device 10 according to the present embodiment can repeatedly generate the response sentence 32 according to the user's input while retaining information on the initial sentence 31 to be acquired and the response sentence 32 to be generated. Details of this will be described below.

[0077] When the response sentence generation unit 113 generates the response sentence 32 after the initial sentence acquisition unit 111 acquires the initial sentence 31, even if the initial sentence 31 is the same, the content of the response sentence 32 may be different each time the response sentence 32 is generated. Therefore, after the initial sentence acquisition unit 111 acquires the initial sentence 31 and the response sentence generation unit 113 generates the response sentence 32, the response sentence generation unit 113 may generate the response sentence 32 again as a second response sentence while retaining information about the first sentence and the first response sentence in response to an input by the user requesting to generate the response sentence 32 again based on the first sentence. The display unit 14 may display the first response sentence and the second response sentence generated by the response sentence generation unit 113.

[0078] The above-mentioned operation of generating the response sentence 32 again based on the first sentence by the response sentence generating unit 113 may be repeated as many times as the user desires. In addition, the display unit 14 may display all of the response sentences 32 generated by the operation of generating the response sentence 32 again based on the first sentence by the response sentence generating unit 113, or a part of the response sentences 32 generated by the operation of generating the response sentence 32 again by the response sentence generating unit 113 selected by the user. That is, after the initial sentence acquisition unit 111 acquires the initial sentence 31 as the first sentence and the response sentence generating unit 113 generates the response sentence 32 as the first response sentence, when the input unit 15 receives an input from the user requesting to generate the response sentence 32 again based on the first sentence, the response sentence generating unit 113 may generate the response sentence 32 again as the second response sentence while the initial sentence acquisition unit 111 holds information on the first sentence and the first response sentence. After the response sentence generation unit 113 generates the second response sentence, when the input unit 15 receives an input from the user requesting to generate a response sentence 32 again based on the first sentence n times, the response sentence generation unit 113 may generate the response sentence 32 again as the (n+1)th response sentence while the initial sentence acquisition unit 111 holds the information on the first sentence and the first response sentence, the second response sentence, ..., the nth response sentence. Here, n is an integer of 1 or more. After the response sentence generation unit 113 generates the response sentence 32 again as the (n+1)th response sentence while the initial sentence acquisition unit 111 holds the information on the initial sentence 31 and the first response sentence, the second response sentence, ..., the nth response sentence, the display unit 14 may display all of the first response sentence, the second response sentence, ..., the nth response sentence, or, for example, a part selected by the user from the first response sentence, the second response sentence, ..., the nth response sentence.

[0079] In addition, after the initial sentence acquisition unit 111 acquires the initial sentence 31 as the first sentence and the response sentence generation unit 113 generates the response sentence 32 as the first response sentence, in response to a user input requesting that response sentence 32 be generated again based on a second sentence different from the first sentence, and input of the second sentence as the initial sentence 31, the initial sentence acquisition unit 111 may retain information regarding the first sentence and the first response sentence, and the sentence analysis extraction unit 112 may first extract, from a plurality of documents registered in advance, a description 33 relating to the content of the second sentence as the second description, and a document 34 from which the second description was extracted as the second document, and then the response sentence generation unit 113 may generate the response sentence 32 as the second response sentence. When the sentence analysis and extraction unit 112 selects from a plurality of documents registered in advance, the description 33 relating to the content of the first sentence as the first description, and the document 34 from which the first description was extracted as the first document, the display unit 14 may display the first sentence and the second sentence acquired by the initial sentence acquisition unit 111, the first description, the first document, the second description and the second document extracted by the sentence analysis and extraction unit 112, and the first response sentence and the second response sentence generated by the response sentence generation unit 113.

[0080] As in the case of the operation of generating the response sentence 32 again based on the first sentence by the response sentence generating unit 113, the operation of generating another response sentence 32 by the response sentence generating unit 113 in response to the input requesting generation of another response sentence 32 based on the initial sentence 31 different from the first sentence and the input of the second sentence as the initial sentence 31 and the extraction of the description related to the content of the initial sentence 31 different from the first sentence and the document from which the description was extracted by the text analysis extraction unit 112 may be repeated as many times as desired by the user. In addition, the display unit 14 may display all of the response sentence 32 generated by the operation of generating another response sentence 32 by the response sentence generating unit 113, the description extracted by the operation of extracting the description and the document from which the description was extracted by the text analysis extraction unit 112, and the document from which the description was extracted, or a part selected by the user. That is, the initial sentence acquisition unit 111 acquires the initial sentence 31 as the first sentence, the sentence analysis extraction unit 112 extracts, from a plurality of documents registered in advance, the description 33 related to the content of the first sentence as the first description, and the document 34 from which the first description was extracted as the first document, and the response sentence generation unit 113 generates the response sentence 32 as the first response sentence. After that, when the input unit 15 receives an input from the user requesting that response sentence 32 be generated again based on a second sentence different from the first sentence, and an input of the second sentence as the initial sentence 31, while the initial sentence acquisition unit 111 retains information regarding the first sentence and the first response sentence, the sentence analysis extraction unit 112 first extracts, from a plurality of documents registered in advance, the description 33 related to the content of the second sentence as the second description, and the document 34 from which the second description was extracted as the second document, and then the response sentence generation unit 113 generates another response sentence 32 as the second response sentence.After the response sentence generating unit 113 generates the second response sentence, when the input unit 15 receives an input by the user requesting the generation of yet another response sentence 32 and the input of the initial sentence 31 repeatedly n times, the sentence analysis extraction unit 112 may extract the description 33 related to the content of the (n+1) sentence as the (n+1) description and the document 34 from which the (n+1) description was extracted as the (n+1) document from among the multiple documents registered in advance, while retaining the information on the first sentence, the second sentence, ..., the nth sentence acquired by the initial sentence acquisition unit 111, and the information on the first description, the second description, ..., the nth description, the first document, the second document, ..., the nth document, the first response sentence, the second response sentence, ..., and the nth response sentence, and the document analysis extraction unit 112 may extract the description 33 related to the content of the (n+1) sentence as the (n+1) description and the document 34 from which the (n+1) description was extracted as the (n+1) document, and the response sentence generating unit 113 may generate yet another response sentence 32 as the (n+1) response sentence. Here, n is an integer of 1 or more. The display unit 14 may display all of the first entry, second entry, ..., nth entry, first document, second document, ..., nth document, first response sentence, second response sentence, ..., nth response sentence, or a portion selected by the user.

[0081] The operation by the response sentence generation unit 113 of generating a response sentence 32 again based on the first sentence and the operation by the response sentence generation unit 113 of generating another response sentence 32 based on an initial sentence 31 different from the first sentence may be repeated by repeating only the operation by the response sentence generation unit 113 of generating a response sentence 32 again based on the first sentence, or may be repeated by repeating only the operation by the response sentence generation unit 113 of generating another response sentence 32 based on an initial sentence 31 different from the first sentence, or the operation by the response sentence generation unit 113 of generating a response sentence 32 again based on the first sentence and the operation of generating another response sentence 32 based on an initial sentence 31 different from the first sentence are selected by the user each time, and when the input unit 15 receives information regarding the operation selected by the user, the operation selected by the user is executed each time, and the operation may be repeated as many times as desired by the user. In addition, the display unit 14 may display the description 33 extracted by the sentence analysis extraction unit 112, the document 34 from which the description 33 was extracted, and all of the response sentence 32 generated by the response sentence generation unit 113, or a portion selected by the user, by having the response sentence generation unit 113 select each time between an operation of generating a response sentence 32 again based on the first sentence and an operation of generating another response sentence 32 based on an initial sentence 31 different from the first sentence, and the selected operation is executed each time, and the operation is repeated as many times as the user desires.

[0082] In the above-mentioned writing analysis generation device 10, the writing analysis extraction unit 112 extracts the description 33 and the document 34 from a plurality of documents registered in advance. The input unit 15 can also accept input of an input document from a user. In this case, the writing analysis extraction unit 112 may extract the description 33 and the document 34 from a plurality of documents registered in advance and an input document accepted by the input unit 15. The input document accepted by the input unit 15 is temporarily used to extract the description 33 and the document 34.

[0083] The text analysis and generation method according to the embodiment of the present invention will now be described with reference to the flow chart shown in FIG.

[0084] In step S501, an initial sentence is acquired via the I / O 13 or the input unit 15.

[0085] In step S502, the text analysis and extraction unit extracts descriptions related to the contents described in the initial text from a plurality of documents registered in advance.

[0086] In step S503, the response sentence generation unit generates a response sentence based on the initial sentence and the extracted description, in accordance with legal or practical information, by quoting the description or a part of the description.

[0087] In step S504, the display unit displays the initial sentence, the response sentence, the description, and the document on the same screen.

[0088] The accuracy of answers provided by the text analysis and generation device according to the embodiment described above was examined by estimating the rate of correct answers using questions from a certain area of ​​the multiple-choice test in the bar exam as initial sentences and the response sentences generated by the text analysis and generation device as answers to the multiple-choice test.

[0089] As a result, in the civil subjects and corporate law areas of the short-answer test of the 2014 Bar Examination, the correct answer rate of the answers by the text analysis and generation device according to this embodiment was 78.6%, while the correct answer rate of the answers by a general generation AI (for example, based on GPT-4) was 35.7%. In addition, for a total of 70 questions in the corporate law area posed in the civil subjects of the short-answer test of the Bar Examination from 2012 to 2014 and the civil law, commercial law, and civil procedure law subjects of the short-answer test of the preliminary bar examination from 2012 to 2016, the correct answer rate of the answers by the text analysis and generation device according to this embodiment was 71.4%, while the correct answer rate of the answers by a general generation AI was 42.9%. The pass rate of the short-answer test of the Bar Examination is usually about 60%, and the correct answer rate of the answers by the text analysis and generation device according to this embodiment was significantly higher than the pass rate of the general generation AI and the pass rate of the short-answer test of the Bar Examination. This shows that the text analysis and generation device according to this embodiment provides a highly accurate response.

[0090] Below, we will describe an example of answer comparison between the text analysis and generation device according to this embodiment and a general generation AI. As the initial sentence, we used the following sentence, which is from the 2014 Bar Examination, Multiple Choice Exam, Civil Subject, Question 43: "Question: (Omitted) Regarding the resolution of the board of directors regarding the dismissal of a representative director, the dismissal takes effect only when the resolution is notified to the representative director." The answer to this question is "False."

[0091] When the above problem by the text analysis and generation device according to the present embodiment is used as the initial text, the answer is "wrong", that is, the correct answer is also generated, and the response text generated is, as shown in FIG. 4, "According to legal precedents, the effect of dismissal of a representative director is automatically generated by a resolution of the board of directors, and notice to the dismissed representative director is not required. (Supreme Court Decision, December 20, 1966, Minshu Vol. 20, No. 10, p. 2160) [1]." In addition, as the document from which the extracted description was extracted, "References: [1] Commentary on the Companies Act, Vol. 8, Institutions (2) (Articles 348-395) (Part 2, Joint Stock Companies, Chapter 4 Institutions, Section 5, Board of Directors, Article 1, Powers, etc., Article 362 (Powers, etc. of Directors), III, Duties of the Board of Directors, 4, Selection of Representative Director, Dismissal, (3) Dismissal" is indicated.

[0092] On the other hand, the answer to the above question by a general generation AI was "correct", i.e., an incorrect answer, along with the following explanation: "If a board of directors resolution is made to dismiss a representative director, the dismissal takes effect at the time the resolution is notified to the representative director. However, this may differ depending on the actual procedures and specific circumstances, so it is advisable to consult an expert if necessary."

[0093] Thus, erroneous information generated by a general large-scale language model may appear very plausible at first glance, and it may be difficult for users who do not have sufficient knowledge of law to determine whether the information is correct or incorrect. Therefore, in the legal field, where accuracy is required, it is important to have the large-scale language model provide an answer based on documents that are legal information and citations of relevant parts of the documents. It is also important to encourage users to obtain an efficient, comprehensive, and complete legal understanding by arranging the answer (response) by the generating AI and the descriptions or documents that serve as the basis for the answer (response) on the same screen so that the user can easily check them at the same time. In the present invention, the present invention provides maximum ease of checking the answers (responses) by the generating AI and the descriptions, documents, and associations on which the answers (responses) are based throughout the invention, thereby promoting the use of AI, including generating AI, in the legal field, where accuracy is particularly important.

[0094] Second embodiment As in the first embodiment, the text analysis and generation device according to this embodiment can extract descriptions related to the recognized contents from legal information as multiple documents registered in advance, analyze the recognized contents in accordance with legal or practical information based on the extracted descriptions, and generate a response text showing the analysis result by quoting the recognized contents. Legal information refers to public information such as "treaties, constitutions and laws and regulations," "judgments and other court documents," "official gazettes, guidelines, public comments, minutes of councils, etc., notices, notifications, circulars, and other government documents," "decisions and rulings, etc.," "legal information related to laws and regulations such as the minutes of the Diet and parliamentary documents," and "local government ordinances, rules, and parliamentary documents," as well as "explanatory books, explanatory articles, explanatory papers, etc. related to the above-mentioned public information," and "statistical information, etc. related to the above-mentioned public information."

[0095] In this embodiment, practical information may be further included in the multiple documents registered in advance. In this embodiment, the initial text is, as an example, a business instruction related to law, and the response text is a legal document necessary for executing the business instruction related to law. The practical information may include, for example, a legal document necessary for executing the business instruction related to law that was created in the past within a specific organization, a practical document describing the content of legal practical processing within the organization that is created based on legal information, a contract document, organizational information, organizational regulations and guidelines, preparatory documents, opinions, records, email and chat information with inside and outside the organization, and disclosure materials. When the text analysis and generation device according to this embodiment generates a legal document necessary for executing a business instruction related to a desired law as an initial text and a response text as a legal document necessary for executing a business instruction related to a desired law, the text analysis and generation device can generate a response text that takes into account practical information within a specific organization by using the practical information as a cited document in addition to the legal information.

[0096] As described in the first embodiment, a plurality of documents registered in advance may be classified by a numerical value representing the legal accuracy of each document or each sentence in a document. For example, the legal accuracy of practical information in a particular organization is lower than that of current laws and regulations. Therefore, a response document generated by taking into account practical information in a particular organization may contain more elements that are not legally accurate than a response document generated by taking into account a document with higher legal accuracy such as current laws and regulations. When generating a response sentence, the text analysis and generation device according to this embodiment may generate both a response document based only on a document or sentence with a higher numerical value representing legal accuracy that does not include a document with a numerical value representing legal accuracy lower than a predetermined value, such as practical information, and a response sentence based on one or more documents including a sentence with a numerical value representing legal accuracy lower than a predetermined value. The display unit 14 may display the difference between the response sentence generated based only on documents with a higher legal accuracy than a predetermined value and the response sentence generated based on one or more documents including a sentence with a numerical value representing legal accuracy lower than a predetermined value.

[0097] For example, the user can modify the response sentence to increase its legal accuracy based on the difference between a response sentence based only on documents whose legal accuracy values ​​displayed on the display unit 14 are higher than a predetermined value and a response sentence generated based on documents including documents whose legal accuracy values ​​are lower than a predetermined value. Furthermore, the text analysis and generation device according to this embodiment can provide feedback on the user's modifications to the response sentence by reflecting the content of the modifications to the response sentence made by the user in the practical information stored in the document storage unit 121.

[0098] For example, the user can narrow down the documents on which the response sentence is based to those that have a certain level of legal accuracy.

[0099] As shown in FIG. 6, the text analysis and generation device according to the present embodiment may include a calculation unit 61 that calculates a numerical value representing legal accuracy and a numerical value representing practical accuracy at a certain point in time for each document or each sentence in a document. For example, the numerical value representing accuracy may be a real value of [0,1], and the higher the accuracy, the closer it is to 1, and the lower the accuracy, the closer it is to 0. The "legal accuracy" is the accuracy determined only based on the "legal information" at a certain point in time. That is, for example, the sentence "employees may have a side job" at the present time can be said to have an accuracy close to 1 based on the freedom of occupation choice under Article 22 of the current Constitution, etc., and based on the current "legal information". However, if Article 22 of the Constitution is amended in the future, the accuracy at a later point in time will be close to 0. The "accuracy in practical information" is the accuracy of the sentence when at least a part of the "practical information" is taken into account. For example, when the "practical information" includes the company's current work rules (which include a provision prohibiting side jobs), the "accuracy of practical information" for the current sentence "Employees are allowed to have side jobs" will be close to 0.

[0100] The text analysis and generation device 60 according to the present embodiment may further include a correction unit 62 that corrects the generated response text so as to improve a numerical value representing the accuracy of legal or practical information. For example, when the numerical value representing the accuracy of the generated response text is determined to be less than a predetermined standard, the correction unit 62 may correct the response text whose numerical value representing the accuracy is determined to be less than a predetermined standard, based on a difference from a response text generated based only on a document having a higher numerical value representing the accuracy. In addition, when correcting the response text, the correction unit 62 may correct the response text based on an instruction in a natural language received from the user by the input unit 15. Furthermore, a document that the response text may refer to may be specified by the user by specifying a numerical value representing the accuracy or meta-information that is the source of the numerical value.

[0101] Furthermore, a feedback unit 63 may be provided that feeds back information including the content corrected by the correction unit 62 to the practical information stored in the document storage unit 121. Hereinafter, the numerical value representing the legal accuracy calculated by the calculation unit is referred to as a score, and the calculation of the score is referred to as scoring.

[0102] The calculation unit 61 may perform scoring for each section using meta information assigned to each of one or more sections when the document is structured. Details of the scoring will be described with reference to FIG. 7. FIG. 7(a) shows a part of the provisions of the XX Law, which is a law enacted on April 1, 2020, and FIG. 7(b) shows a part of the provisions of the current XX Law after Article 100 of the XX Law shown in FIG. 7(a) was amended on August 1, 2023. The XX Law shown in FIG. 7(b) is amended only in Article 100 of the XX Law between April 1, 2020 and August 1, 2023. FIG. 7(c) shows a part of the commentary of the XX Law shown in FIG. 7(a) and FIG. 7(b). The commentary shown in FIG. 7(c) was published on January 1, 2023. Assume that the documents shown in Figures 7(a), 7(b), and 7(c) are structured, and portions of these documents are divided into sections 701a to 703a, 701b to 703b, 701c, and 702c, which are surrounded by dashed lines.

[0103] The time information included in the meta information for section 702a is "Article 100 of the XXX Act, effective April 1, 2020." On the other hand, the time information included in the meta information for section 702b is "Article 100 of the XXX Act, effective August 1, 2023." When comparing the legal accuracy of sections 702a and 702b as of August 1, 2023 and after, section 702b is "legally accurate" since it is the current law after the legal reform. Section 702a is a law from before the legal reform and is not legally accurate.

[0104] On the other hand, the time information included in the meta information for section 703a is "Article 101 of the XX Act, effective April 1, 2020," and the time information included in the meta information for section 703b is "Article 101 of the XX Act, effective August 1, 2023." As of August 1, 2023 and after, section 703b is "legally accurate," but between April 1, 2020 and August 1, 2023, Article 101 of the XX Act has not been amended, and therefore section 703a can also be said to be "legally accurate."

[0105] Furthermore, although Section 701c explains Article 100 of the XX Act before the revision of the law, as of August 1, 2023, Section 701c is not legally accurate because Section 702a, which it cites, is not legally accurate. Section 702c is not legally accurate because Section 703a, which it cites, is legally accurate.

[0106] In this way, the legal accuracy of a section is calculated recursively from the meta information of the section and the legal accuracy of other sections cited by the section. For example, if the legal accuracy of each item constituting the meta information of each section is quantified, the calculation unit can score the legal accuracy of each section as a specific numerical value from the legal accuracy of each item constituting the meta information.

[0107] In the above explanation, only time information was considered as meta information for each section shown in Figure 6, but the legal accuracy of each section may be influenced by other meta information, such as the author, user ratings, etc. In addition, in the above explanation, a point in time after August 1, 2023 was implicitly assumed as the point in time for scoring, but the legal accuracy of each section may be influenced by setting this point in time to, for example, before August 1, 2023.

[0108] The user may use the calculation unit 61 included in the text analysis generation device 60 according to this embodiment to score one or more sections constituting a document, and set a desired score value when the text analysis extraction unit 112 extracts descriptions, documents, and meta-information, and may set the unit to extract descriptions, documents, and meta-information from documents with legal accuracy higher than this score value. Alternatively, the unit may be configured to extract descriptions, documents, and meta-information regardless of the score value, and then display an alert (described later) on the display unit 14 for descriptions or documents with a score value lower than the desired value.

[0109] Furthermore, the user may set desired meta information when the text analysis and extraction unit 112 extracts descriptions, documents, and meta information, and set the text analysis and extraction unit 112 to extract descriptions, documents, and meta information from one or more sections that satisfy the desired meta information. The desired meta information is, for example, a desired author, a desired time point, and the like.

[0110] For example, when a user considers a trial regarding an incident that occurred at a certain point in the past, legally accurate information is required based on legal information at that point in the past. In this case, the user can set a past point in time as the desired meta information, and the text analysis extraction unit 112 can extract descriptions, documents, and meta information.

[0111] An alert may be displayed on the display unit 14 to inform the user that the score values ​​of the extracted descriptions, documents, and meta information are below a given numerical value. The alert may be displayed, for example, by a "!" or highlighting. For example, the content of the alert may be displayed on the screen by clicking the alert, moving the pointer over the alert with the mouse, or the like.

[0112] When the text analysis and extraction unit 112 extracts the description, document, and meta information from the pre-registered document stored in the document storage unit 121, the pre-registered document may have descriptions or sections with substantially different contents even if they have the same degree of legal accuracy, so there may be multiple candidates for the description, document, and meta information extracted by the text analysis and extraction unit 112. This case corresponds to a case where there are multiple interpretations of a certain law and the legal accuracy of the multiple interpretations is the same. In this case, for example, an alert may be displayed on the display unit 14 to inform the user that there are multiple candidates for the description, document, and meta information extracted by the text analysis and extraction unit 112. Alternatively, when the response text generation unit generates a response text, the response text may be generated using, for example, the description, document, and meta information that appear most frequently in the pre-registered document from among the multiple candidates for the description, document, and meta information. At that time, for example, the number of candidates of descriptions, documents, and meta information that correspond to each of three patterns, namely, descriptions with the same meaning as the generated response sentence, descriptions that are contradictory to the generated response sentence, and no related descriptions, among the multiple candidates of descriptions, documents, and meta information, may be displayed on the display unit 14. For example, detailed contents of the multiple candidates of descriptions, documents, and meta information may be displayed on the screen by clicking on the alert or the display of the number of candidates, moving the pointer with the mouse, etc.

[0113] The text analysis and generation device of this embodiment may be configured so that the user can select to what extent the correction unit will take into account the numbers representing legal accuracy and the numbers representing practical information accuracy.

[0114] (Third embodiment) In this embodiment, the initial sentence is not limited in the number of characters, the number of phrases, etc., and may be any form of sentence. For example, the initial sentence may be a written document. According to the writing analysis generation device of this embodiment, for example, a written document is acquired as an initial sentence, and a response sentence is generated. The response sentence generated at this time may be, for example, a document that legally explains the document. Or, for example, if the document does not comply with the rules within the organization based on the law indicated by a part or all of the law or practical information, a document that is modified to comply with the rules within the organization based on the law indicated by a part or all of the law or practical information may be generated. The writing analysis generation device according to this embodiment may be configured so that, when acquiring a written document as an initial sentence and generating a response sentence, the user can select the type of response sentence to be generated, such as a document that legally explains the document, a document that is modified to comply with the law, etc., as described above.

[0115] The documents acquired as the initial documents may include, but are not limited to, documents that are required to comply with the law, such as business documents that describe the contents of legal business processes within an organization, contract documents, internal company information, internal company regulations, preparatory documents, opinions, records, email information, chat information, disclosure materials, etc., that are created based on legal information. For example, due to legal amendments, etc., it may be necessary to amend the above documents, and the text analysis and generation device according to this embodiment can be applied to such cases.

[0116] The initial sentence may be a conversation, a statement, or the like recorded in text form. For example, it may be a statement made in a meeting or the like that has been made into text, or voice data from a meeting or the like may be converted into text data by voice recognition. The text analysis and generation device according to this embodiment may further include a voice / text conversion unit that acquires voice data and converts it into text data to make the initial sentence.

[0117] The initial sentence may also include a non-verbal meaning such as an image or a drawing. For example, the initial sentence may be a diagram or a conceptual illustration in a presentation document, or may be text data representing the gist of the image or drawing. The sentence analysis and generation device according to the present embodiment may further include an image / text conversion unit that acquires image data and converts it into text data to generate the initial sentence.

[0118] Furthermore, the initial sentence may be the output content of any AI, including a generation AI. In the third embodiment, the sentence analysis and generation device may have a calculation unit similar to or similar to that in the second embodiment, calculate the legal accuracy of the output content of any AI used in practice, and reflect the calculation result on the display unit.

[0119] The text analysis and generation device may refer to preregistered documents, descriptions, and meta information that have a certain level of legal accuracy and practical accuracy, generate alternatives and amendments to improve the legal accuracy or practical accuracy of the initial document as a response text, and display and provide the document, description, and meta information that is the basis for the response text. The user may also request the generation of a desired alternative for the amendment of the response text at any time. When generating the amendment or desired alternative, a numerical standard for legal accuracy and additional documents, descriptions, and meta information to be referenced may be separately provided. In addition, the user may instruct the initial text to reflect the alternatives and amendments of the initial text, which is the response text, in the initial text. In this case, the initial text is amended in accordance with the alternatives and amendments as the response text, and may be stored in the document storage unit 121, or, if the source that output the initial text is a system including AI, may be fed back to the system.

[0120] In the third embodiment, the initial text to be obtained does not necessarily have to be actively given by the user. It can be any human or non-human output content generated in the course of a user's work. The text analysis generation device may be constantly operating, and automatically and at any time calculates the legal and practical accuracy of any content generated or received by the user, and if it is calculated that the legal and practical accuracy of such content is below a reference value, it may notify the user at any time by a visual or audio method such as an alert, and at the same time, it may display or suggest alternatives or corrections at any time. In addition, at that time, the description or document on which the alternative or correction is based may be displayed or suggested at the same time.

[0121] In all the embodiments described herein, the text analysis generation device 10 can accept additional input, whether passive or active, as a new initial text after the display unit 14 displays the initial text, description, document, response text, or meta information. In this case, the text analysis generation device can repeatedly respond by extracting description, document, and meta information from the description, document, and meta information displayed by the display unit 14 in addition to multiple documents registered in advance, separately provided input documents, and meta information, and generating response text.

[0122] Just as the initial sentence may be voice data, an image, etc., the response sentence may also be voice data, an image, etc.

[0123] (Fourth embodiment) In the first embodiment, the response sentence generation unit 113 generates the response sentence 32 based on the initial sentence 31 and the description 33 by quoting the description and a part of the description. In the present embodiment, when generating the response sentence 32, the response sentence generation unit 113 may further generate a file required for responding to the initial sentence 31, or may generate a URL (Uniform Resource Locator) that is a link destination for accessing a file required for responding to the initial sentence 31, or may specify a document stored in the document storage unit 121. The file required for responding to the initial sentence 31 may be, for example, a file required for executing a task instruction.

[0124] When the response sentence generation unit 113 determines that the content of the initial sentence 31 indicates that a file needs to be generated to respond, such as a business instruction, it may search for a file format required for the response from among the documents stored in the document storage unit 121. When the response sentence generation unit 113 finds the file format required for the response from among the documents stored in the document storage unit 121, it may extract the found format and generate a response sentence 32 based on the initial sentence 31 and the description 33 in accordance with the extracted format by quoting the description and a part of the description.

[0125] The file, URL, or specified document generated by the response sentence generation unit 113 may be displayed on the screen together with the response sentence 32 when the response sentence 32 is displayed on the screen by the display unit 14. For example, if the initial sentence 31 is a business instruction and the response sentence 32 is a document for executing the business instruction, a file required for responding to the initial sentence 31 or a document stored in the above-mentioned document storage unit 121 may be a file required for executing the business instruction.

[0126] FIG. 8 shows an example in which an initial sentence 31 is a business instruction in the writing analysis generation device 10 according to the present embodiment, and a file required for responding to the initial sentence 31 is generated by the response sentence generation unit 113. In the example shown in FIG. 8, the input contents by the user are written inside a balloon from the user's icon 801, and the response sentence generated by the response sentence generation unit 113 is written inside a balloon from the icon 802 of the writing analysis generation device 10. When the input unit 15 accepts the input of a first initial sentence 803 by the user, the response sentence generation unit 113 generates a response sentence 804 to the first initial sentence 803. When the input unit 15 accepts the input of a next initial sentence 805 by the user, the response sentence generation unit 113 generates a response sentence 806 to the next initial sentence 805. In the example shown in FIG. 8, the next initial sentence 805 is a business instruction with the content "put it into the form of a report while taking into account the format of past reports and upload it to a shared drive." The response sentence generation unit 113 determines that the content of the next initial sentence 805 is a business instruction, and generates a file required for responding to the initial sentence 805 and a link destination for accessing the file required for responding to the initial sentence 805. In the example shown in Fig. 8, the file required for responding to the initial sentence 805 is a "report compiled based on the format of past reports," and a link destination for accessing the file required for responding to the initial sentence 805 is displayed in the response sentence 806.

[0127] As described above, the present invention naturally includes various embodiments not described here. Therefore, the technical scope of the present invention is defined only by the invention-specifying matters related to the scope of the claims appropriate from the above description. [Explanation of symbols]

[0128] 10, 60 Sentence analysis and generation device 11 CPU 12 Storage section 13 I / O 14 Display section 15 Input section 111 Initial sentence acquisition part 112 Sentence analysis and extraction part 113 Response sentence generation unit 121 Document storage unit 20 Documents Sections 201-205, 701a-703a, 701b-703b, 701c, and 702c 206~210 Meta information 31, 803, 805 Initial text 32, 804, 806 Response text 33 Description 34 Documents 35, 36 Entry fields 37 Scroll Bar 38 Calligraphy 41 Mouse Pointer 61 Calculation section 62 Correction section 63 Feedback section 801, 802 Icon

Claims

1. an acquisition unit for acquiring an initial sentence; a text analysis and extraction unit for extracting text and descriptions related to the contents of the initial text from a plurality of documents registered in advance; a response sentence generation unit that generates an answer based on the initial sentence and the description, and generates a response sentence by quoting the description or a part of the description; a display unit that displays the initial sentence, the response sentence, the description, and the document on the same screen; A text analysis and generation device comprising:

2. The apparatus according to claim 1 , wherein the text analysis extraction unit extracts meta-information in addition to the description and the document.

3. The text analysis and generation device according to claim 2 , wherein the response text generation unit generates the response text based on the description or a part of the description, as well as the document and the meta-information.

4. 3. The text analysis and generation device according to claim 2, wherein the display unit displays the meta-information in addition to the initial text, the response text, the description, and the document on the same screen.

5. 2. The text analysis and generation device according to claim 1, wherein the plurality of pre-registered documents are documents that satisfy a predetermined standard.

6. The apparatus of claim 1 , wherein the plurality of pre-registered documents are classified according to a numerical value indicating accuracy.

7. The apparatus of claim 1 , wherein the plurality of pre-registered documents includes legal information.

8. The apparatus according to claim 1 , wherein the plurality of pre-registered documents further includes business information.

9. 2. The apparatus of claim 1, wherein the initial sentence is a modal question and the response sentence is a response to the initial sentence.

10. 2. The text analysis and generation device according to claim 1, wherein the initial text is a business instruction relating to a law, and the response text is a response to the business instruction.

11. The text analysis and generation device according to claim 1, characterized in that, when a mouse pointer is placed on a desired portion of the response text, the display unit highlights and displays the description that served as the basis for generating the desired portion.

12. The text analysis and generation device according to claim 1, characterized in that the display unit displays, in addition to the initial text, the response text, the description, and the document, a fill-in field for inputting additional text on the same screen.

13. The apparatus of claim 1 , wherein the document includes an image of the document's appearance.

14. The apparatus according to claim 1 , further comprising a calculation unit for calculating a value representing legal accuracy and a value representing practical accuracy for each of the documents or each of the sentences in the documents.

15. The apparatus for analyzing and generating text according to claim 1 , further comprising a correction unit for correcting the response text so as to improve a numerical value representing the accuracy of legal or practical information of the response text.

16. The apparatus according to claim 15, further comprising a feedback unit that feeds back information including the content corrected by the correction unit to the practical information.

17. The text analysis and generation device according to claim 1 , wherein the response text generation unit generates a response based on the initial text, the description, and the extracted document.

18. 2. The apparatus according to claim 1, wherein the answer complies with at least one of a law and an internal rule of an organization based on the law.

19. The text analysis and generation device of claim 1, characterized in that when generating the response text, the response text generation unit performs at least one of the following: generating a file related to the response text, generating a URL that is a link destination for accessing the file related to the response text, or specifying a document stored in a document memory unit.

20. Further comprising an input unit, The text analysis and generation device of claim 1, characterized in that after the acquisition unit acquires the initial text and the response text generation unit generates the response text as a first response text, the response text generation unit selects each time from among an operation of generating a response text again based on the initial text and an operation of generating another response text based on an initial text different from the initial text, and the operation accepted by the input unit is repeated the number of times desired by the user.

21. The text analysis generation device of claim 20, characterized in that the display unit is selected by the user each time and the operation received by the input unit is repeated the number of times desired by the user, thereby displaying the description extracted by the text analysis extraction unit, the document from which the description was extracted, and all of the response sentences generated by the response sentence generation unit, or a portion selected by the user.

22. an acquisition step of acquiring an initial sentence; a text analysis and extraction step of extracting descriptions, documents, and meta-information related to the contents of the initial text from a plurality of documents registered in advance; a step of generating an answer based on the initial sentence and the description, the step of generating a response sentence by quoting the description or a part of the description; a display step of displaying the initial sentence, the response sentence, the description, and the document on the same screen; A method for analyzing and generating text, comprising:

23. On the computer, An acquisition function for acquiring the initial sentence; A text analysis and extraction function for extracting descriptions, documents, and meta-information related to the contents of the initial text from a plurality of documents registered in advance; A function for generating an answer based on the initial sentence and the description, the function generating a response sentence by quoting the description or a part of the description; a display function for displaying the initial sentence, the response sentence, the description, and the document on the same screen; A text analysis and generation program characterized by realizing the above.

Images