Document correction program, document correction device and document correction method
The document modification program addresses the limitations of existing technologies by using a sentence generation model to generate tailored modification commands, enhancing the flexibility and accuracy of document corrections, especially in legal documents.
Patent Information
- Application Number
- JP2023073791
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-04-27
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2043-04-27
AI Technical Summary
Existing document correction technologies, such as those described in Patent Document 1, are limited in their ability to make flexible corrections tailored to individual documents, particularly legal documents, as they rely on pre-registered recommended sentences and cannot adapt to the unique content and relationships between clauses.
A document modification program utilizing a sentence generation model, including an acquisition unit, instruction generation unit, and proposal unit, to generate modification commands based on document evaluation, allowing for flexible and tailored corrections.
Enables efficient and flexible document modification, particularly in legal documents, by proposing modifications that consider the content and relationships between clauses, improving the quality and accuracy of document amendments.
Smart Images

Figure 2024158529000001_ABST
Abstract
Description
[Technical Field]
[0001] The present invention relates to a document correction program, a document correction device, and a document correction method. [Background technology]
[0002] The evaluation and creation of documents, especially legal documents such as contracts and regulations, requires a high level of expertise. In light of this background, technologies for checking legal documents and technologies for supporting the revision of legal documents based on the checking have been known.
[0003] For example, Patent Document 1 describes a technology that accepts input of information on review results for legal documents and makes corrections. The technology described in Patent Document 1 uses pre-registered recommended sentences to replace corresponding parts of legal documents with the recommended sentences, thereby correcting the legal documents. [Prior art documents] [Patent documents]
[0004] [Patent Document 1] Japanese Patent Publication No. 2020-080132 Summary of the Invention [Problem to be solved by the invention]
[0005] However, the technology described in Patent Document 1 involves partial replacement of pre-registered recommended sentences, which means that it can only handle content that is anticipated in advance, making it difficult to flexibly modify individual documents. Therefore, the present invention aims to solve the above problems. [Means for solving the problem]
[0006] In order to solve the above problem, a document correction program is provided that utilizes a sentence generation model to suggest corrections to a document based on an evaluation of the document, and causes a computer to function as an acquisition unit, an instruction generation unit, and a suggestion unit, wherein the acquisition unit acquires draft information regarding a draft of the document and evaluation information indicating an evaluation of the draft, the instruction generation unit generates correction instructions regarding the draft based on the draft information and evaluation information and sends them to the sentence generation model, and the suggestion unit sends suggested correction information for correcting the draft based on the output of the sentence generation model.
[0007] With this configuration, it is possible to use a text generation model to assist in revising a document draft based on the evaluation of the draft. This makes it possible to flexibly suggest revisions according to the content of the document, making it easier to revise the document. In particular, even when revising documents that require expertise, such as legal documents, the text generation model makes it easy for users to make revisions themselves, thereby making tasks such as creating and reviewing contracts more efficient.
[0008] In a preferred embodiment of the present invention, the document is a legal document including a plurality of clauses, and the evaluation information includes the results of evaluating the draft with respect to each clause and / or the relationship between the clauses.
[0009] This configuration makes it possible to suggest revisions to a document written by clause based on an evaluation of the content of each clause or the relationships between multiple clauses, thereby supporting more appropriate revisions. For example, if a set of regulations is defined by multiple clauses, the system can appropriately evaluate the regulations that span multiple clauses by treating those clauses as a group and checking the relationships between them. Furthermore, by sending an instruction to revise the draft to the sentence generation model using the evaluation as evaluation information, it is possible to appropriately suggest revisions for such a group of multiple clauses.
[0010] In a preferred embodiment of the present invention, the evaluation information includes an evaluation of any of unnecessary clauses, missing information, and clause contents.
[0011] By adopting such a configuration, it becomes possible to propose appropriate revisions based on an evaluation of whether a clause should be deleted, a clause should be added, or a clause whose content should be revised.
[0012] In a preferred embodiment of the present invention, the command generation unit generates, as the correction command, a sentence that describes instructions regarding the correction for each clause or for each group of multiple clauses that defines a set of regulations, and transmits the generated sentence to the sentence generation model.
[0013] This configuration allows for the appropriate transmission of amendments to legal documents divided into clauses, either for each clause or for each group of clauses that defines a set of provisions, making it easier to execute amendments appropriate for legal documents that include multiple clauses.
[0014] In a preferred embodiment of the present invention, the evaluation information includes a risk level for each of the issues raised about the draft, and the instruction generation unit generates, as the correction instruction, an instruction to correct the draft for issues that are at or above a predetermined risk level, and sends the correction instruction to the writing generation model.
[0015] This configuration makes it possible to propose modifications to high-risk parts, which can then be divided into low-risk parts and parts where the user should consider whether or not to modify them, depending on the risk level.
[0016] In a preferred embodiment of the present invention, the evaluation information includes an evaluation statement that describes an evaluation of the draft in natural language, and the command generation unit generates a sentence based on the draft, the evaluation statement, and a pre-prepared instruction statement as the correction command and transmits the correction command to the sentence generation model.
[0017] With this configuration, natural language sentences can be generated as correction commands, so that correction suggestions can be made using, for example, an interactive sentence generation model that outputs sentences in response to input sentences.
[0018] In a preferred embodiment of the present invention, the computer is further made to function as a re-evaluation unit, which obtains proposed revisions to the draft as an output of the sentence generation model, obtains re-evaluation information indicating an evaluation of the proposed revisions, determines whether the evaluation satisfies a preset criterion, and if the criterion is satisfied, causes the suggestion unit to transmit the proposed revision information, and if the criterion is not satisfied, causes the instruction generation unit to generate a revision instruction for the proposed revision based on the revision information and re-evaluation information and transmit the generated instruction to the sentence generation model.
[0019] This configuration prevents revision proposals that still have problems from being sent as is, and if there are problems, the revision proposals can be reevaluated and revised again, which is expected to improve the quality of the revision proposals.
[0020] In a preferred embodiment of the present invention, the re-evaluation unit causes the suggestion unit and the instruction generation unit to repeatedly obtain re-evaluation information and send correction instructions until the evaluation of the correction proposal satisfies a preset criterion.
[0021] This configuration allows the evaluation and correction to be repeated until the criteria are met, further reducing the occurrence of problems where incomplete parts remain after correction, and enabling higher quality correction suggestions to be made.
[0022] In a preferred embodiment of the present invention, the reevaluation information includes a risk level for each of the issues pointed out for the proposed revision, and the criteria include that the risk levels of all of the issues pointed out for the proposed revision are equal to or lower than a predetermined risk level.
[0023] With this configuration, corrections can be repeated until there are no more issues pointed out that are at or above a predetermined risk level, thereby further improving the quality of correction suggestions.
[0024] In a preferred embodiment of the present invention, when a preset termination condition is met, the re-evaluation unit terminates the repeated re-evaluation and causes the proposal unit to send the proposed revision information, even if the evaluation of the proposed revision does not satisfy the criteria.
[0025] By using such a configuration, infinite loops and extremely repetitive processing can be avoided, and a reduction in processing load can be expected.
[0026] In a preferred embodiment of the present invention, the command generation unit receives an input of a request for correction from a user, generates the correction command based on the request, and transmits the correction command to the sentence generation model.
[0027] This configuration makes it possible to propose modifications that reflect the user's wishes. For example, when making a contract with a specific party, there may be multiple possible modifications, such as making modifications aggressively (to the party's advantage) or making the relationship more equal. By generating a modification command based on the user's wishes, it is possible to propose modifications that better reflect the user's wishes, even in such cases.
[0028] In a preferred embodiment of the present invention, the document is a legal document relating to an agreement with a specific party, and the instruction generation unit acquires background information indicating at least one of the relationship with the party, the type of the agreement, and the wishes of the parties regarding the agreement, generates the correction instruction based on the background information, and sends it to the sentence generation model.
[0029] This configuration makes it possible to propose more appropriate amendments based on the relationship with the other party, the type of agreement, the requests of the parties, etc. For example, by specifying a contract type as the type of agreement, it is possible to propose more appropriate amendments taking the contract type into consideration. Furthermore, when entering into a confidentiality agreement, for example, it is possible to propose amendments that are advantageous to the user by selecting whether the party is the discloser or recipient of information as the relationship with the other party. Furthermore, by selecting whether the party wishes to be aggressive (to provide terms that are advantageous to the party) or to provide terms that are equal, it is possible to propose amendments that reflect the request.
[0030] In a preferred embodiment of the present invention, the draft information is information relating to a draft generated by a model.
[0031] This configuration allows users to complete the document creation process, from drafting to revision, without having to do it themselves, making it easy for users to create legal documents that require expertise, such as contracts.
[0032] In a preferred form of the present invention, the document is a legal document relating to an agreement with a specific party, and the draft information is information about a draft generated by the model based on prior information indicating at least one of the relationship with the party, the type of the agreement, and the parties' wishes regarding the agreement.
[0033] This configuration allows drafts to be created based on the relationship with the other party, the type of agreement, the requests of the parties, etc. This is expected to improve the quality of the final document. [Effects of the Invention]
[0034] According to the present invention, flexible revisions tailored to individual documents can be supported. [Brief explanation of the drawings]
[0035] [Figure 1] 1 is a diagram illustrating a configuration of a document correction system according to an embodiment of the present invention. [Figure 2] FIG. 1 is a diagram illustrating an example of a hardware configuration of a document correction system according to an embodiment of the present invention. [Figure 3] FIG. 1 is a diagram illustrating a functional configuration of a document correction system according to an embodiment of the present invention. [Figure 4] FIG. 10 is a sequence diagram illustrating an example of a processing procedure of the document correction system according to the embodiment of the present invention. [Figure 5] 10 is a display example of an evaluation screen of the document correction system according to the embodiment of the present invention. [Figure 6] 10 is a flowchart showing a processing procedure for a reevaluation process of the document correction system according to the embodiment of the present invention. [Figure 7] 10 is a display example of a correction suggestion screen of the document correction system according to the embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION
[0036] The document correction system of the present invention will be described below with reference to the accompanying drawings, in which preferred embodiments are shown, however this invention may be embodied in many different forms and is not limited to the embodiments set forth herein.
[0037] For example, in this embodiment, the configuration, operation, etc. of the document correction system will be described, but similar effects can also be achieved by a document correction device, a document correction system, or a method executed by a document correction device, a document correction program, etc. In this embodiment, the document correction system takes the form of so-called cloud computing, in which the document correction program is executed on a server (document correction device) connected to the client terminal via a network in order to realize its functions on a client terminal (user terminal device).
[0038] The document correction program may be provided to a computer device as a computer-readable non-transitory storage medium or may be provided as a downloadable program from an external server so that the computer device can function as a document correction device. The document correction program may also be an add-on (plug-in) program for document editing software such as Microsoft Word (registered trademark).
[0039] In this embodiment, the term "unit" may include, for example, hardware resources implemented by a broadly defined circuit and software information processing that can be specifically realized by these hardware resources. In this embodiment, "information" may be represented, for example, by a physical signal value representing voltage or current, a high or low signal value as a binary bit set consisting of 0 or 1, or a quantum superposition (so-called quantum bit), and communication and calculations may be performed on the broadly defined circuit. A broadly defined circuit is a circuit realized by appropriately combining a circuit, a processor, a memory, and the like. For example, the circuit may include a central processing unit (CPU), a graphics processing unit (GPU), a large-scale integration (LSI), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or the like.
[0040] <Definition> First, definitions of terms will be explained. The present invention is a technology for supporting document revision using a sentence generation model. A sentence generation model refers to any model that can output a sentence. A model generated by machine learning can be representatively used as a sentence generation model. In this embodiment, a case will be explained in which a conversational sentence generation model is used, which receives a sentence as an input and generates a sentence as a response to the input sentence.
[0041] More specifically, a sentence generation model is a machine learning model for natural language processing, and is a model with the function of generating sentences based on input information. For example, the sentence generation model is a machine learning model based on a Transformer model with a self-attention mechanism. An example of a Transformer model is a GPT (Generative Pre-trained Transformer) model. By training a sentence generation model using a large dataset, it becomes possible to perform not only general natural language generation tasks but also highly specialized sentence generation tasks. For example, a sentence generation model trained using a large dataset using the GPT-3 or GPT-4 algorithm developed by OpenAI (e.g., a ChatGPT model generated by combining training using a large dataset and reinforcement learning using a reward prediction model) may be applied. Furthermore, although an example of text information being input to the model is described in this embodiment, image information or audio information may also be input.
[0042] In this embodiment, the document is particularly assumed to be a legal document such as a contract, rules, or regulations. Some legal documents include units divided into articles, paragraphs, subsections, etc., and in this embodiment, each of these units is referred to as a "clause." In this embodiment, the legal document is a contract. Examples of contracts include documents that exert contractual effects, such as non-disclosure agreements, sales contracts, and real estate lease agreements. Furthermore, this embodiment is not limited to contracts; for example, the legal document may be a rule, rule, notice, or the like. Examples of rules include documents of legal norms such as work regulations. Examples of notices include documents declaring policies such as privacy. In addition, legal documents include documents that are somehow related to law, such as notices and powers of attorney.
[0043] In the following, a draft refers to a rough draft of a document. In the present invention, information about a draft is called draft information, and the draft information includes information indicating the content of sentences that make up at least a part of the draft. For example, the draft information may include only information about clauses to be revised or clauses related to the revised parts, or the draft information may include information about the entire text of the draft.
[0044] In this embodiment, a revision command is sent to the sentence generation model to output a revision proposal for the draft, which is then used to make a proposal for revising the draft. The revision proposal may be for the entire draft, or may be for a portion of the sentence that is the target of revision, for example.
[0045] The correction command is information indicating a command to cause the writing generation model to output information regarding the correction of the draft. In this embodiment, a conversational writing generation model that takes a sentence as input is used, so a sentence or a sentence in natural language is generated and sent as a correction command. Then, the sentence or a sentence output by the writing generation model as a response is received and a correction is suggested.
[0046] <System configuration> Next, the configuration of the document editing system of this embodiment will be described. FIG. 1 is a diagram showing the configuration of the document editing system of this embodiment. As shown in FIG. 1, the document editing system 0 includes a document editing device 1 and a user terminal device 2, and the document editing device 1 is configured to be able to communicate with a writing generation device 3 via a communication network NW. The document editing device 1 and the user terminal device 2 are also configured to be able to communicate via the communication network NW. The document editing device 1 operates as a server, and the user terminal device 2 operates as a client terminal. The writing generation device 3 is an external server device that provides a writing generation model. Note that the document editing device 1 may be configured to store the writing generation model and be able to edit a document without using an external device, or a configuration may be adopted in which part or all of the processing executed by the document editing device 1 in this embodiment is executed in the user terminal device 2.
[0047] In this embodiment, the communication network NW is an IP (Internet Protocol) network, but there is no limitation on the type of communication protocol, and there is also no limitation on the type and scale of the network.
[0048] It is possible to use a general-purpose server computer or a personal computer as the document correction device 1. It is also possible to configure the document correction device 1 using a plurality of computers.
[0049] A smartphone, a tablet terminal, a personal computer, a wearable device, etc. can be used as the user terminal device 2. The user terminal device 2 has an application (a web browser application or a dedicated application) for making requests to the document correction device 1 and receiving responses, and uses the document correction program executed in the document correction device 1 via this application.
[0050] Based on a request sent from an application on the user terminal device 2, the document correction device 1 sends data in a predetermined format as various display information to the application on the user terminal device 2. The data in the predetermined format is, for example, data in JSON (JavaScript Object Notation) format or XML format. The application on the user terminal device 2 renders the display data based on the received data, and displays information such as evaluation information and correction suggestions on the display of the user terminal device 2.
[0051] For example, when executing a process for displaying revision suggestion information and other information on a web browser application of the user terminal device 2, the web browser application first sends an HTTP (Hypertext Transfer Protocol) request to a web server and receives an HTTP response including the file requested in the HTTP request. The received file is display information including the revision suggestion information, indicates the content or instructions of the web page to be displayed, and is written using a markup language such as HTML (Hypertext Markup Language). A script, which is a program executed on the web browser application, may be attached to the received file. The web browser application draws (renders) a web page in the display area based on the received file. If a script is attached to the file, the browser executes the script and displays the web page according to the execution result.
[0052] <Hardware configuration> 2(a) is a diagram showing an example of the hardware configuration of the document correction device 1. The document correction device 1 includes a control unit 101, a storage unit 102, and a communication unit 103 as the hardware configuration.
[0053] The control unit 101 includes one or more processors such as a CPU, and controls the overall operation and processing of the document correction device 1 by executing the document correction program according to the present invention, OS, browser software, and other applications.
[0054] The storage unit 102 is a hard disk drive (HDD), read only memory (ROM), random access memory (RAM), etc., and stores the document correction program according to the present invention and data used when the control unit 101 executes processing based on the program. The control unit 101 executes processing based on the document correction program stored in the storage unit 102, thereby realizing the functional configuration described below.
[0055] The communication section 103 controls communication with the communication network NW, and performs inputs required to operate the document editing device 1 and outputs related to the operation results.
[0056] 2(b) is a diagram showing an example of the hardware configuration of the terminal device 90 (user terminal device 2). Similarly, the terminal device 90 also includes, as its hardware configuration, a control unit 901, a storage unit 902, a communication unit 903, an input unit 904, and an output unit 905.
[0057] The control unit 901 of the terminal device 90 includes one or more processors such as a CPU, and controls the overall operation and processing of the terminal device 90. The storage unit 902 of the terminal device 90 is an HDD, ROM, RAM, etc., and stores the above-mentioned applications and data used when the control unit 901 executes processing based on a program. The communication unit 903 of the terminal device 90 controls communication with the network. The input unit 904 of the terminal device 90 is a mouse, keyboard, etc., and inputs operation requests from the user / provider to the control unit 901. The output unit 905 of the terminal device 90 is a display, etc., and displays the results of processing by the control unit 901, etc.
[0058] <Functional configuration> Next, the functional configuration of this embodiment will be described with reference to Fig. 3. The document correction device 1 has, as its functional configuration, an acquisition unit 11, an instruction generation unit 12, an evaluation unit 13, and a proposal unit 14. In this embodiment, the evaluation unit 13 functions as the evaluation unit and re-evaluation unit of the present invention. This is a specific implementation of information processing by software stored in the storage unit 102 using hardware (such as the control unit 101). For example, the control unit 101 reads out the document correction program stored in the storage unit 102, expands it into a temporary storage area (such as RAM) in the storage unit 102, executes it, and controls the communication unit 103, etc., thereby realizing the functions of the document correction device 1.
[0059] The acquisition unit 11 acquires draft information related to a document draft and evaluation information indicating an evaluation of the draft. The draft information is acquired based on the draft information transmitted from the user terminal device 2. Specifically, for example, the user can upload a file containing a document draft and acquire draft information by accepting an operation to specify at least a part of the contents of the draft. Note that if the entire draft is to be revised, the operation to specify the part to be revised may be omitted. Alternatively, for example, the configuration may be such that the user uploads a part of the document draft to be revised, and draft information about that part is acquired.
[0060] Here, the user can also use a draft created using the model. In this embodiment, a case will be described in which the same sentence generation model as the sentence generation model used for revising the draft is used as the model for generating the draft. For example, the document revision device 1 receives advance information from the user terminal device 2 and transmits a draft generation command based on that to the model, thereby obtaining the draft generated by the model. Furthermore, a sentence generation model different from the sentence generation model used for revising the draft may be used as the model for generating the draft.
[0061] For example, if the document is a legal document regarding an agreement with a specific party, the prior information may include information such as the relationship with the party, the purpose of the agreement, the type of agreement, the parties' requests regarding the agreement, the desired content of the agreement, and a template for the legal document. Here, the term "agreement" refers to a contract, rule, or regulation. The relationship with the party refers to the party's position in the agreement, such as the discloser or recipient of information in a confidentiality agreement. The type of agreement may include the contract type specified in the legal document, atypical contract, or type of regulation or rule. The parties' requests may include whether the agreement is advantageous to the party or an equal agreement, or any other arbitrary request. This prior information may be input in the form of a user selecting from preset options, or may be text information written freely in natural language.
[0062] The advance information may also be term sheet information indicating the contents of a term sheet used in contract negotiations. A term sheet is a table that summarizes the contents of a proposal or agreement reached with the other party regarding the main items of a contract, and term sheet information includes the contents of the term sheet. The advance information may also be question information based on the questions and question results generated by questioning software that asks questions about the contents to be included in a legal document (e.g., a contract).
[0063] The evaluation information is information indicating the evaluation of the draft. In this embodiment, the evaluation unit 13 of the document editing device 1 evaluates the draft, and the result is acquired by the acquisition unit 11 as evaluation information. However, the present invention is not limited to this, and the evaluation information may be acquired from an external device or the user terminal device 2. Details of the evaluation process for the draft and the evaluation information will be described later.
[0064] Furthermore, the acquisition unit 11 of this embodiment further acquires background information about the background of the contract or the like stipulated in the target legal document. The background of the contract or the like includes information such as the relationship with the other party, the purpose of the agreement, the type of agreement, the requests of the parties regarding the agreement, the content of the agreement to be reached, and a template for the legal document. The background information can be similar to the above-mentioned prior information. In this embodiment, evaluation and generation of correction instructions are performed based on such background information.
[0065] That is, for example, if the document is a legal document regarding an agreement with a specific party, the background information may include information such as the relationship with the party, the purpose of the agreement, the type of agreement, the parties' requests regarding the agreement, the content of the agreement, and a template for the legal document. Here, "arrangement" refers to a contract, rule, regulation, etc. The relationship with the party refers to the position in the agreement, such as the discloser or recipient of information in a confidentiality agreement. The type of agreement may include the contract type specified in the legal document, atypical contract, or type of regulation or rule, and the parties' requests may include wishes such as whether the agreement is advantageous to the party or an equal contract, comments received from the other party to the agreement, and other arbitrary requests.
[0066] This background information may be input in the form of a user selecting from preset options, or may be text information written freely in natural language. Furthermore, the background information may be generated by a text generation model that receives a legal document and an extraction command to extract background information from the legal document. For example, the extraction command is a command to estimate the contract type to which the input legal document corresponds, from among multiple predetermined contract types.
[0067] The command generation unit 12 generates a revision command for the draft based on the draft information and evaluation information acquired by the acquisition unit 11, and transmits the generated command to the sentence generation model. In this embodiment, input to the sentence generation model is performed by transmitting the revision command to the sentence generation device 3 that provides the sentence generation model. The command generation unit 12 in this embodiment generates a revision command as a sentence in natural language using the draft information and evaluation information. The process of generating a revision command will be described in detail later.
[0068] The evaluation unit 13 functions as a re-evaluation unit of the present invention, and receives a proposed revision of the draft from the sentence generation device 3 as an output of the sentence generation model in response to a revision command. Then, it acquires re-evaluation information indicating an evaluation of the proposed revision, and determines whether the evaluation satisfies a preset standard. Note that in this embodiment, the evaluation unit 13 (re-evaluation unit) evaluates the proposed revision and generates re-evaluation information as a result, but it may also be possible to transmit the proposed revision to an external evaluation device and acquire re-evaluation information from the evaluation device as the evaluation result.
[0069] If the evaluation satisfies the criteria, the evaluation unit 13 (re-evaluation unit) causes the proposal unit 14 to send proposed revision information. On the other hand, if the evaluation does not satisfy the criteria, the evaluation unit 13 instructs the instruction generation unit 12 to generate a revision command for the proposed revision based on the revision information and re-evaluation information for the proposed revision. The instruction generation unit 12 then sends the revision command to the sentence generation model, which then generates a proposed revision for the proposed revision.
[0070] The evaluation unit 13 (re-evaluation unit) of this embodiment is configured to repeatedly acquire and evaluate revision proposals until the evaluation of the revision proposals satisfies a predetermined criterion or a termination condition is met. The predetermined criterion can be, for example, that the risk level for each of the findings contained in the re-evaluation information (described later) is equal to or lower than a predetermined risk level. The termination condition can be, for example, to set an upper limit on the number of repetitions or a lower limit on the number of characters changed by the revision. Details of the re-evaluation process will be described later.
[0071] The suggestion unit 14 transmits correction suggestion information based on the output of the sentence generation model in response to the correction command. In this embodiment, in accordance with the determination result by the evaluation unit 13 (re-evaluation unit), display information including the correction suggestion information is transmitted to the user terminal device 2, whereby the user terminal device 2 displays the correction suggestion.
[0072] The suggestion unit 14 of this embodiment generates revision suggestion information including sentences in natural language to support revision of the draft based on the output of the sentence generation model acquired by the evaluation unit 13 (re-evaluation unit). Details of the revision suggestion information and the display contents on the user terminal device 2 will be described later.
[0073] <Processing Procedure> Next, with reference to FIG. 4, the details of the processing from creating a draft of a contract to proposing revisions in the document revision system of this embodiment will be described. FIG. 4 is a sequence diagram showing an example of the processing procedure by the document revision device 1, the user terminal device 2, and the text generation device in the document revision system of this embodiment. Note that the processing shown here is an example, and omissions and modifications may be made without changing the gist of the present invention. Furthermore, the documents to be processed may be other legal documents such as regulations, etc., in addition to contracts. In the following example, both the creation of the draft and the creation of the proposed revisions (described later) are performed using the same text generation model, but they may also be created using different models.
[0074] (1) Drafting Steps S401 to S406 are processes related to creating a draft. Note that these processes are not essential, and the processes from step S407 onwards may be performed using a draft created by the user himself / herself, a draft generated using an external machine learning model, or a draft contract received from a party with whom the contract is to be concluded.
[0075] First, in step S401, the user terminal device 2 accepts input of preliminary information from the user and transmits it to the document editing device 1. Examples of the preliminary information include the relationship with the contracting party, the purpose of the contract, the type of contract, and the parties' requests regarding the contract. Specifically, when concluding a confidentiality agreement when presenting information to a business partner, examples of the relationship with the other party include "discloser of information," the type of contract is the contract type "confidentiality agreement," and the parties' requests include arbitrary requests such as "I want to make an aggressive contract (favorable to my position)." It is also preferable that the names or titles of the contracting parties and their names in the contract (Party A, Party B) be input as preliminary information. Term sheet information indicating the contents of a term sheet used in contract negotiations and question information based on questions about the contents to be included in the legal document and the answers to those questions may also be acquired as preliminary information.
[0076] The input of this prior information is typically assumed to be made by the user selecting from options, but it may also be made by the user freely writing text in natural language.
[0077] Next, in step S402, the command generation unit 12 generates a creation command for causing the sentence generation model to create a draft based on the prior information. Since the sentence generation model of this embodiment is a conversational model that receives a sentence in a natural language as input and generates a sentence as a response, the creation command is generated as a sentence in a natural language.
[0078] More specifically, the command generating unit 12 can generate a creation command by generating a sentence that combines a prepared instruction sentence and a wording specified based on the prior information. Alternatively, for example, a command sentence may be prepared in advance for each option of the prior information and stored in the storage unit of the document editing device 1, a command sentence may be selected from the prior information according to the contract type, and the command generating unit 12 may generate a sentence that combines the instruction sentence and the wording specified based on the prior information as a generation command.
[0079] More specifically, when the prior information is input by selection, first, a word corresponding to the selected prior information is identified. For example, by previously setting corresponding words for each option of the prior information, the word corresponding to the selected prior information can be identified. Then, the command generation unit 12 generates a creation command based on the words and a command statement separately registered in advance in the storage unit of the document correction device 1. On the other hand, when the prior information is input as free text, the creation command can be generated using the input prior information and a command statement prepared in advance.
[0080] Here, it is preferable to store a character string describing a document creation rule corresponding to the prior information as a word corresponding to the prior information. For example, in a legal document, the legal meaning may differ between describing an exception within a sentence of a provision and describing a disclaimer separately after the sentence. By combining words specifying rules for such description methods according to the prior information, an appropriate legal document can be created. Note that the document correction device 1 may determine a document creation rule to be included in the creation command based on the prior information from among multiple document creation rules corresponding to each piece of prior information, or a configuration may be adopted in which multiple document creation rules are included in the creation command and a statement indicating that an appropriate document creation rule from the multiple document creation rules will be used based on the prior information is included in the creation command, thereby determining the document creation rule to be used by the text generation model.
[0081] As an instruction sentence, for example, "Please create a document based on the following conditions," or "Please input prior information indicating the contractual arrangements. Please create an appropriate contract sentence based on the following rules," etc., can be stored in advance in the storage unit of the document editing device 1. Alternatively, an instruction sentence can be prepared for each contract type as a choice of prior information, and information corresponding to the prior information can be included in the instruction sentence, such as "Please create a confidentiality agreement based on the following conditions." Then, by generating a sentence in which the instruction sentence is followed by words based on the prior information, it is possible to generate a creation command for instructing the sentence generation model to create a draft.
[0082] Once the creation command is generated, the process proceeds to step S403, where the command generation unit 12 transmits the generated creation command to the text generation device 3. The transmission of the creation command to the text generation device 3 in step S403 corresponds to the transmission of the creation command to the text generation model. Upon receiving the creation command, the text generation device 3 inputs the creation command into the text generation model. This causes the text generation model to output a draft of the contract (step S404).
[0083] The writing generation device 3 transmits the draft output by the writing generation model to the document editing device 1 in step S405. Then, in step S406, the document editing device 1 transmits the draft to the user terminal device 2. Through the above processing, the user can obtain a draft of the desired contract by simply inputting the advance information. Note that step S406 and steps S407 to S410 described next may be omitted, and the process up to the revision of the draft may be carried out in a consistent manner without waiting for instructions from the user.
[0084] (2) Draft Evaluation Next, the evaluation process of the draft will be described. Steps S407 to S409 are processes related to the evaluation of the draft. This process is for acquiring evaluation information indicating the evaluation of the draft. In this embodiment, as will be described below, the evaluation unit 13 evaluates the draft, but the evaluation information may be acquired by having an external device evaluate the draft, or by receiving the evaluation information from the user terminal device 2 together with the draft information.
[0085] In step S407, the user terminal device 2 transmits an evaluation request to the document editing device 1. The evaluation request includes draft information and background information related to the draft. The draft information transmitted here is information including at least the contents of the draft to be evaluated, and in this embodiment, a document file containing a draft of a contract is uploaded. The background information is information related to the background of the document. For example, the background information includes the relationship with the other party in the agreement indicated by the target legal document, the type of agreement, and the requests of the parties regarding the agreement. The background information can be information with the same content and format as the advance information in draft creation described above.
[0086] In the document correction device 1, the acquisition unit 11 receives an evaluation request, and the evaluation unit 13 evaluates the draft based on the draft information and background information. The draft may be evaluated by any method, such as evaluation using a machine learning model or rule-based evaluation. As an example, in this embodiment, a method of evaluating clauses based on rules in accordance with background information is used.
[0087] In particular, in this embodiment, the content of each clause and the relationships between multiple clauses are evaluated using the contract type and contractual position from the background information, and the title of each clause in the draft. Here, multiple clauses refers to a group of multiple clauses that defines a set of regulations. For example, within a group, the presence or content of clause B may be determined based on the content of clause A, and then clause C may be confirmed based on the result. By specifying that the evaluation process branches depending on the content, it is possible to evaluate the relationships between multiple clauses in a rule-based manner.
[0088] When evaluating a draft, it is evaluated based on whether it contains unnecessary clauses (or clauses that would be disadvantageous to the party in relation to the other party in the background information), whether there are clauses that should be added, and whether the content of the clauses is appropriate.
[0089] In this embodiment, the issues raised in the draft are identified by linking them to the above rules and registering them in advance in the storage unit of the document correction device 1. Specifically, for each issue raised, the above-mentioned evaluation rule, the type of the issue (either deletion, addition, or content modification of a clause), the content of the issue, and the risk level are stored in advance in the storage unit.
[0090] The comments are statements that suggest revisions or point out problems with the draft, such as, "Isn't it necessary to specify a consultation clause?" or "Isn't the confidentiality obligation period set longer than necessary?" In addition, the comments may include explanatory text about the clause in question or template information for that clause. The risk level is information that indicates the degree of risk assumed for each comment. The risk level may be fixed (unchangeable) on the system side, or it may be changeable by user input.
[0091] Examples of risk level settings are as follows. Issues that are likely to become issues after the contract is signed and involve disputes over the company's important rights and obligations can be assigned a risk level of "A." Examples include "intellectual property ownership," "scope of damages," and "definition of confidential information." Issues that are unlikely to become issues even if omitted can be assigned a risk level of "C." Examples include "confidential information exceptions" and "intermediate reporting obligations." Issues that do not fall into either A or C can be assigned a risk level of "B." Risk levels can be assigned for each issue, or they can be determined based on the wording of the clause. For example, risk levels can be set based on whether the clause's wording is favorable or unfavorable for the company.
[0092] As described above, the evaluation unit 13 executes the evaluation process and creates evaluation information indicating the evaluation of the draft. The evaluation information includes, for each of one or more issues identified in the above process, the clause in the draft corresponding to the issue, the content of the issue, and a risk level. Also, when evaluation information is acquired from an external device or the user terminal device 2, evaluation information is acquired for each issue, including the clause in the draft corresponding to the issue, the content of the issue, and a risk level. As described above, the content of the issue includes information written in natural language, and this description corresponds to the evaluation statement in the present invention.
[0093] Note that a contract may be evaluated using a text generation model. The text generation model may be the same as or different from the text generation model used for draft creation or revision. For example, the document revision device 1 acquires draft information and background information, reads out evaluation rules to be used for contract evaluation that are stored in advance based on the background information, generates evaluation commands based on the draft information and the evaluation rules corresponding to the background information, transmits the generated evaluation command to the text generation device 3, and acquires the evaluation result (evaluation information) of the contract generated by the text generation device 3. The evaluation command is generated as a text in natural language. This is expected to reduce the impact on the evaluation due to variations in the writing style and differences in expression in the text. Furthermore, it is expected that an appropriate evaluation will be performed by using an evaluation rule (e.g., an evaluation rule set for each contract type) read out from among multiple evaluation rules based on background information (e.g., contract type).
[0094] Furthermore, evaluation using a sentence generation model and evaluation using a rule base may be performed simultaneously. In this case, if the evaluation using the sentence generation model and the evaluation using the rule base differ, it is preferable to perform processing that prioritizes the evaluation using the rule base. This makes it possible to guarantee a minimum evaluation using the rule base even if there is uncertainty in the evaluation using machine learning. Note that the document correction device 1 may determine, based on the background information, an evaluation rule to be included in the correction command from among multiple evaluation rules corresponding to each piece of background information. Alternatively, the evaluation command may include multiple evaluation rules and include a command to use an appropriate rule from the multiple rules based on the background information, thereby determining the evaluation rule to be used by the sentence generation model.
[0095] When the evaluation unit 13 finishes evaluating the draft, in step S409 the evaluation unit 13 transmits the evaluation information to the user terminal device 2. When the user terminal device 2 receives the evaluation information, the evaluation information is subjected to display processing and the processing result is displayed on the display.
[0096] FIG. 5 shows an example of an evaluation screen displayed on the user terminal device 2 as a result of the evaluation information display process. The evaluation screen W5 displays an outline of the document draft and an outline of the evaluation results at the top of the screen. The contents of the draft identified by the draft information are displayed on the left side of the screen. In addition, a "Read Content Edit" button is provided, and if there is an error in the content identified from the draft information, the content can be edited by selecting this button.
[0097] The right side of the screen displays the evaluation results based on the evaluation information. For each issue, the type of issue (deletion, addition, or revision of the clause) is displayed, along with an explanation and a template of the clause in question.
[0098] Furthermore, for each pointed out item, a differently shaped point out mark W51 is attached for each type of point out, and a serial number for each type is displayed. Similar point out marks W51 are also displayed along with the serial number in the corresponding part of the draft on the left side of the screen, so that the user can recognize the position in the draft that corresponds to the pointed out item, and the corresponding part in the draft is highlighted. In the example in Figure 5, a circle indicates a point out that the content of a clause needs to be revised, and a triangle indicates a point out that a clause is missing.
[0099] The indication mark W51 is displayed in a different color depending on the risk level. By operating the risk filter operation section W52, the indications to be displayed can be filtered according to the risk level. In FIG. 5, the setting is shown to display indications with a risk level of "B" or higher, i.e., indications of A or B.
[0100] Additionally, a batch correction button W53 and an individual correction button W54 are displayed together with the evaluation results. When the user selects these buttons, the correction request of step S410 in Fig. 4 is sent. The range of correction instructions varies depending on the button selected.
[0101] The batch correction button W53 has three options: "Create correction proposals for all items at once," "Create correction proposals for item A," and "Create correction proposals for items AB." When "Create correction proposals for all items at once" is selected, a correction request is sent for all findings. When "Create correction proposals for item A" is selected, a correction request is sent for findings with a risk level of A. When "Create correction proposals for items AB" is selected, a correction request is sent for findings with a risk level of A and findings with a risk level of B. Furthermore, when the individual correction button W54 displayed for each finding is selected, a correction request for that finding is sent.
[0102] Furthermore, in this embodiment, when the Bulk Correction button W53 or the Individual Correction button W54 is selected, further requests for corrections from the parties involved are accepted. For example, for a problem that has multiple possible correction strategies, the user can specifically select one of them to accept a request for correction. Alternatively, when requesting corrections for multiple problems using the Bulk Correction button W53, the user may be prompted to select a general strategy, such as "bullish (highly favorable to the company)," "neutral," or "bearish (more favorable to the company than before the corrections, but more favorable to other companies)." In this case, a correction command may be generated that generates a correction proposal only for the strategy selected by the user, or a correction command may be generated that generates a correction proposal for all strategies. Such requests may also be entered as free text.
[0103] (3) Corrective action Next, the draft revision process will be described. When a revision request is received as described above, the acquisition unit 11 acquires draft information and evaluation information. In addition, the acquisition unit 11 of this embodiment further acquires the background information described above and information on requests for revision that is sent together with the revision request in step S410. In this embodiment, the draft received in step S405, the background information received in step S407, and the evaluation information generated by the evaluation unit 13 in step S408 are stored in a storage unit, and the acquisition unit 11 acquires the draft information, evaluation information, and background information based on these. Note that if the draft is created and evaluated by an external device, the acquisition unit 11 may acquire the draft information and evaluation information from the user terminal device 2 or the external device.
[0104] If the revision request specifies only partial revision of the draft, not the entire scope thereof, the draft information does not need to include the contents of the entire draft, and only the portion related to the subject of revision may be included in the draft information. However, for example, if there is a group of multiple clauses that define a set of regulations, and revisions are made to the pointed out matters regarding the clauses included in that group, it is preferable that the draft information include the contents of all the multiple clauses in the group, rather than just the single clause that corresponds to the pointed out matters.
[0105] Next, in step S411, the command generating unit 12 generates a correction command for causing the sentence generation model to generate a proposed revision of the draft, based on the draft information, evaluation information, background information, and information on requests from the parties regarding revisions acquired by the acquiring unit 11. Since the sentence generation model of this embodiment is a conversational model that receives a sentence in a natural language as input and generates a sentence that will be a response, the correction command is generated as a sentence in a natural language.
[0106] The instruction generation unit 12 of this embodiment generates, as a correction command, a sentence that describes instructions regarding correction for each clause corresponding to the pointed out item. In particular, if there are multiple pointed out items in a single clause or a group of clauses that defines a set of regulations, generating a correction command for each pointed out item may result in multiple contradictory correction proposals. Therefore, the instruction generation unit 12 of this embodiment generates a sentence of correction command for each clause or group of clauses that corresponds to the multiple pointed out items as a correction command for a clause or group of clauses that has multiple pointed out items.
[0107] The generation of the correction command will be specifically described below. First, for background information, when background information is input by selection, the words corresponding to the selected background information are identified. For example, by setting corresponding words in advance for each background information option, the words corresponding to the selected background information can be identified.
[0108] Similarly, when request information is input by selection, the words corresponding to the selected request are first identified. For example, by setting corresponding words in advance for each combination of the target indicated matter and the request options, the words corresponding to the selected request can be identified for each indicated matter.
[0109] Then, the command generation unit 12 generates a correction command based on those words, the draft information, and a command statement separately registered in advance in the storage unit of the document correction device 1. On the other hand, when the background information and the request are input as free text, the correction command can be generated using the input background information and a command statement prepared in advance.
[0110] Here, as with the generation of creation commands in draft creation, it is preferable to store a string of characters describing document creation rules corresponding to the background information as a term corresponding to the background information. For example, in a legal document, the legal meaning may differ between stating an exception within a sentence of a provision and stating a disclaimer separately after the sentence. By combining terms specifying rules for such writing methods according to the background information, an appropriate legal document can be created. Note that the document correction device 1 may determine, based on the background information, which rule to include in the correction command from among multiple rules corresponding to each piece of background information. Alternatively, multiple rules may be included in the correction command and a statement indicating that an appropriate rule from the multiple rules should be used based on the background information may be included in the correction command, thereby determining the rule to be used by the sentence generation model.
[0111] Instructions such as "The following comments were made about the above draft as a result of the review. Please prepare a revised draft based on the following conditions, taking the comments into consideration," or "The following comments were made about the above draft as a result of the review. Please prepare an appropriate contract document based on the following rules, taking the comments into consideration," can be stored in advance in the storage unit of the document correction device 1. Then, a sentence that combines the contents of the draft, the instruction, and words based on the background information and the request can be generated as a creation command to instruct the text generation model to create a draft.
[0112] Once the correction command is generated, the process proceeds to step S412, where the command generation unit 12 transmits the generated correction command to the sentence generation device 3. The transmission of the correction command to the sentence generation device 3 in step S412 corresponds to the transmission of the correction command to the sentence generation model. The correction command for each clause or for each group of clauses generated as described above may be transmitted in multiple batches. For example, it may be configured to first transmit the directive portion of the correction command, followed by words corresponding to the points of criticism or requests. It may also be configured to determine whether to transmit the correction command all at once or in multiple batches depending on the number of characters in the correction command.
[0113] When the writing generation device 3 receives the revision command, it inputs the revision command to the writing generation model. As a result, the writing generation model outputs a revision proposal for the draft (step S413). In this embodiment, the evaluation unit 13 (re-evaluation unit) acquires a revision proposal for each pointed out item. Note that the command generation unit 12 may generate a revision command instructing the system to propose multiple revision proposals for one pointed out item in step S411, and the evaluation unit 13 (re-evaluation unit) may acquire multiple revision proposals for some or all of the pointed out items in step S413.
[0114] Furthermore, a revision command may be sent to the sentence generation model to generate revision reasons based on the evaluation information, and the revision reasons may be acquired in addition to the revision proposal. Furthermore, revision reasons for the company (or for the company's client) and revision reasons for the other party may be generated and acquired. In contract negotiations, the company and the other party repeatedly make revisions to each other, and at this time, it is common to write at least one of revision reasons for explanation within the company and revision reasons for explanation outside the company. Writing revision reasons requires a high level of expertise, but the above configuration can help even users without specialized knowledge to write revision reasons.
[0115] (4) Reevaluation process In step S414, the sentence generation device 3 transmits the proposed revision, which is the output of the sentence generation model, to the document correction device 1. In step S415, the document correction device 1 executes reevaluation by the evaluation unit 13 (reevaluation unit). The evaluation unit 13 (reevaluation unit) acquires the proposed revision and reevaluates the proposed revision in the same procedure as the evaluation process described above. Then, reevaluation information indicating the evaluation of the proposed revision is generated in the same manner as the evaluation information described above. In other words, the reevaluation information is information indicating the evaluation of the proposed revision, and the information included in the reevaluation information is the same items as the evaluation information described above.
[0116] In this case, the reevaluation may be conducted not on the entire draft, but only on the parts related to the issues pointed out above. Specifically, the clauses or groups of clauses indicated in the proposed amendments may be reevaluated, and the reevaluation of other parts may be omitted. This is expected to improve the processing speed.
[0117] The evaluation unit 13 (reevaluation unit) then determines whether the evaluation satisfies a predetermined standard based on the reevaluation information. Specifically, it checks the risk levels of the findings in the reevaluation information and determines that the standard is met if the risk levels of all the findings are equal to or lower than a predetermined risk level.
[0118] If the reevaluation of the proposed revision does not satisfy the criteria, i.e., if there are any findings with a risk level exceeding the predetermined risk level, the process returns to step S411 and repeats the process from generating the revision command to reevaluating in step S415. In this case, in the second and subsequent revision command generation processes, the revision command for the proposed revision is generated using the same procedure as for generating the revision command described above, using the revision command information instead of the draft information and the reevaluation information instead of the evaluation information.
[0119] In principle, the loop process is performed until a preset criterion is met, but if the number of repetitions becomes extremely large or the loop cannot be terminated, there is a concern that the processing load will increase. Therefore, in this embodiment, an upper limit on the number of loop processes is set as the termination condition, and even if the reevaluation does not satisfy the criterion, if the termination condition is met, the reevaluation repetition is terminated, the loop process is terminated, and the process proceeds to step S417. Other termination conditions may also be set, such as when the number of changed characters in a draft before and after revision is below a certain level. This prevents repeated situations in which only minor revisions are made.
[0120] On the other hand, if the evaluation of the proposed modification satisfies the preset criteria, the process enters into a loop interruption process (break process) in the sequence diagram of Fig. 4. In step S416, the proposing unit 14 generates modification proposal information and ends the loop process (Loop process).
[0121] In the interruption process, the proposing unit 14 generates revision proposal information. The revision proposal information includes, for each of the findings in the reevaluation, the clause in the draft corresponding to the finding, the findings, the risk level, and the revision proposal created in the most recent step S413.
[0122] Then, in step S417, the suggestion unit 14 transmits the revision suggestion information generated in step S416 to the user terminal device 2. The user terminal device 2 displays the display processing result of the revision suggestion information on the display, thereby allowing the user to receive suggestions for revising the draft.
[0123] The above-mentioned procedure of the re-evaluation process will be further explained with reference to the flowchart in Fig. 6, particularly with regard to the conditions for repetition. When the document correction device 1 receives the proposed revision, which is the output of the sentence generation model, in step S414 in Fig. 4, the evaluation unit 13 (re-evaluation unit) acquires this proposed revision in step S501.
[0124] Then, in step S502, a reevaluation is performed (corresponding to step S415 in FIG. 4), and in step S503, it is determined whether the evaluation of the proposed revision satisfies the preset criteria. As described above, in this embodiment, the risk levels of the identified items in the reevaluation information are checked, and if the risk levels of all identified items are equal to or lower than a predetermined risk level, it is determined that the criteria are met.
[0125] If the criteria are not met, the process proceeds to step S505, where it is confirmed whether the number of loops (number of corrections) has reached the upper limit. If the upper limit has not been reached ("Y" in step S505), the process proceeds to step S506, where the evaluation unit 13 (re-evaluation unit) causes the instruction generation unit 12 to generate and send a correction instruction again, increments the loop count counter, and returns to step S501.
[0126] If the criteria are met in step S503, or if the number of revisions reaches the upper limit in step S505, the process proceeds to step S504, where the evaluation unit 13 (re-evaluation unit) causes the proposal unit 14 to generate revision proposal information (corresponding to step S416 in FIG. 4). The proposal unit 14 then transmits the revision proposal information to the user terminal device 2 (corresponding to step S417 in FIG. 4), and the process ends. Note that the above processing procedure is an example, and may be changed as desired, for example, by switching the order of judgment between step S503 and step S505.
[0127] (5) Correction proposal screen Next, a correction proposal screen that is displayed on the user terminal device 2 as a result of the correction proposal information display process will be described with reference to Fig. 7. Fig. 7 is a display example of the correction proposal screen in this embodiment. The correction proposal screen W7 in this embodiment is displayed in a similar configuration to the evaluation screen W5 shown above, and the indication mark W71 and risk filter operation section W72 are the same as the indication mark W51 and risk filter operation section W52 in the evaluation screen W5, so their description will be omitted.
[0128] The difference from the evaluation screen W5 is mainly in the content displayed on the right side of the screen, particularly the section related to "Proposed Revisions." The proposed revision screen W7 includes proposed revisions W73 for each issue based on the proposed revisions included in the proposed revision information. The proposed revisions W73 display the proposed revisions and allow the text to be copied. It is also preferable to highlight changes from the draft to make them easy for users to recognize. This allows users to obtain the details and proposed revisions for each issue, allowing even users without extensive legal knowledge to make revisions to legal documents on their own.
[0129] If multiple correction suggestions are obtained for a pointed out item, the correction suggestion information will include multiple correction suggestions for the pointed out item, so it is preferable to display multiple correction suggestions W73 for the pointed out item on the correction suggestion screen W7 so that the user can select from them.
[0130] Note that the method for outputting the revision proposal information is not limited to displaying the revision proposal screen W7 described above, and for example, the proposing unit 14 may transmit a document-format file of the entire draft to which revision proposals have been applied to at least some of the pointed-out items to the user terminal device 2. Furthermore, the user may be prompted to select whether or not to reflect the revision proposals displayed on the revision proposal screen W7, or which revision proposals to reflect, and the proposing unit 14 may further transmit a document-format file in which the entire draft has been revised based on the result of the selection to the user terminal device 2.
[0131] As described above, the document revision system of this embodiment can support users who do not have specialized knowledge in revising drafts of legal documents such as contracts. Furthermore, the quality of draft revisions can be improved by reevaluating the proposed revisions and issuing a command to create a new proposed revision based on the results.
[0132] Furthermore, by generating commands in natural language text, it becomes possible to create revision proposals using a conversational text generation model that takes text as input and generates response text. This allows the model to generate revision proposals that take into account various conditions such as the contract's background and requests, making it possible to provide draft revision proposals that are more in line with the user's situation and wishes.
[0133] Furthermore, a revision command may be sent to the sentence generation model to generate revision reasons based on the evaluation information, and the revision reasons may be included in the revision proposal information. Furthermore, revision reasons may be generated as revision reasons for the company itself (or for the company's client) and revision reasons for the other party. In contract negotiations, the company and the other party repeatedly make revisions to each other, and at this time, it is common to write at least one of revision reasons for explanation within the company and revision reasons for explanation outside the company. Writing revision reasons requires a high level of expertise, but the above configuration can help even users without specialized knowledge to write revision reasons. [Explanation of symbols]
[0134] 0: Document correction system 1: Document editing device 2: User terminal device 3: Sentence generation device 11: Acquisition part 12: Instruction generation section 13: Evaluation section 14: Proposal Department 90: Terminal equipment 101: Control unit 102: Storage section 103: Communications Department 901: Control unit 902: Storage section 903: Communications Department 904: Input section 905: Output section NW: Communication network W5: Evaluation screen W51: Pointing out mark W52: Risk filter operation section W53: Bulk edit button W54: Individual edit button W7: Correction Suggestion Screen W71: Pointing out mark W72: Risk filter operation section W73: Correction proposal
Claims
1. 1. A document correction program for suggesting corrections to a document based on evaluations of the document using a sentence generation model, the program comprising: A computer is caused to function as an acquisition unit, an instruction generation unit, and a suggestion unit; The acquisition unit acquires draft information related to a draft of the document and evaluation information indicating an evaluation of the draft, the instruction generation unit generates a correction instruction for the draft based on the draft information and the evaluation information, and transmits the correction instruction to the writing generation model; The suggestion unit transmits correction suggestion information for correcting the draft based on an output of the sentence generation model.
2. the document is a legal document containing a number of clauses; The document correction program of claim 1 , wherein the evaluation information includes a result of evaluating the draft with respect to each clause and / or relationships between multiple clauses.
3. 3. The document correction program according to claim 2, wherein the evaluation information includes an evaluation regarding any one of unnecessary clauses, missing information, and contents of clauses.
4. The document correction program according to claim 2 or claim 3, wherein the instruction generation unit generates, as the correction instruction, a sentence that describes instructions regarding correction for each clause or for each group of multiple clauses that defines a set of regulations, and transmits the generated sentence to the sentence generation model.
5. The evaluation information includes a risk level for each of the issues pointed out in the draft, The document correction program according to claim 1 , wherein the instruction generating unit generates, as the correction instruction, an instruction to correct the draft for an issue that is at or above a predetermined risk level, and transmits the instruction to the text generation model.
6. the evaluation information includes an evaluation statement in which an evaluation of the draft is described in a natural language; The document correction program according to claim 1 , wherein the instruction generating unit generates a sentence based on the draft, the evaluation sentence, and a prepared instruction sentence as the correction instruction and transmits the correction instruction to the sentence generation model.
7. The computer further functions as a re-evaluation unit; The re-evaluation unit is obtaining a proposed revision of the draft as an output of the sentence generation model; obtaining reassessment information indicative of an evaluation of the proposed revision and determining whether the evaluation satisfies a predefined criterion; If the criteria are met, the suggestion unit transmits the modification suggestion information; 2. The document correction program according to claim 1, further comprising: a command generating unit configured to generate a correction command for the proposed revision based on correction information and reevaluation information for the proposed revision and transmit the generated correction command to the sentence generation model if the proposed revision does not satisfy a criterion.
8. The document correction program according to claim 7 , wherein the re-evaluation unit causes the suggestion unit and the instruction generation unit to repeatedly obtain re-evaluation information and send correction instructions until the evaluation of the correction proposal satisfies a preset criterion.
9. The reassessment information includes a risk level for each of the issues identified with respect to the proposed revision; 9. The document correction program of claim 8, wherein the criteria includes a risk level of all issues for the proposed correction being at or below a predetermined risk level.
10. 10. The document correction program according to claim 8, wherein the re-evaluation unit terminates the repeated re-evaluation and causes the suggestion unit to transmit the correction proposal information when a preset termination condition is met, even if the evaluation of the correction proposal does not satisfy a criterion.
11. 2. The document correction program according to claim 1, wherein the instruction generating section receives an input of a request for correction from a user, generates the correction instruction based on the request, and transmits the correction instruction to the text generation model.
12. The document is a legal document relating to an agreement with a specific party; The document correction program according to claim 1 , wherein the instruction generation unit acquires background information indicating at least one of a relationship with the other party, a type of the agreement, and requests of the parties regarding the agreement, generates the correction instruction based on the background information, and transmits the correction instruction to the text generation model.
13. The document correction program according to claim 1 , wherein the draft information is information about a draft generated by a model.
14. The document is a legal document relating to an agreement with a specific party; The document correction program of claim 13 , wherein the draft information is information about a draft generated by the model based on prior information indicating at least one of a relationship with a counterparty, a type of the agreement, and a request of the parties regarding the agreement.
15. 1. A document correction device for proposing corrections to a document based on an evaluation of the document by utilizing a sentence generation model, comprising: The apparatus includes an acquisition unit, an instruction generation unit, and a suggestion unit, The acquisition unit acquires draft information related to a draft of the document and evaluation information indicating an evaluation of the draft, the instruction generation unit generates a correction instruction for the draft based on the draft information and the evaluation information, and transmits the correction instruction to the writing generation model; The suggestion unit transmits correction suggestion information for correcting the draft based on an output of the sentence generation model.
16. 1. A method for document correction utilizing a sentence generation model to suggest corrections to a document based on evaluations of the document, comprising: The computer obtaining draft information regarding a draft of the document and evaluation information indicating an evaluation of the draft; generating revision instructions for the draft based on the draft information and the evaluation information and sending the revision instructions to the sentence generation model; and transmitting correction suggestion information for correcting the draft based on an output of the sentence generation model.
Citation Information
Patent Citations
Computer program, server device, terminal device and method
JP2022175871A
Contract review program, contract review device, and contract review method
WO2023021685A1
Legal document data correction method, legal document data correction system and legal document data correction program
JP2020080132A
Cited By
Information processing apparatus, information processing method, and program
JP2026032401A
system
JP2026155508A
Information processing device, information processing method, and program
JP7752890B1
