Apparatus, method, and program for constructing a database.

Abstract

We will build a new database for contracts. [Solution] First, the device 100 acquires contract data and metadata relating to a contract having one or more versions (S201). Next, the device 100 creates an instruction to generate an outline of the contract content based on the acquired contract data (S202), and makes a request to the generation AI model including this instruction (S203). The platform 120, upon receiving the request, generates the outline (S204) and transmits the outline to the device 100 (S205). The device 100 then requests the generation of one or more vectors representing all or part of the received outline (S206). The device 100, upon receiving the generated one or more vectors (S207), stores each vector in the database, associating it with the corresponding section of the outline (S208).

Description

[[Technical Field]]

[0001] The present invention relates to an apparatus, a method, and a program for constructing a database for contracts. [[Background Art]]

[0002] In a company, contract operations are in many cases carried out with the involvement of both the business division and the legal department. The business division holds discussions with the counterparty to the contract from a business perspective, and aligns understanding on major matters of the contract content. Under such understanding, one of the contracting parties drafts a contract, and the other reviews the draft. The company that receives the draft contract requests the legal department to examine the draft contract. [[Summary of the Invention]] [[Problem to be Solved by the Invention]]

[0003] When a business division prepares a draft contract, it often refers to similar contracts from the past. For example, even if the name of a contract is "outsourcing agreement" and the name is the same, if there are a large number of contracts with the same name, the burden of finding a contract whose detailed content is similar to that of the draft contract to be prepared is heavy. Also, when the legal department prepares or examines a draft contract, it may also refer to similar contracts from past contracts, but the burden of searching for a contract to be referred to is heavy. Such a burden results from the insufficient provision of databases for contracts that allow searching for highly relevant contracts.

[0004] The present invention has been made in view of such circumstances, and a first object of the present invention is to provide an apparatus, a method, or a program for constructing a new database for contracts.

[0005] A second object of the present invention is to provide various methods using the new database for contracts. [[Means for Solving the Problem]]

[0006] To achieve this objective, a first aspect of the present invention is a method for constructing a database, comprising the steps of: acquiring contract data relating to a contract having multiple versions; creating instructions to generate an outline of the contract content based on the contract data; making a request to a generation AI model including the instructions; receiving an outline generated by the generation AI model, which includes a summary of the contract content; making a request to an embedding model to generate one or more vectors representing part, all, or part and all of the outline; and storing each of the one or more vectors generated by the embedding model in the database in association with the corresponding outline portion.

[0007] Furthermore, a second aspect of the present invention is the method of the first aspect, wherein the summary further includes one or more descriptions relating to one or more provisions that have been changed between the multiple editions.

[0008] Furthermore, a third aspect of the present invention is the method of the first or second aspect, wherein the abstract is no more than a predetermined number of characters.

[0009] Furthermore, a fourth aspect of the present invention is the method of the third aspect, and the above abstract is no more than 3000 characters in Japanese.

[0010] Furthermore, a fifth aspect of the present invention is a method according to any of the first to fourth aspects, wherein the abstract includes the type of contract.

[0011] Furthermore, a sixth aspect of the present invention is a method according to any of the first to fifth aspects, wherein the storage is further associated with a contract identifier that identifies the contract.

[0012] Furthermore, a seventh aspect of the present invention is a method according to any of the first to sixth aspects, further comprising the step of obtaining metadata relating to the contract.

[0013] Furthermore, an eighth aspect of the present invention is the method of the seventh aspect, wherein the storage is carried out in further association with one or more items included in the metadata.

[0014] Furthermore, a ninth aspect of the present invention is the method of the eighth aspect, wherein the one or more items include the type of contract.

[0015] Furthermore, a tenth aspect of the present invention is a method according to any of the first to ninth aspects, wherein the one or more vectors include a single vector representing the summary.

[0016] Furthermore, an eleventh aspect of the present invention is the method of the tenth aspect, wherein the abstract is a summary of the terms of the contract of the last of the multiple versions.

[0017] Furthermore, a twelfth aspect of the present invention is the method of the eleventh aspect, where the last version is the final version.

[0018] Furthermore, a thirteenth aspect of the present invention is a method according to any of the first to twelfth aspects, wherein the one or more vectors include one or more vectors representing each of the one or more descriptions.

[0019] A fourteenth aspect of the present invention is a program for causing a computer to perform a method for constructing a database, the method comprising: acquiring contract data relating to a contract having multiple versions; creating instructions to generate an outline of the contract content based on the contract data; making a request to a generating AI model including the instructions; receiving an outline generated by the generating AI model, which includes a summary of the contract content; making a request to an embedding model to generate one or more vectors representing part, all, or part and all of the outline; and storing each of the one or more vectors generated by the embedding model in the database in association with the corresponding outline locations.

[0020] A fifteenth aspect of the present invention provides an apparatus for constructing a database, configured to: acquire contract data relating to a contract having a plurality of versions; create an instruction to generate an outline of the contract content according to the contract based on the contract data; send a request including the instruction to a generative AI model; receive an outline generated by the generative AI model, the outline including a summary of the contract content; send a request to an embedding model to generate one or more vectors representing part, all, or part and all of the outline; and store each of the one or more vectors generated by the embedding model in the database in association with a corresponding portion of the outline. [Effects of the Invention]

[0021] According to one aspect of the present invention, using contract data having one or more versions, an outline including a summary of the contract content and other desired items is generated using a generative AI model, and a vector for each desired portion of the outline is stored in a database in association with the corresponding portion. This allows the contract to be set as a search target for various purposes. [Brief Description of Drawings]

[0022] [Figure 1] FIG. 1 is a diagram showing a system according to a first embodiment of the present invention. [Figure 2] FIG. 2 is a diagram showing the flow of database construction according to the first embodiment of the present invention. [Figure 3] FIG. 3 is a diagram showing an example of contract data according to the first embodiment of the present invention. [Figure 4] FIG. 4 is a diagram showing an example of an instruction for generating an outline according to the first embodiment of the present invention. [Figure 5] FIG. 5 is a diagram showing a search flow according to a second embodiment of the present invention. [Mode for Carrying Out the Invention]

[0023] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[0024] (First Embodiment) FIG. 1 shows a system according to a first embodiment of the present invention. The apparatus 100 communicates with a user terminal 110 used by a user, a platform 120 that provides a generative AI model, a platform 130 that provides an embedding model, and a database 140 via an IP network such as the Internet to construct a database for contracts.

[0025] In FIG. 1, the generative AI model is provided by the platform 120, but an application for providing the generative AI model may be executed on the apparatus 100. Further, from the viewpoint of security, the apparatus 100 may communicate with the user terminal 110 via a closed network using the IP protocol or other protocols. In this case, it is preferable that the platform 120 is also accessible via the closed network, or the generative AI model is provided on the apparatus 100. The same applies to the platform 130 that provides the embedding model. Although the database 140 is illustrated as a separate apparatus from the platform 130 that provides the embedding model, they may be the same apparatus.

[0026] The device 100 comprises a communication unit 101 such as a communication interface, a processing unit 102 such as a processor or CPU, and a storage unit 103 including a storage device or storage medium such as memory or a hard disk. The device can be configured by executing a program for performing each process or operation in the processing unit 102. The device 100 may include one or more devices, computers, or servers. The program may also include one or more programs and can be recorded on a computer-readable storage medium to form a non-transient program product. The program is stored in the storage unit 103 or in a storage device or storage medium accessible via an IP network from the device 100, and instructions included in the program can be executed by at least one processor in the processing unit 102. Data described below as being stored in the storage unit 103 may be stored in the storage device or storage medium, and vice versa.

[0027] First, the device 100 acquires contract data relating to a contract having one or more versions (S201). This contract data can be received from the user terminal 110 and stored in the storage unit 103 in association with a contract identifier that identifies the contract. The contract data may further include one or more comments for any of the versions, and may be stored in association with a version identifier that identifies the version. The contract data may be stored as, for example, Markdown format data such as JSON format, but is not limited to this.

[0028] Furthermore, the device 100 may acquire metadata relating to the contract as needed (S201). The metadata may include information not included in the contents of the contract, and may include, for example, at least one of the following: the username or user identifier of the user who entered the contract data, the username or user identifier of the user who made changes and registered a new version, the username or user identifier of the user who added comments, the registration date or registration date and time for each version, and the storage location such as the directory or path in which each version is stored. The metadata may also include information included in the contents of the contract, and may include, for example, at least one of the following: the name of the contract such as "Outsourcing Agreement", the contracting party, the contract start date, the contract end date, the signing date, whether or not it is automatically renewed, and whether or not there is an anti-social clause.

[0029] Figure 3 shows an example of contract data according to the first embodiment of the present invention. In actual data, specific text is entered in the blank spaces "". This contract data includes versions 1 through 3, with comments added to versions 2 and 3. Although the example in Figure 3 is a simple example, even in more complex cases, the background and history of clause changes can be described along with the content of the contract. In this example, the contract name, which was explained above as an example of metadata, is included. In this way, some or all of the metadata can be described within the contract data. Also, although this example was explained as including the full text of all versions, it is not necessary to include the full text of all versions once the content of each version of the contract has been determined.

[0030] Furthermore, the metadata may include the type of contract. This type may be identified by the device 100 based on the contents of the contract, or it may be set based on input from the user terminal 110. Such input can be done by text input, selection from a pull-down menu, etc. When identifying based on the contents of the contract, this may be done using an AI model or a generative AI model.

[0031] The type input screen display information can be sent, for example, as an HTML file and loaded by the web browser of the user terminal 110, thereby displaying the type input screen on the user terminal 110's display screen. If a dedicated application is installed on the user terminal 110, the type input screen display information should include the data necessary for that application to display the input screen. Here, "input screen" can take various forms such as a web page, modal window, or popup window when displayed on a web browser, or it can be a screen of a dedicated application when displayed on that application. In any case, any screen that includes an area with input fields for entering the necessary information is considered an input screen. The same technology can be applied to other screens mentioned in this specification.

[0032] The comments mentioned above can also be explained as being included in metadata rather than contract data. More generally, while the above explanation distinguishes between contract data and metadata, if data corresponding to “contract data” and “metadata” as defined herein is included, whether this data exists as separate data or as a single data in any method, program, terminal, device, server, or system (hereinafter referred to as “method, etc.”) does not negate its applicability to these terms.

[0033] Next, the device 100 creates an instruction (prompt) to generate an outline of the contract contents based on the acquired contract data (S202).

[0034] Figure 4 shows an example of instructions according to the first embodiment of the present invention. The instructions describe generating an overview, including a summary, category, and change history, based on contract data given in JSON format. The instructions further describe optimizing the overview for vectorization for vector search and outputting it in Markdown format. Some or all of the instructions can be modified to set values ​​associated with the user or to change descriptions associated with the user by logging in with credentials.

[0035] The instructions allow you to set contract data as a variable. This example shows how to set data in JSON format, but other Markdown formats are also acceptable, and it is not limited to Markdown format as long as it contains the necessary data.

[0036] The instructions also include examples of output formats and descriptions of summaries, classifications, and revision history. The summary may refer to the type, amount, and duration of the contract, and may include other items, or only some of them. The summary is preferably a summary of the final version or other last version, but summaries of earlier versions may be generated additionally or alternatively.

[0037] The classifications are for distinguishing contract content according to predetermined criteria and may include at least one of quantitative classifications and qualitative classifications based on the nature of one or more transactions included in the contract content. Examples of quantitative classifications include amounts and periods. Examples of qualitative classifications include whether or not it is a transaction between multiple parties, and whether or not it includes specific clauses.

[0038] The categories may be generated separately from the summary, or they may be included as part of the summary. Furthermore, the types described as being included in the summary may be generated separately from the summary. Also, the types may be replaced by qualitative categories, or they may be generated separately from qualitative categories. Additionally, categories separate from the summary may not be generated at all.

[0039] Regarding the history of changes, for each change, the changes and their history between earlier and later versions may be included. Instead of a history of changes, instructions may be included to describe the changed clause as a change, and one or more such changes may be generated. More generally, the summary may include one or more descriptions of one or more clauses that have been changed between versions. Alternatively, no such descriptions may be generated.

[0040] Device 100 makes a request to the generating AI model that includes the said instructions (S203). The request to the generating AI model can be made by executing code that includes instructions written in natural language and calling the OpenAI API, and the generating AI model generates a response to the prompt. The OpenAI API is an example, and other APIs may be used. More specifically, the code for making the request is stored in the memory of the device on which the code is executed, and the device can make the request by retrieving this code and executing code that includes instructions obtained by setting the required values ​​to the variables contained in the code.

[0041] Upon receiving the request, platform 120 generates a summary (S204) and transmits the summary to device 100 (S205). The summary received by device 100 should consist of a summary and other desired items as described with reference to Figure 4, and the extent to which these are explicitly described in the instructions created by device 100 can be determined as appropriate. For example, although not explicitly shown in Figure 4, the generated summary may include a list of keywords, and for this purpose, the output format of the instructions for generating the summary may include a description of the list of keywords, if necessary.

[0042] The generated summary or abstract contained therein is preferably no more than a predetermined number of characters, such as 3,000 characters. By keeping it no more than the predetermined number of characters, the main points of the transaction conducted under the contract are described, rather than the minute details of the contract's contents, thereby effectively characterizing the content of the contract.

[0043] Furthermore, by including a list of keywords in the generated summary, the content of the contract can be characterized from multiple perspectives.

[0044] Furthermore, by including one or more descriptions of the modified clauses in the generated summary, it becomes possible to identify the clauses that were discussed during the contract negotiation process, thus characterizing the contract as one that contains clauses likely to require discussion within the user's company.

[0045] The device 100 then requests the generation of one or more vectors representing all or part of the received summary (S206). This request may be made for different platforms 130 depending on the embedding model used, or it may include a specification of the embedding model to be used even for the same platform 130.

[0046] The request may, for example, include the generation of a single vector representing the entire summary. Alternatively, the request may, for example, include the generation of a single vector representing a part of the summary. Examples of such a part include the entire or partial summary, the entire or partial description of one or more amended clauses, the entire or partial classification, and the entire or partial list of keywords. Alternatively, multiple vectors representing the entire summary may be generated, or multiple vectors representing parts of the summary may be generated. Examples of such parts include the entire or partial summary, the entire or partial description of one or more amended clauses, the entire or partial classification, and the entire or partial list of keywords.

[0047] Upon receiving one or more generated vectors (S207), the device 100 stores each of those vectors in the database, associating them with the corresponding summary section (S208). The device 100 may further associate the stored vectors with at least one of the contract identifiers and one or more items included in the metadata. The device 100 can receive input from the user regarding what vectors to generate, store, or enable, and can pre-store conditions for each user or group to which the user belongs in the storage unit 103.

[0048] In this way, by using contract data with one or more versions, an AI model is used to generate an overview of the contract, including a summary of the contract's content and other desired items. By associating the vectors of each desired section of the overview with the corresponding sections and storing them in a database, the contract can be searched for various purposes. This significantly streamlines the process of referencing highly relevant contracts, which previously had to be done manually.

[0049] In this specification, "AI model" refers to a machine learning model that has been trained to predict an output for a given input, and "generative AI model" refers to a large-scale language model (LLM) that has been trained using text data to generate an output not included in the input for a given input. As a generative AI model, an LLM applying a transformer architecture is particularly preferred, but it is expected that the name of the architecture may change as technology advances. Therefore, in this specification, "transformer architecture" includes architectures that use one or more features of a transformer architecture or improvements thereof. In this specification, whether or not "generative AI models" are the same is determined by whether or not the type of generative AI model specified by the user is the same. In the case of the Open AI API, for GPT-4, if the value of the variable "model" is the same, it is expressed as the same generative AI model. Generative AI models used in different processes can be the same, but if they are not the same, they may be provided on the same platform 120 or device 100, or on different devices. Furthermore, it goes without saying that each request for a generative AI model may include multiple requests, and may also include one or more processes performed on the device making the request other than the requests for the generative AI model.

[0050] (Second embodiment) A second embodiment of the present invention relates to a search method using a new database for contracts as described in the first embodiment. The device 100 receives a draft contract from a user terminal 110 (S501), generates one or more vectors corresponding to the draft contract or an outline of the draft contract (S502), and inputs the generated vectors into the database 140 (S503). This allows the device to search for one or more past contracts that have a high relevance or similarity to the draft contract and transmit the results to the user terminal 110 (S504). The outline includes a summary of the draft contract. The outline may also include, for example, one or more keywords or at least some of their synonyms or related terms contained in the content of the draft contract.

[0051] In this specification, the term "draft contract" is not necessarily limited to a document in the form of a contract.

[0052] Furthermore, in the embodiments described above, unless the word "only" is used, such as "based only," "depending only," "in the case of only," or "referencing only," it is assumed in this specification that additional information may also be considered. Also, as an example, the statement "if a, then b" does not necessarily mean "always b in the case of a" or "b immediately after a," unless explicitly stated otherwise. In addition, the statement "each a constituting A" does not necessarily mean that A is composed of multiple components, but includes the possibility that the component is singular.

[0053] Furthermore, it should be noted that the embodiments of the present invention described above are included in the disclosure herein, in any way that they are not inconsistent with each other.

[0054] Furthermore, for the sake of clarity, even if there are aspects of the invention that perform actions different from those described herein by any means, each aspect of the invention is intended to address the same actions as any of those described herein, and the existence of actions different from those described herein does not mean that such methods are outside the scope of the respective aspects of the invention.

[0055] Furthermore, the "start" and "end" shown in Figure 5 are merely examples and do not necessarily mean that the method according to this embodiment will always start or end in the illustrated procedure. [Explanation of symbols]

[0056] 100 devices 101 Communications Department 102 Processing Unit 103 Storage section 110 user terminals A platform that provides 120 generated AI models. Platform providing 130 embedded models 140 Databases

Claims

1. A method for building a database, The computer obtains contract data relating to a contract that has multiple versions, The steps include: the computer creating instructions to generate an outline of the contract based on the contract data; The computer makes a request to the generated AI model, including the instructions, The computer receives a summary generated by the generative AI model, which includes a summary of the contract content and one or more descriptions relating to one or more clauses that have been changed between the multiple versions. The computer makes a request to the embedding model to generate one or more vectors representing part, all, or part and all of the outline, The computer stores each of the one or more vectors generated by the embedding model in the database, associating them with the corresponding outline location. Includes.

2. The method according to claim 1, The above summary must be below the specified character limit.

3. A method according to claim 1 or 2, The above summary includes the type of contract.

4. A method according to claim 1 or 2, The process further includes the step of obtaining metadata relating to the aforementioned contract, The aforementioned storage is performed in association with one or more items included in the metadata.

5. The method according to claim 4, The one or more items mentioned above include the type of contract.

6. A method according to claim 1 or 2, The one or more vectors mentioned above include a single vector representing the summary.

7. The method according to claim 6, The above summary is a summary of the terms of the contract as of the last of the multiple versions mentioned above.

8. A program for causing a computer to perform a method for building a database, wherein the method is Steps to obtain contract data for contracts that have multiple versions, A step of creating instructions to generate an outline of the contract content based on the aforementioned contract data, The steps include making a request to the generated AI model, including the aforementioned instructions, The steps include receiving a summary generated by the generation AI model, which includes a summary of the contract content and one or more descriptions relating to one or more clauses that have been changed between the multiple versions, The steps include: requesting the embedding model to generate one or more vectors representing part, all, or part and all of the above outline; The steps include storing each of the one or more vectors generated by the embedding model in the database, associated with the corresponding section of the summary; Includes.

9. A device for building a database, We obtain contract data for contracts that have multiple versions. Based on the aforementioned contract data, create instructions to generate an outline of the contract content according to the aforementioned contract, and make a request to the generating AI model including the aforementioned instructions. The AI ​​model generates an overview, which includes a summary of the contract and one or more descriptions of one or more clauses that have been changed between the multiple versions. The embedding model is requested to generate one or more vectors that represent part, all, or part and all of the above outline. Each of the one or more vectors generated by the embedding model is configured to be stored in the database in association with the corresponding section of the summary.

Images