How to revise the text of public relations information

Abstract

This invention provides a method and information processing system for quickly and accurately correcting manuscript text for appropriate expression using a large-scale language model. [Solution] The method includes the steps of: step 1, the operator inputs the latest expression guidelines (GL) as training data; step 2, the operator inputs the manuscript text into a large-scale language model (LLM) and instructs it to revise the expression; step 3, the LLM displays the revised text T based on the training data; step 4, the operator instructs further revisions if they deem the revised text T inappropriate; step 5, the LLM generates and displays revised text for the instructed revisions; step 6, the LLM updates the expression GL including information about the revised text that was instructed and generated; and step 7, the process of repeating steps 4' to 6' until the revised text T is deemed appropriate. The manuscript text, revised text, and revised text are displayed separated into sentence components specified for each sentence, and their appropriateness is judged.

Description

Technical Field

[0001] The present invention relates to a method and an information processing system for correcting a text expression manuscript of public relations information of a business entity into an expression that can be evaluated as appropriate by the business entity using a large language model.

Background Art

[0002] For business entities, it is an essential task to internally and / or externally publicize and appeal business information related to their target business fields. In many cases, they express the business information to be publicized (hereinafter referred to as "public relations information") by combining text, figures, images, videos, audio, etc., and display it on paper media, electronic media or the like to explain it to business targets.

[0003] The invention described in Patent Document 1 as the conventional art will be described by taking as an example a case where Company A, a business entity belonging to the field of providing managed services (a method of outsourcing the responsibility for process and function maintenance and demand forecasting to improve operational management and reduce costs), causes the employee a in charge, who is a worker of Company A, to create the manuscript shown in Figure 1 as public relations information in order to explain its business content.

[0004] Main text information was extracted from the manuscript shown in Figure 1 and organized into the table in Figure 2.

[0005] Like many business entities that handle public relations information, Company A conducts a preliminary internal review from various perspectives to determine whether the content and expression of the manuscript created by the employee a in charge are appropriate when explaining to business targets.

[0006] In particular, text expressions are carefully pre-reviewed, as in order to achieve appropriate communication with business targets, they are required to be socially appropriate from the perspective of social trends such as legal, cultural, domestic and foreign economic situations, and the business entity's own cultural trends, and to have appeal to business targets from the viewpoint of the business entity.

[0007] In the past, when AI technology was underdeveloped, the pre-screening process relied almost entirely on manual labor. Social trends, business culture trends, and other factors were addressed by a vast and diverse set of expression guidelines created by numerous internal and external experts in each trend theme.

[0008] However, with the rapid development of social media and AI technology in recent years, the traditional manual approach is no longer sufficient to cope with the expansion of trend information on each trend theme and the diversification of business targets. As a result, the development of expression guidelines and pre-screening processes are no longer proceeding smoothly and quickly.

[0009] On the other hand, with the rapid development of social media and AI technology in recent years, large-scale language models (LLMs), which are generative AIs that correspond to text, have been introduced, particularly in the development of expression guidelines and pre-screening of text. This has led to the smoother and faster development of expression guidelines and pre-screening, without relying on manual methods.

[0010] For example, Patent Document 1 discloses the invention described below.

[0011] ■Invention described in Patent Document 1■ Steps include receiving questions entered by the user, A large-scale pre-trained language model generates a set of candidate answers to the aforementioned questions. Select an answer from the aforementioned set of candidate answers as the target answer, The steps include displaying the target answer to the user, Upon receiving a feedback request submitted by the user regarding the target response, Generate a feedback page, A step of displaying a feedback page to the user, The content of the aforementioned feedback page includes the aforementioned set of candidate answers, steps, Upon receiving the update request submitted by the user to the feedback page, The answer indicated by the update request in the aforementioned set of candidate answers is identified as the new target answer. The steps include displaying a new target answer to the aforementioned user, A response feedback method to be applied to large-scale language models, including those that include this feature.

[0012] Here, the large-scale language model in the disclosed invention of Patent Document 1 is loaded with the expression guidelines used by Company A for pre-examination as training data, and the large-scale language model is then configured to perform the following: When employee A enters the question, "Revise the expression of the manuscript text information created by employee A based on the training data," The large-scale language model that received the question said: Candidate answers to the question are generated, and the target answer is selected from the candidate answers and displayed, for example, as shown in Figure 3-1 (Reference 1, paragraph 0027, Reference 1, Figure 3A). When employee A believes that some modification is needed to the target answer and requests an update from the large-scale language model, all newly generated candidate answers in response to employee A's update request are displayed, for example, as shown in Figure 3-2, and employee A selects a satisfactory answer from among them (Reference 1, paragraph 0039, Reference 1, Figures 3B-C). [Prior art documents] [Patent Documents]

[0013] [Patent Document 1] Japanese Patent Publication No. 2024-178148 [Overview of the project] [Problems that the invention aims to solve]

[0014] However, as shown in Figures 3-1 and 3-2, if the text expressions extracted from the manuscript, the target answer, and the multiple candidate answers provided as feedback are displayed as text, employee A must visually scan each sentence from the first line, compare the text expressions extracted from the manuscript with the target answer or the multiple candidate answers, and judge the appropriateness of the target answer or each candidate answer. This increases the workload of correcting the text expressions in the manuscript, which impairs the speed and accuracy of the correction work.

[0015] The present invention aims to provide a method for revising the text expression of public relations information and an information processing system that can quickly and accurately revise the draft text expression of public relations information of a business entity to an expression that the business entity deems appropriate, using a large-scale language model. [Means for solving the problem]

[0016] The present invention [1] The following steps are taken to revise the draft text of the public relations information of the business entity to an expression that the business entity deems appropriate, using a large-scale language model: Step 1: The latest expression guideline information regarding the manuscript text expression, which is the text expression of the aforementioned manuscript, is input into the large-scale language model as training data and stored in an information storage medium accessible to the large-scale language model; Step 2: The operator inputs the manuscript text expression into the large-scale language model and instructs it to revise the manuscript text expression; Step 3: The large-scale language model modifies the manuscript text expression entered in step 2 based on the training data stored in the information storage medium, and displays the modified text expression to the operator. With respect to the modified text expression displayed in step 3, If the worker determines it is appropriate, the worker will not instruct the large-scale language model to make any modifications, or will instruct not to do so. Step 4: If the operator determines that the text is inappropriate, the operator instructs the large-scale language model to revise the modified text representation; Step 5, in which the large language model generates and displays a re-revised text representation obtained by revising the revised text representation based on the revision instructed in Step 4 and the training data, and responds to the worker; The large language model, at least information related to the instruction by the worker in Step 4, and information related to the re-revised text representation generated in Step 5, newly generates expression guideline information to update the expression guideline, as Step 6; For the re-revised text representation, when the worker determines that it is appropriate, the worker does not instruct the large language model to make revisions, or instructs that no revision is needed, when the worker determines that it is inappropriate, the worker instructs the large language model to revise the re-revised text representation, as Step 4'; Step 5', in which the large language model generates and displays a re-revised text representation obtained by revising the re-revised text representation based on the revision instructed in Step 4' and the training data, and provides response feedback to the worker; The large language model, at least information related to the instruction by the worker in Step 4', and information related to the re-revised text representation generated in Step 5', newly generates expression guideline information to update the expression guideline, as Step 6'; and Step 7 of repeating the Step 4', Step 5' and Step 6' until the worker determines that the corresponding re-revised text representation is appropriate in Step 4', the method being a text representation revision method for public relations information, In Step 3, the large language model, to the worker, displays the original text representation by dividing it into sentence constituent elements specified by the worker on a sentence-by-sentence basis, The generated revised text and the further revised text are displayed sentence by sentence, separated into the constituent elements of the sentence. In step 4 and step 4', The worker shall determine whether the revised text expression and the further revised text are appropriate or inappropriate for each component of the sentence, The components of the aforementioned sentence that are deemed appropriate will not be instructed to be modified by the large-scale language model. The large-scale language model is instructed to modify the components of the aforementioned sentence that it has determined to be inappropriate. In step 5 and step 5', The large-scale language system displays the revised text representation for each component of the sentence, A method for modifying the textual expression of public relations information to be used as a response to the aforementioned worker (hereinafter also referred to as "Invention 1"), and [2] At least, An information display medium that displays the original text expression, revised text expression and further revised text expression in steps 1 to 6, 4' to 6' and 7 described in the preceding paragraph [1], An information storage medium that stores the expression guidelines in steps 1 to 6, 4' to 6' and 7 described in the preceding paragraph [1], The present invention relates to an information processing system for modifying public relations information (hereinafter also referred to as "the present invention 2"), comprising a large-scale language model that performs steps 1 to 6, 4' to 6' and 7 described in the preceding paragraph [1].

[0017] Hereinafter, "Invention 1" and "Invention 2" will be collectively referred to as "the present invention." [Effects of the Invention]

[0018] According to the present invention, it is possible to provide a method for correcting the text expression of public relations information and an information processing system that can quickly and accurately correct the draft text expression of public relations information of a business entity to an expression that the business entity can evaluate as appropriate, using a large-scale language model. [Brief explanation of the drawing]

[0019] [Figure 1] This is an example of a draft of public relations information created by an operator in this invention.

[0020] [Figure 2] This is an example of textual expression extracted from the draft of the public relations information shown in Figure 1.

[0021] [Figure 3-1] This is an example of a target response displayed in response to instructions for revision of text expressions extracted from a public relations information draft by an operator, according to the invention disclosed in Patent Document 1.

[0022] [Figure 3-2] This is an example of multiple target responses displayed as feedback after receiving instructions from an operator to revise their target response, according to the invention disclosed in Patent Document 1.

[0023] [Figure 4-1] This is an example of a method for displaying text information extracted from a manuscript, sentence by sentence.

[0024] [Figure 4-2] This is an example of a method for displaying text information extracted from a manuscript, divided sentence by sentence into sentence components specified by the operator. (a) An example where each component is displayed using only text information, and clicking on a component instructs the large-scale language system to modify that component. (b) An example where buttons are placed on each component to instruct the large-scale language system to mark it as appropriate (OK) or inappropriate (NG). (c) An example where buttons are placed on each component to instruct the large-scale language system to mark it as appropriate (check icon (1)) or inappropriate (trash can icon (2)).

[0025] [Figure 5-1] This is an example of an embodiment in which, for one of the sentences that make up the text information extracted from the manuscript, the operator specifies the constituent elements of the sentence and displays them, while keeping a record of whether the manuscript text information and the corrected text information are appropriate (check icon (1)) or inappropriate (trash can icon (2)).

[0026] [Figure 5-2] This is an example of an embodiment in which, for one of the sentences that make up the text information extracted from the manuscript, the operator specifies the constituent elements of the sentence and displays them, while keeping a record of whether the manuscript text information and the corrected text information are appropriate (check icon (1)) or inappropriate (trash can icon (2)).

[0027] [Figure 6] This is an example of a system flow performed by the present invention 2. [Modes for carrying out the invention]

[0028] Invention 1 [Business entity and worker in Invention 1] In Invention 1, "business entity" refers to any entity that conducts activities (hereinafter referred to as "business") to achieve its business objectives, regardless of whether it is for profit or non-profit, regardless of its organizational form, such as an individual business or a corporate organization, and whose business objective is to handle goods and / or services, and whose essential business is to publicize and promote business information internally and / or externally in order to achieve its business objectives.

[0029] In Invention 1, a worker belonging to an entity is a person responsible for the public relations activities of the entity's business information, and may be a department or individual that performs the public relations activities of the entity, a team or individual that constitutes the public relations department (for example, a company employee in charge), a contractor or agent for the public relations activities of the entity, or, in the case of a sole proprietor, the sole proprietor.

[0030] In the present invention 1, to the extent that the effects of the present invention 1 are not impaired, a large-scale language system and / or other generative AI may perform or support the worker's work under the worker's supervision.

[0031] The worker is responsible for modifying the draft text of the public relations information using a large-scale language model to an expression that the business entity deems appropriate. Since the information regarding these modifications is added to the expression guidelines, which constitute the training data used by the large-scale language model, and thus forms the latest expression guidelines, it is preferable that the worker possesses a certain level of knowledge regarding the field to which the public relations information belongs, the modification work performed by the large-scale language model, and the content of the expression guidelines.

[0032] [Structure of public relations information in Invention 1] The public relations information in this invention is expressed by combining text, figures, images, videos, audio, etc., and is conveyed to business targets in a manner that can be viewed by displaying it on paper media, electronic media, etc., and can be stored as electronic data in a storage medium with flexible input and output capabilities.

[0033] The present invention applies to text information expressed by text.

[0034] The text information to which this invention applies consists of units of sentences (in Japanese, units of text separated by periods, and in English, "Sentence"), and may consist of one sentence or multiple sentences.

[0035] The object of this invention is to enable workers in a business entity to quickly and accurately interpret text information. The number of characters per sentence is, In the case of Japanese, 10 to 1000 characters is preferred, 20 to 500 characters is more preferred, 30 to 300 characters is even more preferred, 40 to 200 characters is even more preferred, and 70 to 150 characters is even more preferred. In the case of English, 20-300 words is preferred, 30-150 words is more preferred, and 40-80 words is even more preferred.

[0036] [Large-scale language model in Invention 1] Large-scale language models (LLMs) are generative AI models that are expected to continue improving in functionality as text-based AI. They are constructed using large amounts of data and deep learning techniques, and are modeled using the probability of occurrence of sentences and words. By providing them with a large amount of text data, they can learn context and the meaning of words, and from the results of this learning, they can generate new text data by arranging words and sentences that are considered to have a high probability of following a particular word ("What is a Large-Scale Language Model (LLM)? Mechanism, Types, and Applications" (https: / / www.hitachi-solutions-create.co.jp / column / technology / llm.html)).

[0037] Examples of large-scale language models that can be used for the present invention 1 include OpenAI's GPT4, preferably OpenAI's latest model, and models provided by Amazon, Meta, Google, etc.

[0038] In this invention, the text information is displayed to the operator in a written language such as Japanese. However, in order to display it in a written language on the internet, such as on the Web, it is preferable to convert the text information in a written language into a language that is easy for a large-scale language model to process, such as by converting it into a markup language such as HTML.

[0039] [Training data in Invention 1] In Invention 1, the latest expression guideline information regarding the expression of manuscript text scheduled for pre-examination is input into a large-scale language model as training data.

[0040] Since Invention 1 targets public relations information in diverse business fields, expression guidelines have been established for each business field.

[0041] The expression guidelines not only explain and illustrate appropriate and up-to-date terminology, phrases, and other expressions used in the business field to which the content of the manuscript belongs, but also organize the background of that business field, such as the historical changes in expression, and expression know-how such as taboo words and their alternatives in that business field.

[0042] The public relations information in this invention is expected to be displayed on the internet and disseminated worldwide. For example, expressions used in the relevant business field may be organized into a so-called whitelist (a list defining safe expressions) and / or blacklist (a list defining dangerous expressions). A large-scale language model may then be allowed to access or execute only expressions on the whitelist, prohibit access or execution of expressions on the blacklist, or replace expressions on the blacklist with expressions on the whitelist.

[0043] When the content of a previously printed document is digitized and used for the first time as a representation guideline for a large-scale language model, that becomes the latest representation guideline at that time. Subsequently, if new information is added to the representation guideline, the representation guideline at the time of the last addition becomes the latest representation guideline. [Steps 1-7 in Invention 1] Invention 1 includes the following steps 1-6, 4'-6', and 7, which modify the draft text of a public relations information of a business entity to an expression that the business entity can deem appropriate, using a large-scale language model.

[0044] (Step 1) This step involves inputting the latest expression guideline information regarding manuscript text expression into a large-scale language model as training data, and storing it in an information storage medium accessible to the large-scale language model.

[0045] The information storage medium may be a database built into or externally installed in a device that constitutes a large-scale language model, or it may be an external cloud.

[0046] As mentioned above, expressions used in the relevant business field may be organized into a so-called whitelist (a list defining safe expressions) and / or blacklist (a list defining dangerous expressions) using, for example, a CSV file, and added to the training list.

[0047] Additionally, prompts for instructing large-scale language models may be created in YAML format beforehand and added to the teacher list.

[0048] (Step 2) This step involves the operator inputting the manuscript text into a large-scale language model and instructing it to revise the manuscript text.

[0049] When inputting manuscript text into a large-scale language model, it is preferable to convert it into a language format that is easy for the large-scale language model to process, such as converting it into a markup language like HTML.

[0050] (Step 3) Step 2 involves a large-scale language model correcting the manuscript text expression entered in Step 2 based on training data stored in the information storage medium, and then displaying the corrected text expression to the operator (for example, as shown in Figure 4-2(a)).

[0051] (Step 4) For the corrected text expression displayed in Step 3, If the worker deems it appropriate, the worker will not instruct the large-scale language model to make modifications. If the operator determines that the text is inappropriate, the operator instructs the large-scale language model to correct the modified text expression (for example, the instruction is displayed as shown in Figure 4-2(a)).

[0052] Furthermore, if the worker deems it appropriate, they may instruct the large-scale language model not to make any modifications, rather than instructing it to do so.

[0053] (Step 5) This step involves a large-scale language model generating and displaying a revised text representation (for example, as shown in Figure 4-2(b)) based on the corrections instructed in step 4 and the training data, and then providing the response to the operator.

[0054] (Step 6) Large-scale language models, At a minimum, information regarding the worker's instructions in step 4, To include information about the revised text representation generated in Step 5, This step involves generating new expression guideline information and updating the existing expression guidelines.

[0055] Information regarding instructions given by the worker in Step 4 includes, for example, the expression the worker instructed to be corrected in Step 4, the corrected expression, and the reason for the correction. Information regarding the revised text expressions generated in Step 5 includes, for example, keeping the expressions instructed in Step 4 and the revised text expressions as a set.

[0056] Step 6 may include information deemed necessary for the latest expression guidelines, in addition to information regarding worker instructions in Step 4 and information regarding the revised text expression generated in Step 5.

[0057] (Step 4') For revised text expressions, If the worker deems it appropriate, the worker will not instruct the large-scale language model to make modifications. If the worker determines that the text is inappropriate, the worker instructs the large-scale language model to revise the corrected text representation.

[0058] Furthermore, if the worker deems it appropriate, they may instruct the large-scale language model not to make any modifications, rather than instructing it to do so.

[0059] (Step 5') This step involves a large-scale language model generating and displaying a revised text representation based on the corrections instructed in step 4' and the training data, and then responding to the operator (for example, as shown in Figure 5-2(b)).

[0060] (Step 6') Large-scale language models, At a minimum, information regarding the worker's instructions in step 4', To include information about the revised text expression generated in step 5', This step involves generating new expression guideline information and updating the existing expression guidelines.

[0061] Information regarding the worker's instructions in step 4' includes, for example, the expression the worker instructed to be corrected in step 4', the corrected expression, and the reason for the correction.

[0062] Information regarding the revised text expression generated in Step 5' includes, for example, keeping the expression instructed in Step 4' and the revised text expression as a set.

[0063] Step 6' may include information deemed necessary for the latest expression guidelines, in addition to information regarding the worker's instructions in Step 4' and the revised text expression generated in Step 5'.

[0064] Furthermore, regarding the instructions given by the worker in steps 4 and 4', in order to prevent variations in the evaluation of "appropriate" and "inappropriate" by different workers, the large-scale language model may be given an inference function regarding appropriate instructions (for example, by introducing temperature parameters, etc.), which takes into account not only the learning function through the accumulation of training data for the expression guidelines in steps 6 and 6', but also information on the worker's instructions when other manuscript text expressions were modified.

[0065] (Step 7) Steps 4', 5', and 6' are performed as follows: Step 4' is repeated until the operator determines that the revised text expression is appropriate (for example, until it matches the representation of the revised text manuscript in Figures 5-1 and 5-2).

[0066] [Display configuration according to Invention 1] In the method of the present invention, In step 3, a large-scale language model is used by the worker. The manuscript text is displayed sentence by sentence, separated into the sentence components specified by the worker, The generated corrected text and the further corrected text are displayed sentence by sentence, separated into sentence components. In Step 4 and Step 4', The worker will determine whether the revised text expression and the further revised text are appropriate or inappropriate for each sentence component, The large-scale language model will not be instructed to modify the sentence components that it deems appropriate. For sentence elements deemed inappropriate, the large-scale language model is instructed to make corrections and feedback is requested. In Step 5 and Step 5', The revised text is displayed for each sentence component, and the worker receives feedback on their response.

[0067] (Display method in Step 3) In the method of the present invention, in step 3, a large-scale language model is provided to the operator, The manuscript text is displayed sentence by sentence, separated into the sentence components specified by the worker.

[0068] For example, employee A of Company A created the draft shown in Figure 1 as public relations information, In Step 2, we assume that the textual representations (sentences consisting of two sentences) of the main textual information organized in the table in Figure 2 were input into the large-scale language model.

[0069] For example, in the invention disclosed in Patent Document 1, as shown in Figure 3-1, the entered text expression (a sentence consisting of two sentences) is displayed to employee a (as shown in Figure 3-2), In the present invention 1, as shown in the right column of Figure 4-1, each sentence is divided into two parts ("Sentence 1" and "Sentence 2") and displayed to employee a.

[0070] In the present invention 1, it may be further specified that each sentence be divided and displayed into sentence components specified by the operator, as shown in Figures 4-2, 5-1, and 5-2, so that the operator can easily compare the original text expression, the revised text expression, and the further revised text expression for each sentence component.

[0071] For example, for each sentence in the original text and the revised text, If we consider a sentence according to the grammar of a written language, and break it down into its constituent elements, in the case of Japanese, As constituent elements of a sentence, it is acceptable to divide it into meaningful phrases or into words, which are the smallest units of a sentence, so that the constituent elements of the sentence in the original text and the revised text can be compared.

[0072] For example, when a sentence in Japanese, Chinese, or other languages ​​is expressed using ideograms, Treating each sentence in the original text and the revised text as a simple string of characters, Each string can be divided into two or more characters that are common to all of the constituent characters (hereinafter referred to as "common strings") and other characters (hereinafter referred to as "different strings"), and the common strings and different strings can be compared as constituent elements of a sentence.

[0073] Furthermore, if the common string and the differing string include symbols that are used in pairs, such as "?", "!", "*", ";", ":", and " / ", or parentheses such as () and "", these symbols may be displayed separately (as shown in Figure 4-2).

[0074] Furthermore, one way to display the text by dividing it into its constituent elements as specified by the operator is: We take into account the information obtained by morphological analysis of the part-of-speech information of each corrected section, You may also include the sentences before and after the sentence to be corrected in the input, and consider the context and formatting information to create a structure that incorporates consistency in expression.

[0075] In the embodiment of the present invention 1, as shown in Figure 4-2, each sentence is divided into common strings and differing strings and displayed to employee a.

[0076] (Display format in Step 4 and Step 4') In the method of the present invention, in steps 4 and 4', The worker judges whether the revised text expression and the further revised text are appropriate or inappropriate for each sentence component (as shown in Figure 4-2), The large-scale language model will not be instructed to modify the sentence components that it deems appropriate. For sentence elements deemed inappropriate, the system instructs a large-scale language model to make corrections and requests feedback.

[0077] Furthermore, the worker may judge whether the expression in the manuscript text is appropriate or inappropriate for each component of the sentence (as shown in Figures 4-2(b) and (c)).

[0078] If a worker determines that a particular element of a revised text expression or a sentence in the revised text is inappropriate, for example,

[0079] (Example 1) (For example, as shown in Figure 4-2(a)) By clicking on the constituent elements of the specific sentence,

[0080] (Example 2) (For example, as shown in Figure 4-2(b)) Place OK or NG buttons adjacent to the components of the specific sentence, and click them (for example, Figure 4-2(b)), or,

[0081] (Example 3) (For example, as shown in Figure 4-2(c)) Place a trash can icon next to the element of the specific sentence, and click it. You could also instruct the large-scale language model to make corrections and request feedback, Along with OK and NG buttons, or a trash can icon (2), a "check icon (1)" may be placed (for example, as in Figure 4-2(c)) that can be clicked when a specific component is deemed appropriate.

[0082] In the present invention 1, the large-scale language model can not only be instructed to judge whether a particular component of a sentence is appropriate or inappropriate, but it can also be instructed by inputting a prompt that expresses the specific content of the correction in text.

[0083] For example, employee A can input the following prompt into a large-scale language model, include it in the expression guidelines, and generate new expression guidelines.

[0084] ■Example of a prompt■ The following are examples of inappropriate actions after the corrections: Please do not revise it, as it may change the meaning. ##Example of inappropriate content after correction## • True one-stop service → Consistent

[0085] [Examples] Company A, a business entity belonging to the field of managed services, instructed its employee, 담당자 'a', to create the draft shown in Figure 1 as public relations information in order to explain its business activities.

[0086] Employee A extracted the main textual expressions from the manuscript shown in Figure 1 and organized them in the table in Figure 2. Using a large-scale language model that implements the method of the present invention, employee A performed the task of revising the text to an expression that Company A would deem appropriate through a preliminary review.

[0087] This embodiment was carried out under the following conditions. (1) As a large-scale language model, we used OpenAI's GPT4. (2) Text information was converted into HTML, a markup language, and input into a large-scale language model. (3) In advance, for large-scale language models, In the expression guidelines information, expressions used in the field of managed services are, A whitelist is a list of expressions that are safe to use, and / or We keep a blacklist of expressions that are considered dangerous to use. Only whitelisted expressions are permitted access and / or execution, or I instructed them to replace expressions in the blacklist with those in the whitelist.

[0088] Employee A, responsible for managing services, updated the expression guidelines created in the past during the process of pre-screening public relations information for customers in the managed services sector using traditional methods such as manual processes and computer or large-scale language models. This updated guideline was then compiled by adding the latest white data, black data, and a collection of paraphrases of black data into white data, along with necessary instructions (prompts) for the large-scale language model in YAML format. This updated guideline was then entered into a database accessible to the large-scale language model as training data (Step 1).

[0089] Employee A input the manuscript text expressions, which were organized into the table in Figure 2, into the large-scale language model and also instructed it to revise the manuscript text expressions (Step 2).

[0090] In Step 2, the large-scale language model modified the input manuscript text sentence by sentence based on training data. The modified text was displayed alongside the original manuscript text for comparison, as shown in Figures 5-1(a) and 5-2(a) (Step 3).

[0091] Employee A, upon viewing the displays in Figures 5-1(a) and 5-2(a), judged whether the original text expression and the revised text expression were appropriate or inappropriate for each component and clicked the corresponding icon (in Figures 5-1(a) and 5-2(a), the clicked icons remain, and the unclicked icons are deleted) (Steps 4 and 4').

[0092] In this example, common strings were not modified; only differing strings were corrected.

[0093] Here, in the large-scale language model, employee A is in charge of: Clicking the icon corresponding to "appropriate" (check icon (1)) will not instruct the large-scale language model to modify the clicked component. Clicking an icon that indicates "inappropriate" (trash can icon (2)) instructs the large-scale language model to modify the clicked component and requests feedback (Figures 5-1(b), 5-2(b)).

[0094] The large-scale language model generates revised text representations for the components clicked as "inappropriate," based on the instructed corrections and the latest training data. These are displayed as response feedback as shown in Figures 5-1(c) and 5-2(c), and the confirmed corrections are reflected in the original text representation, for example, as white text in a black box (Steps 5 and 5').

[0095] The correction to sentence 1 of the example (upper right column of Figure 4-1) is completed in Figure 5-1(c) (Step 4 → End), The correction to statement 2 of the example (bottom right of Figure 4-1) is not completed in Figure 5-2(c), and requires further correction work (Figure 5-2(d)) before being completed in Figure 5-2(e) (Step 4' → End).

[0096] Furthermore, large-scale language models are Information regarding instructions given by the worker in Step 4 and Step 4', Information regarding the revised text expression generated in Step 5 and Step 5', In addition to the existing expression guideline information, new expression guideline information is generated and the expression guidelines are updated (Steps 6 and 6').

[0097] Information Processing System of the Present Invention The information processing system of the present invention includes at least, An information display medium for displaying the original text expression, revised text expression, and further revised text expression in steps 1 to 6, 4' to 6', and 7 of the method of the present invention, An information storage medium for storing the expression guidelines in steps 1 to 6, 4' to 6' and 7 of the method of the present invention, This is an information processing system for modifying public relations information, comprising a large-scale language model that performs steps 1-6, 4'-6', and 7 of the method of the present invention.

[0098] In this invention, the information display medium can be a display medium such as a personal computer, smartphone, tablet, or a display used in virtual reality (VR).

[0099] In the present invention, the information storage medium may be a database built into or externally mounted on a device that constitutes a large-scale language model, or it may be an externally located cloud.

[0100] In the present invention, the large-scale language model may be, for example, a deep learning model with millions to billions of parameters that has been pre-trained on a large-scale corpus.

[0101] In this invention, the terminal medium for inputting information into the large-scale language model may be built into the device constituting the large-scale language model, or it may be externally mounted.

[0102] Figure 6 shows an example of the system flow when the present invention 2 performs steps 1 to 6, 4' to 6', and 7 of the method of the invention.

[0103] The modification of the manuscript text expression of public relations information according to the present invention 2 is: A large-scale language model displays the original text expression, the revised text expression, or the further revised text expression (Figure 4-2) for each sentence and for each sentence unit specified by the worker. The operator judges the appropriateness of each sentence unit and instructs the large-scale language model to make corrections. The large-scale language model then provides feedback with suggested corrections until all sentence units are appropriate, and the corrected content is finalized. For workers, this reduces the workload of revising the text of public relations materials, allowing them to complete the revision work quickly and accurately. [Explanation of Symbols]

[0104] 1. A checkmark icon indicates that it has been judged as "appropriate". 2. The trash can icon indicates that the content has been deemed "inappropriate".

Claims

1. The following steps involve using a large-scale language model to revise the draft text of a business entity's public relations information to an expression that the business entity deems appropriate: Step 1: The latest expression guideline information regarding the manuscript text expression, which is the text expression of the aforementioned manuscript, is input into the large-scale language model as training data and stored in an information storage medium accessible to the large-scale language model; Step 2: The operator inputs the manuscript text expression into the large-scale language model and instructs it to modify the manuscript text expression; Step 3: The large-scale language model modifies the manuscript text expression entered in step 2 based on the training data stored in the information storage medium, and displays the modified text expression to the operator; With respect to the modified text expression displayed in step 3, If the worker determines it is appropriate, the worker will not instruct the large-scale language model to make any modifications, or will instruct not to do so. Step 4: If the operator determines that the text is inappropriate, the operator instructs the large-scale language model to revise the modified text representation; Step 5: The large-scale language model generates and displays a revised text expression in which the revised text expression has been modified based on the modifications instructed in step 4 and the training data, and responds to the operator; The aforementioned large-scale language model At a minimum, information relating to the instructions given by the worker in step 4, Includes information regarding the revised text expression generated in step 5, Step 6: Generate new expression guideline information and update the expression guidelines; With respect to the aforementioned revised text expression, If the worker determines it is appropriate, the worker will not instruct the large-scale language model to make any modifications, or will instruct not to do so. If the operator determines that the text is inappropriate, the operator instructs the large-scale language model to revise the revised text representation in step 4'; Step 5' involves the large-scale language model generating and displaying a revised text expression based on the revisions instructed in step 4' and the training data, and responding to the operator; The aforementioned large-scale language model At a minimum, information relating to the instructions given by the worker in step 4', Includes information regarding the revised text expression generated in step 5', Step 6': Newly generate expression guideline information and update the expression guideline; and Step 4', Step 5', and Step 6' are performed as follows: A method for revising the text expression of public information, comprising step 4' above, followed by step 7, which is repeated until the worker determines that the revised text expression is appropriate, In step 3, the large-scale language model provides the operator with The aforementioned manuscript text expression is displayed sentence by sentence, separated into the sentence components specified by the operator, The generated revised text and the further revised text are displayed sentence by sentence, separated into the constituent elements of the sentence. In step 4 and step 4', The worker shall determine whether the revised text expression and the further revised text are appropriate or inappropriate for each component of the sentence, The components of the aforementioned sentence that are deemed appropriate will not be instructed to be modified by the large-scale language model. The large-scale language model is instructed to modify the components of the aforementioned sentence that it has determined to be inappropriate. In step 5 and step 5', The large-scale language system displays the revised text representation for each component of the sentence, A method for revising the textual expression of public relations information to be provided to the aforementioned worker.

2. A method for modifying the text representation of public relations information according to claim 1, wherein the information to be input to or instructed to the large-scale language model is converted into a markup language and input to or instructed to the large-scale language model.

3. The above sentence is expressed in a language that includes ideographic characters, The sentence components specified by the aforementioned worker are, When each of the sentences in the aforementioned manuscript text expression, the aforementioned revised text expression, and the aforementioned further revised text expression are considered as mere strings of characters, The common string that is shared by the two or more characters that make up each string, A method for modifying the text representation of public information according to claim 1 or 2, wherein the differing string is one other than the common string.

4. In the aforementioned expression guidelines information, the expressions used in the business field to which the public relations information belongs are: In the aforementioned business field, A whitelist is a list of expressions that are safe to use, and / or It is organized into a blacklist, which is a list of expressions that are considered dangerous to use. A method for modifying the text representation of public information according to claim 1 or 2, wherein the large-scale language model is permitted to access and / or execute only expressions on a whitelist, or expressions on the blacklist can be replaced with expressions on the whitelist.

5. at least, An information display medium for displaying the original text expression, revised text expression and further revised text expression in steps 1 to 6, 4' to 6' and 7 described in claim 1, An information storage medium for storing the expression guidelines in steps 1 to 6, 4' to 6' and 7 of claim 1, An information processing system for modifying public relations information, comprising a large-scale language model that performs steps 1 to 6, 4' to 6' and 7 as described in claim 1.

Images