Multidimensional language parameter-controlled input processing device, input processing method, input processing program, and recording medium

Abstract

This invention provides an input processing device, method, and program that reduce nuance discrepancies caused by various gaps between the sender and receiver in natural language communication (language barriers, positional barriers, cultural barriers, expertise barriers, relationship barriers, and purpose barriers). [Solution] The input processing device includes: a parameter mapping means that acquires operation information including a combination of continuous operations, discrete operations, or natural language instructions, and maps the operation information to points or weighted combinations in a multi-dimensional parameter space representing a part of the writing style, etc.; a candidate generation means that generates language expression candidates including translation candidates or writing style conversion candidates based on the parameters; a candidate update means that regenerates the entire list of candidates or updates only a part of them differentially; an auxiliary information generation means that generates auxiliary information that visualizes the differences in writing style between the source text and the candidates, etc.; and a presentation control means that synchronizes the updating and presentation of language expression candidates and auxiliary information while within a single operation context.

Description

【Technical Field】 【0001】 The present invention relates to translation processing technology related to natural language. In particular, it relates to nuance control in multilingual translation and conversion / adjustment within the same language. 【Background Art】 【0002】 In recent years, with the progress of machine translation and text generation technology based on large language models (LLMs), input interfaces that support communication between different languages and conversion / adjustment of style, tone, and honorific level within the same language have been widely used (Patent Document 1). 【0003】 In this technical field, especially in multilingual translation, while maintaining the intention of the original text, adjusting the formality, politeness, intensity of emotion, sentence length, and nuance according to the reader in the target language is an important process for a more natural translation. 【0004】 Also, in business communication, it is essential to select and distinguish finer nuances according to the relationship with the other party and the situation. However, in Japanese business communication, there are still many translation examples that are unnatural as business documents or overly casual expressions. Furthermore, while the text expression is too rough as a business email to a customer, excessive use of honorifics in the text is also a problem (Patent Document 2). 【0005】 Regarding such a gap between the "self-evaluation" of honorifics and the "impression on the recipient side", conventional input support systems and proofreading tools mainly focus on detecting grammar errors and typos, and do not sufficiently provide means to handle the degree of honorifics and formality of style as continuous indicators and adjust expressions based on them. 【0006】 Furthermore, with previous translation systems, if a user wanted to change the nuance of the translation result, they had to rewrite the original text itself using extreme expressions and then re-translate it. For example, if they wanted to express "thank you" in a more casual way, they had no choice but to change the original text to extreme expressions such as "Seriously grateful!" or "I'm so grateful!" and input it directly into the translation system, which was cumbersome (Patent Document 3). 【0007】 Such extreme and non-standard source texts tend to be translated into the target language as equally extreme and unnatural expressions, often resulting in outcomes that deviate from the intended meaning. 【0008】 In this context, most conventional translation and input assistance systems employ a two-stage pipeline structure that clearly separates the "input phase" and the "conversion phase." Specifically, after the user inputs the source text, processing is executed through explicit operations such as translation and conversion buttons, and the results are presented statically and all at once. 【0009】 In the above configuration, the translation result is fixed to a single value, and the user is not provided with a means to continuously adjust the level of formality, politeness, sentence length, emotional intensity, etc. Therefore, if the translation result does not convey the expected nuance, the user is forced to repeat the trial-and-error process of revising the original text and running the translation again, which hinders efficient communication. 【0010】 Furthermore, when adjusting multidimensional expression parameters such as writing style (casual / formal), degree of honorifics, tone (assertive / indirect), sentence length (concise / detailed), and target audience (expert / general consumer), users typically had to either select a preset on a separate settings screen or repeat trial-and-error input multiple times. 【0011】 Furthermore, with the recent advancements in large-scale language models (LLMs), stylistic adjustment using prompts has become widely used. Users provide natural language instructions (prompts) such as "make it more formal," "make it a little more casual," or "make it sound like you're talking to a friend," and the LLM then generates output as a result, thus adjusting the nuances. However, this LLM + prompt method had the following structure (Patent Document 4). 【0012】 Firstly, there is the problem of the invisibility of nuance space. In conventional LLM+ prompting methods, users provide natural language instructions, but it is difficult to predict in advance how much nuance change those instructions will bring about. 【0013】 Secondly, there is the problem of the inefficiency of regenerating the entire output. In the conventional LLM+ prompt method, even if you only wanted to change the nuance slightly, it was necessary to change the prompt and regenerate the entire output. 【0014】 Thirdly, there is the problem of the difficulty of simultaneous control of multiple axes. The nuances of linguistic expression are composed of multiple independent axes such as style, politeness, tone, emotion, sentence length, and target audience. However, with the conventional LLM+prompt method, it was difficult to control these axes independently and simultaneously. 【0015】 Fourthly, there is the problem of discontinuity in the output. In the conventional LLM+ prompting method, even a small change in the prompt can lead to a large change in the output, sometimes resulting in unpredictable and unstable output behavior. 【0016】 Fifth, there are limitations to discrete style selection. Existing translation services offer discrete style options such as "formal" and "casual." However, in actual business communication and everyday conversation, continuous nuance adjustments are required. 【0017】 These issues stem from the fact that the conventional LLM+ prompt method is based on a "prompt trial-and-error paradigm." [Prior art documents] [Patent Documents] 【0018】 [Patent Document 1] Japanese Patent Publication No. 2017-068879 [Patent Document 2] Special Publication No. 2007-532995 [Patent Document 3] U.S. Patent Application Publication No. 2013-0197898 [Patent Document 4] European Patent Application Publication No. 3884419 [Overview of the Initiative] [Problems that the invention aims to solve] 【0019】 This invention has been made in view of the above-mentioned prior art, and aims to reduce the discrepancies in nuance caused by various gaps (language barriers, positional barriers, cultural barriers, expertise barriers, relationship barriers, and purpose barriers) that exist between the sender and receiver in natural language communication. [Means for solving the problem] 【0020】 An input processing device for inputting source text, comprising: parameter mapping means that analyzes operation information including one of a continuous operation, a discrete operation, or a natural language instruction performed by a user, and associates the operation information with a continuous quantity, a predetermined point, or a weighted combination of multiple reference points in a multi-dimensional parameter space composed of attributes of language expression representing at least one of style, politeness, tone, emotion, sentence length, and target audience; candidate generation means that generates language expression candidates including translation candidates or style conversion candidates based on the results of the mapping between the source text and the parameter space; and presentation control means that updates and presents the language expression candidates according to the parameters mapped by the mapping means. 【0021】 An input processing device, wherein the parameter association means associates a plurality of discrete operation inputs with different points in the parameter space in a plurality of dimensions, the candidate generation means generates a plurality of language expression candidates corresponding to the plurality of different points, and the presentation control means presents the plurality of language expression candidates. 【0022】 An input processing method for generating language expression candidates for the original text, comprising: a parameter association step of acquiring operation information given by a user and associating the operation information with a continuous quantity or a predetermined point in a parameter space in a plurality of dimensions representing attributes of language expressions; a candidate generation step of generating language expression candidates including translation candidates or style conversion candidates based on the original text and the association result in the parameter space; and a presentation control step of updating and presenting the language expression candidates according to the associated parameters. 【0023】 An input processing device for inputting an original text, comprising: an operation reception means for acquiring operation information that can change a continuous quantity by a user and starting and maintaining a single operation context; a parameter determination means for continuously updating parameters in a plurality of dimensions representing attributes of language expressions based on the operation information acquired by the operation reception means; a candidate generation means for generating language expression candidates including translation candidates or style conversion candidates based on the original text and the parameters; and a presentation control means for continuously updating and presenting the language expression candidates according to the parameters while within the single operation context. 【0024】 An input processing device, further comprising a point selection parameter determination means for treating the discrete value as a predetermined point in the parameter space in the plurality of dimensions when the operation information acquired by the operation reception means is a discrete value and updating the language expression candidates based on the selected point. 【0025】 An input processing device for inputting source text, comprising: an operation receiving means for acquiring operation information that can change a continuous quantity by the user, and for starting and maintaining a single operation context; a parameter determination means for continuously updating multi-dimensional language parameters, including at least two of style, tone, politeness, emotion, sentence length, and target audience, based on the operation information acquired by the operation receiving means; a candidate generation means for generating language expression candidates, including translation candidates or style conversion candidates, based on the source text and the language parameters; and a presentation control means for continuously updating and presenting the language expression candidates according to the language parameters while within the single operation context. 【0026】 An input processing device further comprises, when the operation information acquired by the operation receiving means is a discrete value, a point selection parameter determination means that treats the discrete value as a predetermined point in the language parameter space of multiple orders and updates the language expression candidate based on the selected point. 【0027】 An input processing device for inputting source text, comprising: an operation receiving means for acquiring operations by a user that can change a continuous quantity, and for starting and maintaining a single operation context; a parameter determination means for determining multidimensional language parameters, including at least two of style, tone, politeness, emotion, sentence length, and target audience, based on the operation information acquired by the operation receiving means, and for continuously updating points in a multidimensional parameter space; a candidate generation means for generating language expression candidates, including translation candidates or style conversion candidates, based on the source text and the language parameters; a candidate update means for selectively performing a partial update process to regenerate the entire candidate or update only the part related to the change, depending on the amount of change in the language parameters; an auxiliary information generation means for analyzing at least one of style differences, semantic differences, tone differences, changes in politeness, and appropriateness of honorifics between the source text and the language expression candidates, and generating auxiliary information including nuance information expressed as at least one of a numerical score, graph, icon, color, or annotation text; and a presentation control means for synchronously updating and presenting the language expression candidates and the auxiliary information while within the single operation context. 【0028】 The input processing device generates candidate language expressions by translating the source text into an intermediate language, performing nuance adjustments based on language parameters in the intermediate language, and translating the adjusted intermediate language expression into a target language. 【0029】 An input processing device wherein the target language is the same language as the source text, and the candidate generation means generates stylistic conversion candidates within the same language by back-translating the adjusted intermediate language expression into the language of the source text. 【0030】 An input processing device wherein the auxiliary information generation means evaluates the appropriateness of honorific expressions in the candidate language expression and, if it detects at least one of double honorifics, insufficient honorifics, or inappropriate honorific forms, includes information in the auxiliary information that suggests an appropriate alternative expression. 【0031】 The input processing device includes an operation receiving means which acquires operation information from at least one of the following: a slide operation on a touch panel, a drag operation with a mouse, a rotation operation with a wheel device, or a change in volume or pitch in voice input. 【0032】 This enables nuance control in both multilingual and intralingual communication. In other words, while conventional translation systems fix the translation result to a single, unified format, making it difficult for users to continuously adjust formality, politeness, tone, sentence length, and target audience, this system defines these as axes in a multidimensional parameter space, allowing for the updating of these parameters and the presentation of updated candidates through continuous operations within a single operation context. 【0033】 The input processing device includes a presentation control means which performs the update of the language expression candidates and the presentation of the auxiliary information, based on the change in the language parameters, within 100 milliseconds. This enables nuance control, parameter updating, and presentation to be performed without stress. 【0034】 The input processing device uses as the intermediate language a language that satisfies at least one of the following conditions: the amount of training data for the language model is abundant, the syntactic structure of the language is relatively clear, or it is easy to function as a reference point for a common semantic space in a multilingual model. 【0035】 The input processing device includes an operation receiving means which determines the continuation of the single operation context by dynamically changing the threshold of the operation interval according to the type of operation, the user's operation pattern, or the characteristics of the application. 【0036】 An input processing device wherein the candidate update means sets different weights for a plurality of language parameter axes and selects the partial update process or the whole regeneration based on the weighted change amount. 【0037】 An input processing device wherein the presentation control means switches between one of the following modes based on user settings or application settings: a mode in which all of the auxiliary information is always presented, a mode in which the auxiliary information is selectively presented in response to a user request, or a mode in which only a part of the auxiliary information is presented. 【0038】 An input processing method for inputting source text, comprising: an operation reception step in which an operation reception means acquires operation information that can change a continuous quantity by the user, and starts and maintains a single operation context; a parameter determination step in which a parameter determination means determines multidimensional language parameters, including style, tone, politeness, emotion, sentence length, and at least two of the target audience, based on the operation information acquired by the operation reception step, and continuously updates points in a multidimensional parameter space; a candidate generation step in which a candidate generation means generates language expression candidates, including translation candidates or style conversion candidates, based on the source text and the language parameters; and a candidate update means, The system includes: a candidate update step that selectively performs a candidate either overall regeneration or partial update process that updates only the parts related to the change, depending on the amount of change in the language parameters; an auxiliary information generation step in which an auxiliary information generation means analyzes at least one of the differences in style, semantic differences, tone differences, changes in politeness, and appropriateness of honorifics between the source text and the language expression candidates, and generates auxiliary information that includes nuance information representing the analysis results as at least one of a numerical score, graph, icon, color, or annotation text; and a presentation control step in which a presentation control means synchronously updates and presents the language expression candidates and the auxiliary information while they are within the single operation context. 【0039】 An input processing method for inputting source text, comprising: an operation reception step in which an operation reception means acquires operation information that can change a continuous quantity by the user, and starts and maintains a single operation context; a parameter determination step in which a parameter determination means determines multidimensional language parameters, including style, tone, politeness, emotion, sentence length, and at least two of the target audience, based on the operation information acquired by the operation reception step, and continuously updates points in a multidimensional parameter space; a candidate generation step in which a candidate generation means generates language expression candidates, including translation candidates or style conversion candidates, based on the source text and the language parameters; and a candidate update means updates the language parameters The input processing method is performed by a computer, which includes: a candidate update step that selectively performs a candidate update process that either completely regenerates the candidates or updates only the parts related to the change, depending on the amount of change; an auxiliary information generation step in which an auxiliary information generation means analyzes at least one of the differences in style, semantic differences, tone differences, changes in politeness, and appropriateness of honorifics between the source text and the language expression candidates, and generates auxiliary information that includes nuance information, which represents the analysis results as at least one of a numerical score, graph, icon, color, or annotation text; and a presentation control step in which a presentation control means updates and presents the language expression candidates and the auxiliary information synchronously while they are within the single operation context. 【0040】 The input processing device includes a parameter mapping means for determining a plurality of style reference points and calculating parameter values ​​as a weighted combination between the plurality of style reference points based on user operation information, wherein the weighted combination is performed as one or more combinations of a linear combination of a plurality of prompt templates, interpolation of a plurality of style embedding vectors, synthesis of a plurality of model adapter parameters, and weighting of a plurality of style tags. 【0041】 The system includes: a parameter receiving unit that receives a language processing request including a multidimensional parameter vector from a client device; a prompt mapping unit that maps the multidimensional parameter vector to at least one of a prompt, a system message, or a generation control parameter for a large-scale language model; a language model execution unit that executes the large-scale language model using the mapped prompt, system message, or generation control parameter and generates language expression candidates; and a response generation unit that transmits the generated language expression candidates to the client device. 【0042】 An LLM service provider server, wherein the prompt mapping unit generates the prompt by selectively combining or interpolating a plurality of prompt templates according to the values ​​of each dimension of the multidimensional parameter vector. 【0043】 An LLM service provider server, wherein the language model execution unit dynamically adjusts at least one of the sampling temperature, top P value, or iteration penalty at the time of generation based on the multidimensional parameter vector. 【0044】 An input field monitoring unit monitors the input fields of an existing application and detects the input of source text; a discrete input conversion unit obtains discrete style selection operations from the existing application and converts the discrete selections into predetermined points in a multidimensional parameter space; The LLM service provider server includes: a parameter receiving unit that receives a language processing request including a multidimensional parameter vector from a client device; a prompt mapping unit that maps the multidimensional parameter vector to at least one of a prompt, a system message, or a generation control parameter for a large-scale language model; a language model execution unit that executes the large-scale language model using the mapped prompt, the system message, or the generation control parameter and generates language expression candidates; and a response generation unit that sends the generated language expression candidates to the client device. The LLM service provider server includes a linkage API management unit that transmits points in the multidimensional parameter space to acquire language expression candidates, and an insertion control unit that inserts the acquired language expression candidates into the input field or candidate display area of ​​the existing application. 【0045】 The user interface integrated gateway further receives continuous operation input from the user, continuously updates the points in the multidimensional parameter space as relative movement amounts from the predetermined points, and transmits the updated parameters to the LLM service provider server. 【0046】 An input processing device for inputting source text, comprising: parameter mapping means that analyzes operation information including one of a continuous operation, a discrete operation, or a natural language instruction performed by a user, and associates the operation information with a continuous quantity, a predetermined point, or a weighted combination of multiple reference points in a multidimensional parameter space composed of attributes of language expression representing at least one of style, politeness, tone, emotion, sentence length, and target audience; candidate generation means that generates language expression candidates including translation candidates or style conversion candidates based on the results of the mapping between the source text and the parameter space; and presentation control means that updates and presents the language expression candidates according to the parameters mapped by the mapping means; and a multidimensional parameter vector from a client device An LLM service provider server comprises: a parameter receiving unit that receives language processing requests; a prompt mapping unit that maps the multidimensional parameter vector to at least one of a prompt, a system message, or a generation control parameter for a large-scale language model; a language model execution unit that executes the large-scale language model using the mapped prompt, the system message, or the generation control parameter and generates language expression candidates; and a response generation unit that transmits the generated language expression candidates to the client device. The input processing device generates a multidimensional parameter vector based on user operations and transmits it to the LLM service provider server, and the LLM service provider server generates language expression candidates based on the multidimensional parameter vector and returns them to the input processing device. 【0047】 The system comprises: semantic-style separation means for separating and extracting semantic structure M and style information S from the source text; style conversion means for converting only the style information S based on multi-dimensional language parameters under the constraint that the semantic structure M is not changed; and reconstruction means for combining the retained semantic structure M and the converted style information S' to generate a linguistic representation of the target language. 【0048】 An input processing device for inputting source text, comprising: an operation receiving means for acquiring continuous or discrete operation information from a user and starting and maintaining a single operation context; a parameter determination means for continuously updating a multidimensional parameter vector P representing at least a portion of the style, politeness, tone, emotion, sentence length, and target audience based on the operation information acquired by the operation receiving means; a candidate generation means for generating language expression candidates based on the source text and the parameter vector P; and a parameter change amount calculation that calculates a parameter change amount ΔP based on the difference between the previous value P_old and the current value P_new of the parameter vector P. The system includes: a means; a candidate update means that performs a whole regeneration process to regenerate the entire language expression candidate if the parameter change amount ΔP is greater than or equal to a predetermined first threshold, and a partial update process that updates only a part of the language expression candidate differentially if the parameter change amount ΔP is less than the first threshold; an auxiliary information generation means that analyzes the difference in writing style, level of politeness, appropriateness of honorifics, sentence length ratio, etc., between the original text and the language expression candidate, and generates the analysis results as auxiliary information; and a presentation control means that synchronously displays the language expression candidate updated by the candidate update means and the auxiliary information within the single operation context. 【0049】 The system includes: a request receiving unit that receives a language processing request from a client device that includes information about the source text, the previously generated language expression candidates, and a multidimensional parameter vector P; a server-side parameter change amount calculation unit that calculates a parameter change amount ΔP based on the difference between the previous value P_old and the current value P_new included in the language processing request; a candidate generation processing unit that, when the parameter change amount ΔP is greater than or equal to a predetermined threshold, executes a whole regeneration process to newly generate the entire language expression candidate based on the source text and the current value P_new, and when the parameter change amount ΔP is less than the predetermined threshold, executes a partial update process to replace only the part of the previously generated language expression candidate that is affected by the parameter change; and a response transmission unit that transmits the language expression candidates generated or updated by the candidate generation processing unit and auxiliary information as necessary to the client device. 【0050】 An auxiliary information generation device for visualizing nuance differences between source text and candidate language expressions, comprising: acquisition means for acquiring source text and at least one candidate language expression corresponding to the source text; analysis means for analyzing at least two of the following between the source text and the candidate language expression: differences in writing style, level of politeness, appropriateness of honorifics, differences in tone, sentence length ratio, lexical changes, and emotional tone; auxiliary information generation means for generating nuance auxiliary information that expresses the analysis results by the analysis means as at least one of numerical scores, graphs, icons, color coding, or annotation text; and presentation means for presenting the candidate language expression and the nuance auxiliary information synchronously on a display device. 【0051】 An input processing device wherein the operation receiving means acquires discrete style selection operations, given as pressing of a plurality of preset buttons in addition to continuous operations, as operations within the single operation context, and the parameter determination means maintains a predetermined reference point for a multidimensional parameter vector for each of the preset buttons, and updates the parameter vector P to the vicinity of the reference point in response to the button press. 【0052】 The input processing device includes a parameter change amount calculation means which, upon each press of the preset button, calculates a parameter change amount ΔP based on the difference between the parameter vector P_old immediately before the press and the parameter vector P_new immediately after the press, and a candidate update means which performs a partial update process or a full regeneration process for the language expression candidate according to the parameter change amount ΔP. 【0053】 These methods enable a balance between responsiveness and output stability through the selective execution of partial update processing and overall regeneration processing based on the amount of parameter change. Based on the amount of change ΔP in the multidimensional parameter space, if ΔP is large, the entire candidate is regenerated, and if ΔP is small, only polite expressions, sentence endings, modifiers, etc. are differentially updated. This reduces the number of tokens generated on the server side and the amount of data sent and received between the client and server while maintaining responsiveness to user operations. 【0054】 An input processing device wherein the operation receiving means acquires a single slider operation, and the parameter determination means takes the position of the slider as input and updates the multidimensional parameter vector P by simultaneously determining values ​​for multiple axes, including style, politeness, tone, and emotion, based on a predefined mapping function. 【0055】 An input processing device further comprises axis definition means for a user to register the names of nuance axes and conversion policies for language expressions corresponding to the nuance axes, wherein the multidimensional parameter vector P includes, in addition to axes predetermined by the system, elements corresponding to at least one nuance axis determined by the user by the axis definition means. 【0056】 This enables convergent behavior and approximate reversibility. In this invention, the multidimensional parameter space is defined as a continuous and bounded space, and the output language expression candidates are designed to change continuously in response to continuous changes in the parameters. 【0057】 Furthermore, when the parameters are returned to their original point neighborhood, the candidates also return to their original representation neighborhood, achieving approximate reversibility, which allows users to explore the parameter space with confidence. 【0058】 An input processing device further comprises: an instruction input means for inputting nuance adjustment instructions in natural language; and an instruction interpretation means for analyzing the nuance adjustment instructions in natural language and updating at least some components of the multidimensional parameter vector P according to the content of the nuance instructions. 【0059】 Visualizing these nuance changes makes it easier for users to understand the results of their actions not only subjectively but also objectively and quantitatively. By presenting stylistic differences, politeness levels, honorific language appropriateness scores, sentence length ratios, vocabulary change lists, and emotional change indicators as numerical scores, graphs, color coding, and annotated text, users can understand at a glance how much each axis has shifted. 【0060】 The input processing device comprises an auxiliary information generation means that calculates a numerical score indicating the degree of honorifics, politeness level, or formality of the writing style for the source text, a parameter determination means that automatically sets the initial value of the multidimensional parameter vector P based on the numerical score, and a user interface that displays the initial position of the sliders for each axis at the position corresponding to the initial value. 【0061】 This enables a consistent nuance adjustment function across diverse devices and applications through a technology structure independent of input method. Because different input methods—such as keyboard input, mouse / touch operation, gestures, voice input, existing discrete style selection UIs, and natural language instructions—can be uniformly treated as inputs to a multidimensional parameter space, a common architecture can be applied to smartphone IMEs, browser extensions, standalone applications, chatbots, and more. [Effects of the Invention] 【0062】 This invention makes it possible to reduce nuance discrepancies caused by various gaps (language barriers, positional barriers, cultural barriers, expertise barriers, relationship barriers, and purpose barriers) that exist between the sender and receiver in natural language communication. [Brief explanation of the drawing] 【0063】 [Figure 1] This is a functional block diagram of an input processing device relating to one embodiment of the present invention. [Figure 2] This figure shows variations in the presentation area for language expression candidates and auxiliary information according to one embodiment of the present invention. [Figure 3] This figure shows a dynamic update sequence in a single operation context according to one embodiment of the present invention. [Figure 4] This diagram shows the flow of the operation reception process related to one embodiment of the present invention. [Figure 5]This is a conceptual diagram showing an example of the configuration of a multidimensional parameter space related to one embodiment of the present invention. [Figure 6] This figure shows an example of applying one embodiment of the present invention to an IME (Input Method Editor). [Figure 7] This figure shows an example of the screen configuration of an independent application according to one embodiment of the present invention. [Figure 8] This is a block diagram showing the application integration configuration for one embodiment of the present invention. [Modes for carrying out the invention] 【0064】 In this invention, the following terms are defined as follows: A linguistic expression candidate is a broad concept that includes translation candidates to a target language different from the source language, and stylistic conversion candidates within the same language in which nuances such as style, tone, level of honorifics, emotion, sentence length, and target audience have been adjusted. 【0065】 The source language refers to the language in which the source text entered by the user is written. The intermediate language refers to a language different from the source language and target language, used to abstract and normalize semantic structure and style information. 【0066】 The target language refers to the language to which the text has been translated, and includes the same language as the source language (in the case of stylistic transformation within the same language). A multidimensional parameter space refers to a parameter space composed of multiple continuous values ​​representing attributes of linguistic expression, such as style, politeness, tone, emotion, sentence length, and target audience. 【0067】 Auxiliary information refers to nuance-based supplementary information that quantifies or visualizes stylistic differences, levels of politeness, appropriateness of honorifics, sentence length ratios, vocabulary changes, emotional tone, etc., between the original text and each candidate language expression. 【0068】 A single operation context is a series of operations performed by a user from start to finish. It is distinguished as a set of operations because the intervals between operations are within a predetermined time, or because the changes in operations are continuous. 【0069】 Embodiments of the present invention will be described in detail with reference to the drawings. The present invention can also be expressed as an input processing method that executes each means in a time series. Furthermore, it can be realized as an input processing program for causing a computer to execute the input processing method, and as a computer-readable recording medium on which the program is stored. However, the present invention is not limited to these embodiments. 【0070】 (Embodiment 1) First, the basic configuration and integrated processing pipeline will be described. As shown in Figure 1, the input processing device 10 of this embodiment includes an operation reception means 11, a parameter determination means 12, a candidate generation means 13, a candidate update means 14, an auxiliary information generation means 15, and a presentation control means 16. Each of these means is realized either as a dedicated hardware circuit or by the CPU executing an input processing program. 【0071】 These six methods work together holistically within a single continuous operation context to achieve simultaneous input and nuance adjustment. In conventional technology, the input phase and conversion phase were separated, making it impossible to operate these methods in an integrated manner. However, this invention achieves this integration by introducing the concept of a single operation context. 【0072】 More specifically, the operation reception means 11 maintains the context from the time the user starts an operation until it ends. During this time, the parameter determination means 12 continuously updates the parameters, and the candidate generation means 13 and candidate update means 14 generate and update candidates in real time. The auxiliary information generation means 15 analyzes the nuance differences, and the presentation control means 16 presents them synchronously. In this way, the six means work together within a single context, enabling simultaneous execution of input and nuance adjustment. 【0073】 The operation reception means 11 acquires operation information from an interface capable of detecting a continuous amount of operation, such as a touch panel, mouse, trackpad, gesture sensor, wheel device, or microphone (not shown). The operation information includes the start time of the operation, the end time of the operation, coordinate position, movement trajectory, movement speed, movement direction, pressure intensity, voice parameters, and a specific gesture pattern. 【0074】 The operation reception means 11 monitors the operation interval to determine the continuity of operations. Specifically, if the time interval between consecutive operation events is within a predetermined time (e.g., 250 milliseconds), those operations are treated as a single operation context. Furthermore, if the trajectory of operations is smooth and continuous, it is also determined to be a single operation context. 【0075】 At the start of a single operation context, the operation receiving means 11 generates a context identifier and supplies it to the subsequent processing means. This context identifier allows for the independent management of multiple operation sequences, even if they occur simultaneously. 【0076】 For example, in the case of voice input, the volume or pitch of speech can be mapped to the emotion axis, and the duration of speech can be mapped to the sentence length axis, thereby creating a configuration that controls multidimensional parameters using only microphone input. 【0077】 The parameter determination means 12 determines language parameters as points in a multidimensional parameter space consisting of a style axis, a politeness axis, a tone axis, a sentence length axis, an emotional axis, and a target reader axis, based on the operation information supplied from the operation reception means 11. 【0078】 Furthermore, multiple dimensions can be controlled simultaneously from a single continuous operation. For example, in a continuous operation such as touch operation, by associating the horizontal movement amount with the style axis and the vertical movement amount with the politeness axis, two parameters can be controlled simultaneously with a two-dimensional operation. Moreover, by associating the pressure intensity with the emotion axis, simultaneous three-dimensional control becomes possible. 【0079】 Each axis parameter is expressed as a normalized value within the range of 0 to 1. For example, on the style axis, 0 represents casual and 1 represents formal. On the politeness axis, 0 represents no honorifics and 1 represents the highest level of honorifics. On the tone axis, 0 represents assertive and 1 represents indirect expression. On the sentence length axis, 0 represents concise and 1 represents detailed. 【0080】 The parameter determination means 12 continuously updates these parameter values ​​based on the operation trajectory and speed. The update interval is set, for example, in the range of 16 milliseconds to 100 milliseconds to achieve responsiveness suitable for human visual feedback recognition. 【0081】 The candidate generation means 13 generates language expression candidates, including translation candidates or stylistic transformation candidates, based on the source text and the parameters at that time, using a neural network-based language model, a statistical translation model, a rule-based transformation model, or a combination thereof. 【0082】 The language model used in the candidate generation means 13 may be a model installed in the client device. Alternatively, the configuration may involve querying a model located on a server device via a network. When using a large-scale language model (LLM), candidates corresponding to the parameters can be generated by supplying parameter values ​​to the language model as part of the prompt. 【0083】 For example, consider a case in Japanese-to-English translation where the style parameter is 0.8 (leaning towards formal) and the politeness parameter is 0.6 (moderate level of politeness). The candidate generation means 13 supplies the source text to the language model with the instruction, "Translate using formal business English and polite expressions," and obtains candidates. 【0084】 The candidate update means 14 monitors the change in parameters ΔP from the parameter determination means 12. Here, ΔP is defined as the distance between the previous parameter value and the current parameter value in the multidimensional parameter space. The distance can be calculated using the Euclidean distance, Manhattan distance, or Mahalanobis distance. 【0085】 For example, in a 6-dimensional parameter space (style, politeness, tone, sentence length, emotion, target audience), let P_prev = (p1, p2, p3, p4, p5, p6) be the previous parameter vector and P_curr = (q1, q2, q3, q4, q5, q6) be the current parameter vector. The Euclidean distance is calculated using the following formula. 【number】 【0086】 If ΔP is greater than or equal to a predetermined threshold (e.g., 0.1), the candidate update means 14 requests the candidate generation means 13 to perform a complete regeneration. In a complete regeneration, language expression candidates are generated entirely from scratch based on the source text and the current parameters. This process is computationally intensive but is necessary to accommodate significant changes in parameters. 【0087】 On the other hand, if ΔP is less than the threshold, the candidate update means 14 performs a partial update process. In the partial update process, only the parts of the previously generated language expression candidates that are affected by the parameter change are updated differentially. Polite expressions, sentence endings, and specific vocabulary are among the items that are updated. This process makes it possible to reduce the processing load while maintaining responsiveness. 【0088】 Specifically, the candidate update means 14 analyzes the scope of influence of parameter changes and identifies the relevant parts within the language expression candidates. For example, if only the politeness axis changes by 0.1 in translation, only polite expressions such as "please" and "would you" will be updated. If the style axis changes, vocabulary selection (for example, from "check" to "confirm," or from "about" to "regarding") will be updated. 【0089】 The candidate update means 14 is designed to maintain a target latency (e.g., within 100 milliseconds) from parameter change to screen update, regardless of whether it performs a full regeneration or a partial update. In particular, by using partial update processing, it is possible to ensure real-time responsiveness even to continuous parameter changes while reducing the number of tokens generated on the server side and the amount of data sent and received between the client and server. 【0090】 Furthermore, the criteria for selecting between full regeneration and partial update are not limited to these threshold judgments. For example, it is also possible to set different weights for each axis and make a decision based on the weighted change. 【0091】 Furthermore, it is possible to configure the system to change the judgment criteria according to the type of axis that has changed (for example, a configuration in which a change in the writing style axis requires a complete regeneration, but a change in the politeness axis can be handled with a partial update). In addition, it is possible to configure the system to dynamically adjust the judgment criteria according to the processing load and network conditions. The candidate update means 14 selectively performs a complete regeneration and a partial update using one or a combination of these judgment criteria. 【0092】 The auxiliary information generation means 15 compares the original text with the generated candidates and generates auxiliary information including nuance information by performing morphological analysis, syntactic analysis, semantic analysis, tone analysis, and honorific appropriateness analysis. 【0093】 Here, the auxiliary information includes at least one or more of the following information: style difference score, politeness level, honorific appropriateness score, honorific correction proposal, intonation index, sentence length ratio, vocabulary change list, and emotional change index. 【0094】 More specifically, the style difference score quantifies the difference in style between the original text and the candidate. For example, in the range from 0 to 1, 0 represents casual and 1 represents formal. The politeness level is an index indicating the degree of honorific expressions in the candidate. For example, it is evaluated on a 5-level scale (no honorifics, basic polite language, business polite language, high-level honorifics, maximum honorifics). 【0095】 The honorific appropriateness score evaluates the appropriateness of honorific expressions in the candidate. More specifically, it detects and evaluates the detection of excessive honorifics (double honorifics such as "ossyarareru", "mouwasete itadakuku", etc., excessive polite language such as "omeshiagari ni narareru", etc.), the detection of insufficient honorifics (expressions that do not reach the required honorific level in a business scenario, for example, expressions such as "riyou shimashita", "wakari mashita" in an email to a business partner), the detection of inappropriate honorific forms (confusion between humble and respectful language, inappropriate use of "sasete itadakuku", expressions considered inappropriate in business), and the presentation of the recommended honorific level (the range of the recommended honorific level suitable for the current business scenario and the honorific level of the current candidate). 【0096】 The honorific correction proposal proposes appropriate alternative expressions for the detected inappropriate honorific expressions. For example, "sasete itadakuku" → "itashimasu", "ossyarareru" → "ossyaru", "riyou shimashita" → "shouchi shimashita", "mouwasete itadakuku" → "mouwaimasu", "omeshiagari ni narareru" → "meshiagaru", etc., but it is not limited to these. 【0097】 The intonation index is an index indicating the intonation (assertive, neutral, euphemistic) of the candidate. It is calculated based on the end expression and the usage frequency of auxiliary verbs. The sentence length ratio is the ratio of the number of characters or words in the original text and the candidate. For example, when the original text has 50 characters and the candidate has 80 characters, the sentence length ratio is 1.6. 【0098】 The vocabulary change list contains information on vocabulary substitutions from the source text to the candidate words. In particular, in translation, it shows the correspondence between the source words / phrases and their corresponding translated words. The sentiment change index represents the difference in emotional tone between the source text and the candidate words. It quantifies the degree of positivity / negativity and the intensity of emotion. 【0099】 These supplementary information items are represented as numerical scores, bar graphs, indicators, annotation text, color coding, etc. For example, the level of politeness can be displayed as a bar graph, and vocabulary changes can be shown by color coding the corresponding parts in the original text and the candidate text. Information regarding the appropriateness of honorific language is presented in the form of warnings, cautions, and correction suggestions, either as text or icons. 【0100】 The auxiliary information generation means 15 updates the auxiliary information each time the candidates are updated. This allows the user to grasp the nuances that change with parameter manipulation in real time. In particular, even if the user mistakenly believes that they are using appropriate honorifics, the auxiliary information allows them to obtain an objective evaluation. 【0101】 The manner in which supplementary information is presented varies depending on the situation. For example, in addition to presenting detailed supplementary information at all times, it may also include presenting supplementary information selectively upon user request, or presenting only a portion of the supplementary information. 【0102】 The presentation control means 16 synchronously updates and presents candidates and auxiliary information based on the output from the candidate generation means 13 and the auxiliary information generation means 15 while a single operation context is ongoing. 【0103】 The display area may be any of the upper area 41, lower area 42, left area 43, or right area 44 of the screen, as shown in Figure 2. It may also be an overlay window or an external display. The display control means 16 is designed to operate independently of the display position or display format. 【0104】 The presentation control means 16 synchronizes the presentation of candidates and auxiliary information by coordinating their update timing. More specifically, when a new candidate is supplied from the candidate generation means 13, the presentation control means 16 waits for the corresponding auxiliary information to arrive from the auxiliary information generation means 15 before simultaneously displaying both on the screen. This prevents inconsistencies between candidates and auxiliary information. 【0105】 The display control means 16 has a function to control the update frequency. For example, if parameters are changing rapidly, updating the screen for each change may cause the display to flicker, impairing the user experience. Therefore, the display control means 16 can adjust the update interval within a range of 16 milliseconds to 100 milliseconds to achieve a smooth display. 【0106】 In this embodiment, the processing sequence within a single operation context is executed as follows, as shown in Figure 3. 【0107】 First, user input 51 occurs. Next, the parameter determination means 12 updates the language parameters 52. Based on the updated parameters, the candidate generation means 13 generates language expression candidates 53. In parallel with this, the auxiliary information generation means 15 generates auxiliary information 54. Finally, the presentation control means 16 displays these 55. 【0108】 This processing cycle runs continuously, following user actions, as long as the single operation context persists. This allows the user to see changes in language expression in real time, synchronized with their actions. 【0109】 The execution interval of processing cycles should ideally be set within the range of 16 to 100 milliseconds, considering both system processing power and user experience. This range is based on the characteristics of human visual feedback perception. An update frequency of 60 Hz (approximately 16 milliseconds) can provide a smooth visual experience. Furthermore, a response time of 100 milliseconds or less represents the upper limit at which users can perceive the causal relationship between their actions and feedback. 【0110】 Here, the entire processing cycle is executed integrally within a single operation context. In conventional technology, the input phase and the conversion phase were separated, making it impossible to realize such a continuous processing cycle. In this invention, by introducing the concept of a single operation context, a series of processes—operation input → parameter update → candidate generation → auxiliary information generation → display—can be repeatedly executed as long as the user's operation continues. 【0111】 As shown in Figure 4, the processing flow of the operation reception means 11 is as follows. In step 21, when the operation reception means 11 detects an operation start event, it starts a single operation context. Here, an operation start event is an event such as a touchdown, mouse button press, gesture start, or voice input start. 【0112】 In step 22, the operation reception means 11 acquires a series of operations. The acquired operation information is supplied to the parameter determination means 12 at predetermined intervals (for example, every 16 milliseconds). 【0113】 In step 23, the operation reception means 11 determines whether to continue the operation. If the operation is interrupted for a predetermined time (e.g., 500 milliseconds) or if an explicit termination event (touch-up, mouse button release) is detected, the process proceeds to step 24. 【0114】 In determining the continuation of a single operation context, the operation interval threshold is not limited to a fixed value, but may be dynamically changed depending on the type of operation, the user's operation pattern, or the characteristics of the application. For example, an adaptive configuration can be used, where the threshold is set shorter for users who perform fast operations and longer for users who perform slow operations. 【0115】 Furthermore, the context can be configured to terminate when an explicit termination operation (e.g., double-tap, specific gesture) is detected, or to terminate the previous context and start a new context when new source text input begins. 【0116】 In step 24, the single operation context ends, and the presentation control means 16 confirms and presents the final candidate. The user can output the candidate to an input field or an external application by selecting the presented final candidate. 【0117】 In this processing flow, steps 22 and 23 are repeatedly executed while a single operation context is maintained. During this time, a series of processes—parameter determination, candidate generation, auxiliary information generation, and presentation—are performed sequentially. 【0118】 As shown in Figure 5, the parameter space is constructed as a multidimensional space including the style axis 31, the politeness axis 32, the tone axis 33, and the sentence length axis 34. Furthermore, including the emotion axis and the target audience axis, it becomes a space of six or more dimensions. 【0119】 The stylistic axis 31 represents a continuous range from casual to formal. In translation, for example, in English, a stylistic axis value of 0.0 corresponds to colloquial and casual expressions (e.g., "wanna", "gonna"). A value of 0.3 corresponds to everyday expressions, and a value of 0.7 corresponds to business expressions. A value of 1.0 corresponds to formal document expressions (e.g., "therefore", "hereby"). 【0120】 The politeness axis 32 represents the degree of politeness. In translation, for example in English, a value of 0.0 on the politeness axis corresponds to a direct expression (e.g., "Check this"). 0.3 corresponds to a basic polite expression (e.g., "Please check this"). 0.7 corresponds to a very polite expression (e.g., "Could you please check this"). 1.0 corresponds to the most polite expression (e.g., "I would be most grateful if you could kindly check this"). 【0121】 The tone axis 33 represents a continuous range from assertive to indirect. In translation, for example, in English, a tone axis value of 0.0 corresponds to an assertive expression (e.g., "This is correct"), 0.5 corresponds to a neutral expression (e.g., "This appears to be correct"), and 1.0 corresponds to an indirect expression (e.g., "This might possibly be correct"). 【0122】 The sentence length axis 34 represents a continuous range from concise to detailed. In translation, a sentence length axis value of 0.0 indicates a concise translation containing only the minimum necessary information. A value of 0.5 indicates a standard level of detail. A value of 1.0 indicates a detailed translation including background explanations and supplementary information. 【0123】 For example, if the sentence length axis is adjusted towards shortening (towards 0.0) in translation, the expression "I would like to confirm the details regarding the upcoming meeting scheduled for next week" will be shortened to "Please confirm the meeting details." Conversely, if it is adjusted towards detail (towards 1.0), a more detailed expression with background explanations and supplementary information will be generated. 【0124】 The Emotion Axis 35 represents the intensity of the emotional tone contained in the expression. In translation, for example in English, an Emotion Axis value of 0.0 corresponds to a neutral expression with suppressed emotion (e.g., "I understand the situation"). An Emotion Axis value of 0.5 corresponds to an expression with moderate emotion (e.g., "I really appreciate your help"). An Emotion Axis value of 1.0 corresponds to an expression with strong emotion (e.g., "I'm absolutely thrilled about this opportunity!"). 【0125】 The target audience axis represents the adjustment of expression according to the attributes of the intended audience. In translation, a target audience axis value of 0.0 results in expression that heavily uses specialized terminology for experts. 0.5 results in expression suitable for general business professionals. 1.0 results in simple expression suitable for general consumers and non-experts. 【0126】 These six axes represent typical configurations of the multidimensional parameter space in the present invention. However, the present invention is not limited to these axes. Depending on the situation, additional axes such as industry specialization, regionality, and generational fit can be added. Furthermore, it is possible to use only some of the axes for specific applications. 【0127】 Each axis can be controlled independently, and it is also possible to change multiple axes simultaneously. For example, the style axis and the politeness axis can be controlled simultaneously by a two-dimensional slide operation on a continuous operation interface such as a touch panel. Furthermore, it is possible to configure the system to control a third axis (e.g., the tone axis) using pressure intensity. In addition, by associating the speed of the gesture with the sentence length axis, four-dimensional simultaneous control is possible. 【0128】 In the multidimensional parameter space, any point represents a specific nuance of a linguistic expression. Through a series of user operations, points in this space are continuously moved. The candidate generation means 13 generates linguistic expression candidates corresponding to the moved points. This allows the user to achieve multidimensional nuance adjustment through a single series of operations. 【0129】 The parameter determination means 12 continuously updates the values ​​of these axes according to the trajectory and speed of the operation. Based on the parameters at that time, the candidate generation means 13 generates or updates candidates. 【0130】 This paper describes various implementation forms of user interfaces and presentation areas for input processing devices. The present invention is not limited to a specific user interface and can be implemented in various forms. 【0131】 As shown in Figure 6, it can be implemented as an IME at the operating system (OS) level. The functions of the input processing device 10 are provided as an IME module. When the user activates the IME in the text input field, the operation reception means 11, parameter determination means 12, candidate generation means 13, candidate update means 14, auxiliary information generation means 15, and presentation control means 16 start operating. 【0132】 In IME implementations, language expression candidates and auxiliary information are presented as a candidate window. The candidate window can be located near the text cursor, at a fixed position at the bottom of the screen, or at any position set by the user. 【0133】 The advantage of implementing an IME lies in the fact that the functions of the present invention can be used in any application compatible with the OS. A consistent user experience is provided in any application that accepts text input, such as email clients, web browsers, word processors, and messaging applications. 【0134】 As shown in Figure 7, it can also be implemented as a standalone application. In this implementation, the original text input area, operation input area, language expression candidate display area, and auxiliary information display area are arranged within a dedicated application window. 【0135】 Furthermore, in standalone application implementations, there is a high degree of flexibility in screen layout. 【0136】 For example, the screen can be split horizontally, with the original text input area on the left and the language expression candidate display area on the right. Alternatively, the screen can be split vertically, with the original text and candidates displayed side-by-side at the top and detailed supplementary information at the bottom. 【0137】 Independent application implementations can provide richer auxiliary information and more advanced user interfaces. For example, they can include an operation panel that visually represents a multidimensional parameter space, allowing users to intuitively adjust parameters. 【0138】 It can also be implemented as a web application that runs on a web browser. Using web technologies such as HTML, CSS, and JavaScript, the functions from the operation reception means 11 to the presentation control means 16 are realized on the browser. 【0139】 In web application implementations, the language model used by the candidate generation means 13 is typically located on a server device. The client-side browser sends operation information and parameters to the server and receives and displays language expression candidates and auxiliary information from the server. 【0140】 The web application implementation allows users to utilize the functions of the present invention using only a web browser, without installing any software. Furthermore, by using a large-scale language model on the server side, high-quality language expression candidates can be generated regardless of the processing power of the client device. 【0141】 The present invention can also be implemented as a mobile application for smartphones and tablet devices. In this case, an intuitive operation interface using a touch panel is provided. 【0142】 In mobile application implementation, multidimensional parameter control is possible by leveraging the characteristics of touch operation. For example, in a slider area on the screen, it is possible to simultaneously control the style axis with a horizontal swipe and the politeness axis with a vertical swipe. It is also possible to control the sentence length axis with a pinch gesture and the emotion axis with the intensity of a long press. 【0143】 In mobile application implementation, screen size constraints necessitate careful consideration of how auxiliary information is displayed. For example, a configuration could be implemented where only language expression candidates are displayed under normal circumstances, and auxiliary information is expanded and displayed only when the user performs a specific action (e.g., double-tap). 【0144】 The presentation area for language expression candidates and auxiliary information by the presentation control means 16 can be positioned in various locations depending on the situation. For example, as an inline display, the candidate window can be displayed directly below or near the text cursor. 【0145】 Next, as a side panel display, candidates and auxiliary information can be displayed in a panel fixed to the left or right side of the screen. Subsequently, as an overlay display, a semi-transparent overlay window can be displayed on top of the current application window. This form is useful in IME implementations when displaying more detailed auxiliary information. 【0146】 Furthermore, as an external display, candidates and auxiliary information can be displayed on an external display or secondary monitor separate from the main display. This configuration is useful when you want to secure a larger workspace during translation or document creation work. 【0147】 These variations in presentation areas can be selected or combined according to the user's preferences and work environment. Furthermore, they are not limited to any specific presentation area. 【0148】 It should be noted that the correspondence between the operating means and the language parameter axes shown here is merely one example. Here, an example is shown where horizontal movement is mapped to the style axis, but the same operation can also be mapped to the politeness axis. The correspondence between the operating means and the parameter axes can be arbitrarily changed depending on the application settings, user preferences, or device characteristics, and the present invention is not limited to any specific correspondence. 【0149】 Here, we describe the configuration of multidimensional nuance adjustment via an intermediate language. This is a nuance-controlled generation technique for linguistic expression that is applicable to both multilingual translation and conversion of style and honorific levels within the same language. In particular, by using an intermediate language, the naturalness and accuracy of the translation can be improved. 【0150】 The candidate generation means 13 performs a process of translating the source text into an intermediate language. The intermediate language is a language used in translation from the source text to the target language to abstract and normalize the semantic structure, emotional intensity, and formality. 【0151】 Natural languages ​​such as English, French, Spanish, Simplified Chinese, Traditional Chinese, and Korean can be used as intermediate languages. However, these are not the only intermediate languages ​​that can be used; abstract representations generated by multilingual models can also be used as intermediate languages. 【0152】 During the translation process into an intermediate language, the semantic core of the source text is clarified. For example, translating the Japanese phrase "Arigato! Takkatta yo!" into English results in "Thank you! That really helped!". In this process, ambiguities and non-standard expressions unique to Japanese are normalized to a standard semantic structure in the intermediate language. 【0153】 The following factors are considered when selecting an intermediate language: Firstly, the amount of training data available for the language model is abundant. Secondly, the syntactic structure of the language is relatively clear, resulting in low semantic ambiguity. Thirdly, the language is likely to function as a reference point in a common semantic space within a multilingual model. For example, English is widely suitable as an intermediate language because it has the most abundant training data for many multilingual models and its syntactic structure is relatively clear. However, the selection of an intermediate language is determined based on the combination of source and target languages, the characteristics of available language models, and processing efficiency, and is not limited to a specific language. 【0154】 After being translated into the intermediate language, the candidate generation means 13 adjusts the nuances of the intermediate language expression based on the multidimensional language parameters supplied by the parameter determination means 12. For example, consider the case where English is used as the intermediate language and the style axis is adjusted toward casual. The intermediate language expression "Thank you! That really helped!" is adjusted as follows depending on the value of the style axis: If the style axis is 0.7, it becomes "Thanks! Really helpful!" If the style axis is 0.5, it becomes "Thanks a ton! You're a lifesaver!" If the style axis is 0.3, it becomes "Dude, thanks! Total lifesaver!" 【0155】 In this way, by adjusting nuances in the interlanguage, precise nuance control becomes possible within the natural expressive space of the target language. Since there is no need to rewrite the original text with extreme expressions as in conventional techniques, there is less confusion in meaning, and more natural and appropriate expressions can be obtained. 【0156】 The adjusted intermediate language expression is provided to the user in one of at least three output modes: output in the intermediate language as is, output after translation into the source language, or output after translation into a third target language. 【0157】 Outputting in the intermediate language directly outputs the adjusted intermediate language expression as a language expression candidate. For example, if the source language is Japanese and the intermediate language is English, the adjusted English expression will be presented to the user as is. This mode is more effective when the user wishes to communicate in the intermediate language (e.g., English). 【0158】 The output after back-translation to the source language is a method of outputting the adjusted intermediate language expression by back-translating it back to the source language. For example, if the source language is Japanese and the intermediate language is English, the user is presented with the result of back-translating the adjusted English expression back into Japanese. This method is particularly effective for stylistic conversion within the same language. For example, if the Japanese original text "Thank you! You saved me!" is translated into English, adjusted in a more casual style in English, and then back-translated back into Japanese, expressions such as "Thank you! You really saved me!" or "I'm so grateful! You saved me!" can be obtained. Compared to directly rewriting the original text as "I'm so grateful! You saved me!" and translating it, these expressions have less semantic variation and result in more natural Japanese expressions. 【0159】 The output after translation into a third target language is produced by translating the adjusted intermediate language expression into a third target language that is different from both the source language and the intermediate language. For example, if the source language is Japanese, the intermediate language is English, and the target language is French, the user will be presented with the result of translating the adjusted English expression into French. This method is effective in multilingual translation. By using an intermediate language, the accuracy of nuances can be improved compared to direct translation from the source language to the target language. 【0160】 These output formats include, but are not limited to, the three forms described above. For example, a configuration that presents multiple output formats to the user simultaneously is also possible. The selection of the output format can be determined automatically by the user's selection or by the application's settings. 【0161】 Multidimensional nuance adjustment via an interlanguage offers the following technical advantages. First, the accuracy of nuance adjustment is improved through the normalization of semantic structure. Non-standard and ambiguous expressions contained in the source text are normalized to a standard semantic structure during the translation process into the interlanguage. By performing nuance adjustment on this normalized semantic structure, the predictability and consistency of the adjustment results are improved. 【0162】 Next, it becomes possible to control nuances within the natural expressive space of the target language. By adjusting nuances in the interlanguage, variations in expression are generated that remain within a natural range for the target language. Conventional methods of rewriting the original text into extreme expressions tended to result in extreme and unnatural expressions in the target language as well, but this problem does not occur with the interlanguage method. 【0163】 Furthermore, intermediate languages ​​function as bridges between multiple languages. For example, in translation from Japanese to Korean, using English as an intermediate language can sometimes yield more natural translation results than direct translation between Japanese and Korean. This is due to the abundance of training data available for multilingual models that include English. 【0164】 Furthermore, this embodiment is a linguistic expression nuance control generation technology applicable to both multilingual translation and conversion of style and honorific levels within the same language. In particular, real-time continuous adjustment of translation results in multilingual translation is one of the main application areas of this embodiment. 【0165】 When translating the Japanese phrase "Let me confirm the details regarding the meeting" into English, in this embodiment, the user can switch between translation candidates with nuances such as "I want to check about the meeting" when the stylistic axis is 0.3 and "I would appreciate it if I could confirm the details regarding the meeting" when the stylistic axis is 0.7, through a single sequential operation. 【0166】 The Japanese sentence "I would like to buy this product" can be transformed from "I would like to buy this product" (no honorifics) to "We would be grateful if you would purchase this product, and we sincerely request your cooperation" (highest level of honorifics) by adjusting the level of politeness. 【0167】 Furthermore, it can be implemented in either a standalone configuration or a client / server configuration. In the client / server configuration, the operation reception means 11, parameter determination means 12, and presentation control means 16 operate on the client device, while the candidate generation means 13, candidate update means 14, and auxiliary information generation means 15 operate on the server device. 【0168】 Furthermore, it can be implemented as an input processing program that causes a computer to execute the functions of each means. The input processing program can be recorded on a computer-readable recording medium and distributed. 【0169】 Comparing the discrete style selection UI of existing translation services (such as DeepL and Google Translate) with the configuration integrating the continuous parameter control function of the present invention, in this embodiment, the discrete value acquisition unit acquires discrete style selection values ​​from the existing UI, and the point selection parameter determination means sets these discrete values ​​as predetermined points (base points) in a multidimensional parameter space. Subsequently, the relative movement amount from the base point is calculated in response to the user's continuous operations, and the points in the parameter space are continuously updated. This realizes an intuitive two-stage workflow in which the general direction is determined by discrete selection and the fine nuances are refined by continuous adjustment. 【0170】 This embodiment describes a method for updating multidimensional parameters using only natural language instructions, without requiring the user to operate a graphical UI such as sliders. 【0171】 The system configuration consists of a natural language instruction receiving unit, an instruction analysis unit, a parameter mapping unit, and a candidate generation unit. More specifically, the natural language instruction receiving unit receives instructions from the user in natural language (e.g., "more politely," "speak angrily," "make it sound like a research paper," etc.) as text or audio. 【0172】 The instruction analysis unit analyzes natural language instructions using a large-scale language model (LLM) or the like, and extracts the target parameter axis, direction of change, and amount of change. The parameter mapping unit converts the analysis results into continuous quantities or predetermined points in a multidimensional parameter space. The candidate generation unit generates language expression candidates based on the source text and the mapping results in the parameter space. 【0173】 Next, let's explain an example of processing using these system configurations. For example, if the instruction is "Make this more like an academic paper, but not too strict," the parameters will be judged as follows: writing style becomes more academic (+0.7), tone becomes less strict (-0.3), and sentence length becomes slightly longer (+0.2). As a result of the processing, a candidate is generated in which the colloquial expressions of the original text are converted into academic expressions, and the assertive tone is softened. 【0174】 Next, a specific algorithm example for the partial update process in candidate update means 14 is shown. When the parameter change amount of the sentence length axis is less than the threshold, a partial update is performed using the following algorithm. 【0175】 First, the candidate sentence T is parsed to extract the set of modifying phrases M = {m1, m2, …, mn}. Next, a deleteability score S_delete(mi) is calculated for each modifying phrase mi. Furthermore, the deleteability score can be calculated using, for example, the following formula. 【number】 【0176】 Next, the modifying phrases are sorted in descending order of S_delete. Then, if the parameter change ΔP_length < 0 (shortening), the modifying phrases are deleted in descending order of S_delete until the target sentence length ratio is reached. 【0177】 If the parameter change on the thoroughness axis is below the threshold, a partial update is performed using the following algorithm. 【0178】 First, the candidate sentence T is morphologically analyzed to extract honorific-related tokens. Next, the honorific dictionary D_keigo is consulted to obtain the current honorific level L_current for each token. Then, the target honorific level L_target = L_current + ΔP_politeness is calculated. Furthermore, the entry closest to L_target is searched for in the honorific dictionary D_keigo, and the token is replaced. The honorific dictionary D_keigo has, for example, the following structure: D_keigo = {"to send": {0.0: "to send", 0.3: "I will send", 0.5: "I will send", 0.7: "I will send", 0.9: "I will send it"}, ...}. 【0179】 This section explains semantic-invariant nuance transformation, which adjusts only the style while preserving the semantic structure. 【0180】 For comparison, let's explain conventional pivot translation: it follows this processing flow: Source text → [Translate] → Intermediate language → [Translate] → Target language. In this method, the intermediate language is used only as a by-product of translation, and semantic structure and style information are processed together. 【0181】 Next, the semantic-invariant nuance conversion of this embodiment will be explained. The following processing flow is executed: Original text → [Semantic structure extraction] → M (Semantic structure) + S (Style information). After that, M is frozen (invariant), and only S is transformed based on parameters to become S'. Subsequently, the process is M + S' → [Reconstruction] → Target language expression. Here, M (Semantic structure) represents the propositional content, factual relationship, and logical structure, and S (Style information) represents the writing style, politeness, tone, emotion, sentence length, and target audience. 【0182】 Furthermore, disentanglement of semantic structure M and style information S involves separating and extracting the semantic structure M and style information S from the source text. Freezing of semantic structure M means that the semantic structure M is not changed at all during parameter adjustment, and independent transformation of style information S means that only style information S is manipulated in a multidimensional parameter space. 【0183】 As a result, in conventional pivot translation, the intermediate language is used as a byproduct of a two-stage translation process: source text → intermediate language → target language. This means that semantic structure and style information are mixed together during processing, and the semantic content changes simultaneously with style adjustments. In this embodiment, however, the semantic structure M and style information S are explicitly separated, and M is frozen during parameter adjustments, thereby suppressing semantic drift caused by style adjustments. 【0184】 Furthermore, by freezing the semantic structure M and independently transforming the style information S, semantic invariance is guaranteed, meaning that even if style axes such as writing style, politeness, tone, emotion, sentence length, and target audience are continuously changed for the same semantic content, the propositional content and factual relationships remain unchanged. This allows for the safe application of style adjustment functions in fields where semantic precision is required, such as contracts, legal documents, and technical documents. 【0185】 Here, we will describe the server-side configuration and client integration configuration in more detail. As shown in Figure 8, this system consists of a client-side input processing unit 10, a UI integration gateway 300, and a server-side LLM service provider server 120. 【0186】 The LLM service provider server 120 is equipped with a parameter receiving unit 121, a prompt mapping unit 122, a language model execution unit 123, and a response generation unit 124. 【0187】 The parameter receiving unit 121 receives a language processing request from the client device that includes a multidimensional parameter vector. The multidimensional parameter vector is represented, for example, as a 6-dimensional vector P = (p_style, p_politeness, p_tone, p_length, p_emotion, p_audience), where each element is a normalized value in the range of 0 to 1. 【0188】 The prompt mapping unit 122 maps the received multidimensional parameter vector to at least one of a prompt, system message, or generation control parameter for a large-scale language model. 【0189】 Furthermore, five approaches are employed as mapping methods: template selection and interpolation, direct parameter embedding, style tagging, style embedding vector interpolation, and model adapter parameter synthesis. 【0190】 The template selection and interpolation method pre-determines multiple prompt templates (e.g., T_casual, T_business, T_formal) and selectively combines or interpolates these templates based on parameter values. For example, if the style axis value is 0.6, it generates prompts that interpolate T_business and T_formal in a 6:4 ratio. 【0191】 The direct parameter embedding method embeds parameter values ​​directly within the system message. For example, you might input, "Please translate the following sentence. Output it with a formality level of 0.7, politeness level of 0.8, and tone of voice neutral (0.5)." 【0192】 The style tag method involves adding multiple style tags (for example, [FORMAL:0.7][POLITE:0.8][NEUTRAL_TONE]) to the beginning or end of a prompt, and the LLM interprets these tags to adjust the style. 【0193】 The style embedding vector interpolation method involves pre-training style embedding vectors corresponding to the stylistic axis, politeness axis, tone axis, emotional axis, etc., and then linearly combining or interpolating multiple style embedding vectors based on the received multidimensional parameter vector. For example, a linear combination of the casual embedding E_casual and the formal embedding E_formal can be used. 【number】 A style vector is generated and provided as additional input to the language model. 【0194】 The model adapter parameter synthesis method prepares a set of model adapter parameters, such as LoRA (Low-Rank Adaptation) and adapter layers, for each of the multiple styles, and then synthesizes multiple adapter parameters with weights based on the received multidimensional parameter vector. For example, the business style adapter A_business and the casual style adapter A_casual are linearly combined according to the values ​​of the style axis. 【number】 Enable it and run the language model. 【0195】 The language model execution unit 123 executes a large-scale language model using mapped prompts, system messages, or generation control parameters to generate candidate language representations. The language model execution unit 123 has the function of dynamically adjusting at least one of the generation sampling temperature (temperature), top_p value (top_p), or repetition penalty (repetition_penalty) based on a multidimensional parameter vector. 【0196】 For example, if the emotion axis value is high (close to 1.0), the sampling temperature is increased (e.g., to 0.9) to generate more diverse expressions, and if the emotion axis value is low (close to 0.0), the sampling temperature is decreased (e.g., to 0.3) to generate more stable output. Also, if the sentence length axis value is high, the repetition penalty is set low to allow for detailed explanations, and if the sentence length axis value is low, the repetition penalty is set high to encourage concise output. 【0197】 The response generation unit 124 transmits the generated language expression candidates to the client device. In addition to language expression candidates, the response may include auxiliary information (such as a style score and politeness level). 【0198】 The user interface integrated gateway 300 is middleware that mediates between the existing application 310 and the LLM service provider server 120, enabling the use of multidimensional nuance adjustment functionality without significantly modifying the existing application. The user interface integrated gateway 300 also includes an input field monitoring unit 321, a discrete input conversion unit, a collaborative API management unit 322, and an insertion control unit 323. 【0199】 The input field monitoring unit 321 monitors the input fields of the existing application 310 and detects the input of source text. The monitoring method employs one or a combination of the following: obtaining the content of the input field using the accessibility API, monitoring the clipboard, detecting keystrokes using keyboard hooks, and monitoring the DOM using browser extensions. 【0200】 The discrete input conversion unit obtains discrete style selection operations (e.g., DeepL's "Formal" and "Casual" buttons) from the existing application 310 and converts the discrete selection into predetermined points in a multidimensional parameter space. For example, if "Formal" is selected, a base point such as P = (0.8, 0.7, 0.5, 0.5, 0.3, 0.5) is set. 【0201】 The linked API management unit 322 sends points in the multidimensional parameter space to the LLM service provider server 120 and obtains language expression candidates. It also accepts continuous operation input from the user (for example, operation of the overlay slider), continuously updates the points in the multidimensional parameter space as relative movement amounts from a predetermined point (base point), and sends the updated parameters to the LLM service provider server 120. 【0202】 The insertion control unit 323 inserts the acquired language expression candidates into the input field or candidate display area of ​​the existing application 310. The insertion method can be one of the following: paste via clipboard, direct input using the accessibility API, character input via keyboard emulation, or DOM manipulation via a browser extension. 【0203】 The multidimensional nuance adjustment system comprises an input processing unit 10 and an LLM service provider server 120. The input processing unit 10 generates a multidimensional parameter vector based on user input and transmits it to the LLM service provider server 120. The LLM service provider server 120 generates language expression candidates based on the multidimensional parameter vector and returns them to the input processing unit 10. 【0204】 This system configuration allows for a division of labor where the client side performs only lightweight operation reception and parameter determination processing, while the server side handles computationally intensive language model processing. This enables the use of high-quality nuance adjustment functions even on devices with limited computing resources, such as smartphones. 【0205】 A typical processing sequence in this embodiment is shown below. First, the user enters the source text in the existing application 310 (for example, DeepL). Next, the input field monitoring unit 321 detects the input of the source text and notifies the UI integration gateway 300. Subsequently, the user presses the "Formal" button in the existing application 310. 【0206】 Furthermore, the discrete input conversion unit converts the "formal" selection into a base parameter P_base = (0.8, 0.7, ...), and the linked API management unit 322 sends the source text and P_base to the LLM service provider server 120. The parameter receiving unit 121 of the LLM service provider server 120 then receives the request. 【0207】 Subsequently, the prompt mapping unit 122 converts P_base into a prompt, the language model execution unit 123 executes the LLM to generate language expression candidates, and the response generation unit 124 sends the language expression candidates back to the client. 【0208】 Then, the insertion control unit 323 inserts the language expression candidate into the candidate display area of ​​the existing application 310, and the user makes fine adjustments by operating the overlay slider (continuous operation corresponding to claim 26). The linked API management unit 322 calculates the relative movement amount ΔP from P_base, 【number】 The calculation is performed. Then, the linked API management unit 322 sends P_new to the LLM service provider server 120 to obtain updated language expression candidates, and these are repeated until the user is satisfied. 【0209】 (Embodiment 2) In this embodiment, the multidimensional nuance adjustment mechanism of the present invention is applied to the dynamic transformation of educational content such as textbooks and reference books. 【0210】 In this embodiment, the explanatory text from a textbook is input as the source text, and educational nuance axes such as "clarity," "amount of examples," "level of specialization," and "target grade level" are defined as components of a multidimensional parameter vector P. By operating sliders or preset buttons corresponding to these axes, learners or teachers can generate in real time candidate language expressions that explain the same concept from different perspectives. 【0211】 For example, the definition of a "function" in mathematics can be transformed from an abstract set-theoretic description to an explanation using familiar examples such as vending machines and game controllers. The auxiliary information generation means may analyze the differences in abstraction levels, specialization levels, and the amount of examples inserted between the original text and the generated explanatory text, and visualize the differences in explanatory content, such as "reconstructing abstract definitions with familiar examples" and "replacing technical terms with everyday language," as annotation text or icons. 【0212】 This allows learners to intuitively understand which parts have been rephrased and how. This embodiment is ideal for textbook publishers, public educational institutions, EdTech companies, and others that need to adjust explanatory texts on demand according to learners' level of understanding and background knowledge. 【0213】 Furthermore, the nuance axes for educational content are not limited to those predefined by the system; textbook editors and teachers may be allowed to add their own axes, such as "exam-oriented" or "inquiry-based learning-oriented." 【0214】 Furthermore, the auxiliary information generation means analyzes the level of specialization and abstraction of existing textbook texts and automatically sets initial values ​​for each axis based on the results, allowing learners to objectively grasp the difficulty level of their own writing and textbook texts, and intuitively manipulate the up and down directions from there. 【0215】 (Embodiment 3) In this embodiment, the multidimensional parameter control and partial update mechanism of the present invention is applied to response generation of a conversational AI assistant. The conversational AI assistant referred to herein also includes a large-scale language model-based chatbot that presents answers and suggestions in a continuous conversational session with a user. 【0216】 In this embodiment, the server-side conversation management unit assigns a uniquely determined session ID to each dialogue session and maintains a reference parameter vector P_session associated with that session ID. Each component of P_session includes at least axes such as "degree of consideration," "degree of criticism," "conciseness," "degree of empathy," "formality," and "expertise," and the initial values ​​are determined by the user or system administrator operating sliders or preset buttons on the settings screen at the start of the session. 【0217】 In each turn of the dialogue, the client-side user interface accepts a turn-specific nuance adjustment amount ΔP_answer, separate from the P_session set per session. For example, after the user reviews the initial response provided by the AI ​​assistant, they specify ΔP_answer by pressing preset buttons such as "more concise," "more strict," or "more empathetic," or by manipulating the corresponding slider. The input processing unit is: 【number】 The final parameter vector P_answer is calculated, and the answer text is regenerated or partially updated based on this P_answer. 【0218】 Here, a threshold determination based on the parameter change amount ΔP may be used to perform a partial update process that replaces only parts such as honorific expressions, sentence endings, and emphasis expressions if ΔP_answer is small, and a full regeneration process that regenerates the entire answer if ΔP_answer is large. This allows for the automatic optimization of the allocation of computational resources depending on whether the goal is to fine-tune only the tone of the response while maintaining the flow of the dialogue, or to significantly rephrase the answer, including its content. 【0219】 The auxiliary information generation means may visualize scores such as "criticalness," "empathy," and "conciseness" for each response as gauges, icons, labels, etc., and may add annotations such as "This response is highly critical" or "This response is not very empathetic and is factually based and concise." This allows users to objectively understand the style tendencies of the AI ​​assistant's responses and readjust parameters as needed. 【0220】 Furthermore, the axes in this embodiment are not limited to those predefined by the system, and may be configured to allow users or system administrators to define additional nuance axes such as "humor level," "tolerance for technical terms," ​​and "intensity of emotional expression." 【0221】 Furthermore, instead of using sliders, users may input instructions in natural language such as "be a little more frank" or "please be quite direct in your feedback." The instruction interpretation means analyzes these natural language instructions and updates each component of P_session or ΔP_answer, allowing for nuance adjustments without users having to be aware of explicit axis names or numerical values. 【0222】 Furthermore, the initial value estimation means may analyze the user's question and past dialogue history to estimate the user's average level of politeness and formality, and then automatically set the initial position of the sliders on each axis. This allows the user to intuitively manipulate how and to what extent they change the nuances of their speech, starting from their usual way of speaking, and to objectively learn their own communication style. [Explanation of symbols] 【0223】 10 Input Processing Unit 11 Operation reception means 12 Parameter determination means 13 Candidate generation means 14 Candidate update means 15 Auxiliary information generation means 16 Presentation control means 20 Original Text 21 Original text input field 22 Translation candidate display section 23 Auxiliary information display section 24 Parameter control section 30 Multidimensional parameter space 31 Stylistic axis 32. The core principle of thoroughness 33 Intonation Axis 34 Sentence length axis 35 Emotion axis 36 Target Audience 40 Language Expression Candidates 41 Upper presentation area 42 Lower presentation area 43 Left side presentation area 44 Right presentation area 50 Supplementary Information 51 Operation Input 52. Language parameter update 53 Language expression candidate generation 54 Auxiliary information generation 55 displays 80 Language Models 120 LLM service provider servers 121 Parameter receiving unit 122 Prompt mapping section 123 Language Model Execution Unit 124 Response generation unit 300 User Interface Integration Gateway 310 Existing applications 320 Application Integration Department 321 Input field monitoring unit 322 Integration API Management Department 323 Insertion Control Unit

Claims

[Claim 1] An input processing device for inputting original text, A parameter mapping means analyzes operation information, which includes one of the following: a continuous operation, a discrete operation, or a natural language instruction performed by the user, and associates the operation information with a continuous quantity in a multi-dimensional parameter space composed of attributes of style, politeness, tone, emotion, sentence length, and a linguistic expression representing at least one of the target audience, a predetermined point, or a weighted combination of multiple reference points. A candidate generation means generates language expression candidates, including translation candidates or stylistic transformation candidates, by supplying the aforementioned source text and parameter values, which are the result of the correspondence within the parameter space, to a language model as part of a prompt. The system includes a presentation control means that updates and presents the candidate language expression according to the parameters associated by the parameter matching means, The parameter mapping means includes means for determining a plurality of style reference points and calculating parameter values ​​as weighted combinations between the plurality of style reference points based on user operation information. The weighted combination is performed as one or more combinations of the following: linear combination of multiple prompt templates, interpolation of multiple style embedding vectors, synthesis of multiple model adapter parameters, or weighting of multiple style tags. Input processing device. [Claim 2] A parameter receiving unit that receives a language processing request including a multidimensional parameter vector from a client device, A prompt mapping unit that maps the multidimensional parameter vector to at least one of a prompt, system message, or generation control parameter for a large-scale language model by predetermining multiple prompt templates according to the values ​​of each dimension of the parameter vector, and generating prompts by selectively combining or interpolating the templates based on the parameter values, A language model execution unit that executes the large-scale language model using the mapped prompt, the system message, or the generation control parameters and generates language expression candidates, The system includes a response generation unit that transmits the generated language expression candidates to the client device. LLM service provider server. [Claim 3] An LLM service provider server according to claim 2, The prompt mapping unit generates the prompt by selectively combining or interpolating multiple prompt templates according to the values ​​of each dimension of the multidimensional parameter vector. LLM service provider server. [Claim 4] An LLM service provider server according to claim 2 or 3, The language model execution unit dynamically adjusts at least one of the sampling temperature, top P value, or iteration penalty during generation based on the multidimensional parameter vector. LLM service provider server. [Claim 5] The input field monitoring unit monitors the input fields of an existing application and detects the input of source text, A discrete input conversion unit that obtains discrete style selection operations from the aforementioned existing application and converts the discrete selections into predetermined points in a multidimensional parameter space, The LLM service provider server includes: a parameter receiving unit that receives a language processing request including a multidimensional parameter vector from a client device; a prompt mapping unit that predetermines a plurality of prompt templates according to the values ​​of each dimension of the parameter vector, and generates prompts by selectively combining or interpolating the templates based on the parameter values, thereby mapping the multidimensional parameter vector to at least one of prompts, system messages, or generation control parameters for a large-scale language model; a language model execution unit that executes the large-scale language model using the mapped prompts, system messages, or generation control parameters and generates language expression candidates; and a response generation unit that sends the generated language expression candidates to the client device; and an interoperation API management unit that transmits points in the multidimensional parameter space to the LLM service provider server and acquires language expression candidates; and an insertion control unit that inserts the acquired language expression candidates into the input field or candidate display area of ​​the existing application. User interface integration gateway. [Claim 6] A user interface integration gateway according to claim 5, The aforementioned API management unit further receives continuous operation input from the user, continuously updates the points in the multidimensional parameter space as relative movement amounts from the predetermined point, and transmits the updated parameters to the LLM service provider server. User interface integration gateway.

Images