User-state-adaptive interface-based ai legal consultation and prediction system

Abstract

The present invention relates to a user-state-adaptive AI legal consultation and prediction system comprising: a user state analysis unit for generating a user state vector by analyzing, in real time, an emotional state and a cognitive level of a user from voice information and text information collected from a speech of the user; an adaptive dialogue interface unit for dynamically adjusting the tone of a response and question difficulty on the basis of the user state vector; a universal legal feature extraction unit for extracting, from the collected voice information and text information, universal legal feature metadata, which is not dependent on the legal system of a specific country, and determining whether necessary metadata for each case type is satisfied; and a prediction analysis unit for calculating, on the basis of the extracted metadata, the likelihood of a winning case on legal grounds and the actual probability of bond collection as quantitative indicators independent of each other, wherein the adaptive dialogue interface unit generates an additional question on the basis of whether the general-purpose legal feature metadata is missing.

Description

User State Adaptive Interface-Based AI Legal Consultation and Prediction System

[0001] The present invention relates to artificial intelligence-based legal consultation and prediction technology, and more specifically, to a user state-adaptive interface-based AI legal consultation and prediction system that provides a conversational interface that analyzes the user's emotional state and cognitive level in real time and adapts to them, while simultaneously extracting universal legal feature values ​​that do not directly depend on the legal system of a specific country to predict the likelihood of winning a case and the actual possibility of debt recovery.

[0002] The present invention belongs to a field of technology that fuses Human-Computer Interaction (HCI), Natural Language Processing (NLP), emotion recognition technology, knowledge structuring technology, and predictive analytics technology. In particular, it relates to an integrated artificial intelligence system that combines adaptive interface technology, in which conversational structure and expression methods are dynamically modified in response to user state, and legal analysis technology based on a universal case structure.

[0003] With the recent advancement of artificial intelligence technology, AI-based consultation systems are being actively introduced in the field of legal counseling. Conventional legal consultation systems have primarily focused on keyword matching, standardized question-and-answer sessions, or rule-based analysis grounded in the legal provisions of specific countries.

[0004] However, most existing legal consultation systems provide answers in a standardized format without reflecting the user's emotional state or level of understanding in real time. In particular, despite the fact that legal issues are closely related to the parties' emotional tension, anxiety, anger, or confusion, conventional technology has failed to adequately reflect these dynamic characteristics of Human-Computer Interaction (HCI). As a result, problems arise such as reduced efficiency in information delivery during the consultation process or inefficient collection of additional information.

[0005] Furthermore, many existing systems are designed to rely directly on the legal provisions, case law systems, or litigation types of a specific country. In such cases, since the system structure itself is subordinate to that legal system, it is difficult to apply to the legal systems of other countries, leading to problems that limit international expansion. Additionally, classification structures centered on legal provisions tend to focus on resulting legal classification rather than structuring the essential facts of a case, thereby limiting their ability to provide universal analysis.

[0006] Furthermore, conventional technology tends to focus solely on predicting the likelihood of winning a case or the establishment of legal liability, often failing to separately analyze substantive enforceability, such as whether claims can actually be recovered after a judgment. Consequently, users face the difficulty of distinguishing between the legal probability of winning and the actual feasibility of resolving the issue.

[0007] Furthermore, existing systems often have learning structures limited to case law data from specific countries or jurisdictions, and lack sufficient standardized feature structures for the integrated comparison and analysis of cases across nations. Consequently, there are limitations in normalizing and analyzing case structures across diverse legal systems using common standards.

[0008] Against this backdrop, there is a growing need for a new type of AI-based legal consultation and prediction system that analyzes users' emotional states and cognitive levels in real-time during conversations and provides an interface structure capable of immediately adapting to such changes. Simultaneously, this system should analyze cases based on a universal case structure and feature framework that is not directly dependent on the legal provisions of specific countries, and be able to quantify and distinguish between the legal probability of winning and the actual probability of debt recovery. Furthermore, there is a demand for an scalable analytical structure capable of normalizing and utilizing precedent and case data from multiple countries into a common, universal feature structure.

[0009] The present invention aims to overcome the limitations of conventional technology, which fails to reflect real-time changes in a user's emotional state and cognitive level during a conversation, and to provide an adaptive interface structure capable of dynamically adjusting the tone of response, expression style, and question difficulty by continuously analyzing the user's state even during a conversation.

[0010] The present invention aims to provide a legal analysis structure applicable to various legal systems by moving away from analysis structures directly dependent on the legal provisions or litigation types of specific countries and by characterizing cases based on universal case structures such as monetary transfers, the occurrence of obligations, and non-performance of obligations.

[0011] The present invention aims to solve the problem of conventional technology that fails to clearly distinguish between the legal probability of winning a case and the actual possibility of recovering the debt after a judgment, and to provide the user with independent quantitative indicators for calculating the legal judgment possibility and the possibility of execution or recovery, respectively.

[0012] The present invention aims to overcome the limitations of existing prediction methods that fail to systematically reflect the logical consistency between user statements and the reliability of evidence, and to provide an analysis structure capable of quantifying the consistency of collected information and the reliability level of evidence to reflect in a prediction model.

[0013] The present invention aims to resolve the scalability limitations of learning structures restricted to case data of a specific country or a single jurisdiction, and to provide an scalable database-based prediction structure capable of integrated analysis by normalizing case data from multiple countries into a common universal feature schema.

[0014] The present invention relates to a user state adaptive AI legal consultation and prediction system that interacts based on user voice and text data, implemented by a computing device comprising one or more processors and a memory storing program code executed by said one or more processors, comprising: a user state analysis unit that generates a user state vector by analyzing the user’s emotional state and cognitive level in real time from voice information and text information collected from said utterances; an adaptive conversational interface unit configured to dynamically adjust the tone of response and question difficulty based on said user state vector; a universal legal feature extraction unit configured to extract universal legal feature metadata not dependent on the legal system of a specific country from said collected voice information and text information and to determine whether essential metadata for each case type is satisfied; and a prediction analysis unit configured to calculate the legal probability of winning a case and the substantive probability of debt recovery as mutually independent quantitative indicators based on said extracted metadata; wherein the adaptive conversational interface unit is configured to generate additional questions based on whether said universal legal feature metadata is missing.

[0015] Additionally, the adaptive conversational interface unit is configured to switch the tone, speech rate, or voice intonation of the response to a first control mode when the emotion dimension value included in the user state vector is greater than or equal to a preset first threshold, and to provide the response in a standardized second control mode when the emotion dimension value is less than the first threshold.

[0016] Additionally, the adaptive conversational interface unit is configured to activate a multi-stage question mode that sequentially decomposes and presents multiple legal judgment elements when the cognitive level dimension value included in the user state vector is below a preset second threshold, and to activate an integrated question mode that includes multiple judgment elements when the cognitive level dimension value exceeds the second threshold.

[0017] In addition, the general-purpose legal feature extraction unit is configured to have a set of essential metadata items for each predefined case type, extract relationships between entities and case structures from the collected voice information or text information, and map them to the essential metadata items.

[0018] Additionally, the general-purpose legal feature extraction unit is configured to identify unmet items among the essential metadata items and generate additional information request signals corresponding to the unmet items, and the adaptive conversational interface unit is configured to dynamically reconfigure the question generation priority based on the additional information request signals.

[0019] Additionally, the predictive analysis unit includes: a logical consistency evaluation block that calculates a logical consistency score by determining whether there is a discrepancy in time information or event information between the collected response data; and an evidence reliability scoring block that calculates an evidence reliability score based on the type and number of the collected evidence materials.

[0020] In addition, the predictive analysis unit is configured to form the general-purpose legal feature metadata, the logical consistency score, and the evidence reliability score into a feature vector, and to calculate the legal probability of winning the case as a first quantitative indicator in the form of a percentage by matching the feature vector with a pre-trained historical case database.

[0021] In addition, the prediction analysis unit includes a debt recovery probability calculation block configured to calculate the actual debt recovery probability after winning as a second quantitative indicator by applying weights to each item and summing them up for a plurality of items representing execution probability factors; and the first quantitative indicator and the second quantitative indicator are calculated independently of each other.

[0022] Additionally, the user state analysis unit is configured to repeatedly update the user state vector based on voice information or text information additionally collected during the conversation, and the adaptive conversation interface unit is configured to re-select the interface control mode based on the updated user state vector.

[0023] The present invention relates to a user state adaptive AI legal consultation and prediction method that interacts based on user voice and text data, performed by a computer device, comprising: a step of generating a user state vector including a plurality of emotional dimension values ​​and cognitive level dimension values ​​by analyzing the user’s emotional state and cognitive level in real time from voice information and text information collected from the user; a step of dynamically adjusting the tone of the response and the difficulty of the question by comparing the emotional dimension values ​​and cognitive level dimension values ​​included in the user state vector with a preset threshold value; a step of extracting universal legal feature metadata not dependent on the legal system of a specific country from the voice information and text information, and identifying missing items by determining whether essential metadata items for each case type are satisfied; a step of generating additional questions and reconstructing a conversation path based on the missing items; a step of calculating a logical consistency score and an evidence reliability score based on the universal legal feature metadata; and a step of calculating the legal probability of winning a case as a first quantitative indicator by configuring the universal legal feature metadata, the logical consistency score, and the evidence reliability score into a feature vector and matching it with a pre-trained past case database. It includes a step of evaluating execution feasibility factors to calculate the actual possibility of debt recovery after winning as a second quantitative indicator; and the steps from the generation of the user state vector to the calculation of the first and second quantitative indicators are repeatedly performed according to input information additionally collected during the conversation, and are updated in real-time or near-real-time.

[0024] Additionally, the vector generation step further includes a preprocessing step that, prior to performing natural language processing or sentiment analysis, distinguishes between fixed noise and non-fixed noise included in the speech information and removes background noise by performing adaptive filtering with different filtering parameters according to the type of noise.

[0025] The present invention has the effect of improving the quality of user-customized interaction by dynamically adjusting the response tone, expression style, and question difficulty by reflecting changes in the user's emotional state and cognitive level in real time during conversation.

[0026] The present invention can provide a highly scalable legal consultation and prediction structure applicable to various legal systems by analyzing cases based on a universal case structure and feature schema that is not directly dependent on the legal provisions of a specific country.

[0027] The present invention has the effect of supporting users in making comprehensive decisions by distinguishing between the legal judgment possibility of a case and the actual problem-solving possibility, by separating and calculating the legal probability of winning and the actual probability of debt recovery into independent quantitative indicators.

[0028] The present invention can provide more sophisticated and highly reliable prediction results beyond simple event type classification by quantifying the logical consistency between user statements and the reliability of evidence data and reflecting them in predictive analysis.

[0029] The present invention has the effect of enabling the implementation of an AI-based legal consultation and prediction platform capable of comparing case structures between countries and global expansion by normalizing case data from multiple countries into a common universal feature structure and performing integrated analysis.

[0030] FIG. 1 is a block diagram schematically showing the overall configuration of a user state adaptive AI-based legal consultation and prediction system (10) according to one embodiment of the present invention.

[0031] FIG. 2 is a block diagram showing the detailed configuration of a user state analysis unit (210) according to one embodiment of the present invention.

[0032] FIG. 3 is a block diagram showing the detailed configuration of an adaptive conversational interface unit (220) according to one embodiment of the present invention.

[0033] FIG. 4 is a block diagram showing the detailed configuration of a general-purpose legal feature extraction unit (230) according to one embodiment of the present invention.

[0034] FIG. 5 is a block diagram showing the detailed configuration of a prediction analysis unit (240) according to one embodiment of the present invention.

[0035] Specific details of the embodiments are included in the detailed description and drawings.

[0036] The advantages and features of the present invention and the methods for achieving them will become clear by referring to the embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below but may be implemented in various different forms. These embodiments are provided merely to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention, and the present invention is defined only by the scope of the claims. Throughout the specification, the same reference numerals refer to the same components.

[0037] FIG. 1 is a block diagram schematically showing the overall configuration of a user state adaptive AI-based legal consultation and prediction system (10) according to one embodiment of the present invention.

[0038] FIG. 1 is a block diagram schematically showing the overall configuration of a user state adaptive AI-based legal consultation and prediction system (10) according to one embodiment of the present invention.

[0039] Referring to FIG. 1, a user state-adaptive AI-based legal consultation and prediction system (10) may include a user terminal device (100) and a computing device (200) configured to communicate with the user terminal device (100). The user terminal device (100) may include an input interface unit (110) for collecting voice or text input from the user, an output interface unit (120) for outputting consultation responses and analysis results provided from the computing device (200), and a communication unit (130) for transmitting and receiving data with the computing device (200). Accordingly, the user can input their legal problem situation in voice or text form and receive analysis results and prediction information regarding it.

[0040] According to one embodiment of the present invention, a computing device (200) may include a user state analysis unit (210) that analyzes the emotional state and cognitive level of a user based on information collected from a user terminal device (100), an adaptive conversational interface unit (220) that dynamically adjusts the tone of response, expression method, and difficulty of questions based on the analysis results, a universal legal feature extraction unit (230) that extracts universal legal feature values ​​not dependent on the legal system of a specific country, and a predictive analysis unit (240) that quantitatively calculates the possibility of proceeding with a case, the probability of winning, and the actual possibility of recovering a debt based on the extracted feature values. Additionally, the computing device (200) may further include a communication unit (250) for exchanging data with the user terminal device (100).

[0041] According to one embodiment of the present invention, the user state-adaptive AI-based legal consultation and prediction system (10) aims to provide a conversational interface that adapts to the user's psychological state and level of understanding, going beyond simple legal information retrieval or standardized consultation, and at the same time, to provide an integrated AI legal consultation and decision support service that structures and analyzes universal legal feature values ​​that do not directly depend on country-specific legal provisions and predicts the possibility of resolving the case from various angles.

[0042] According to one embodiment of the present invention, the user state analysis unit (210), the adaptive conversational interface unit (220), the universal legal feature extraction unit (230), and the predictive analysis unit (240) are depicted as functional blocks and do not mean physically independent hardware modules, but may be implemented in the form of program code or instruction sets executed by one or more processors included in the computing device (200) in one embodiment. Specifically, the computing device (200) may include one or more processors and memory, and the memory may store program code, learning models, rule sets, metadata templates, or databases for performing the functions of each unit (210, 220, 230, 240). The processor may be configured to perform user state analysis, adaptive conversational control, universal legal feature extraction, and predictive analysis functions by executing the program code stored in the memory. Additionally, the memory may be embedded in the computing device (200), but in other embodiments, it may be implemented as an external storage device or a cloud-based storage system that communicates with the computing device (200).

[0043] In one embodiment, the universal legal issue extraction unit (230) may be configured to have a predefined universal case type-specific essential metadata template. The universal case type-specific essential metadata template may not be directly dependent on the legal provisions or litigation types of a specific country, and may define feature items that must be secured in correspondence with universal case types such as money transfer, occurrence of obligation, setting of deadlines, expiration of deadlines, non-performance of obligations, etc.

[0044] The universal legal issue extraction unit (230) can determine whether the information collected from the user satisfies each item of the required metadata template and can be configured to identify in real time if there is missing required metadata. Additionally, the adaptive conversation interface unit (220) can be configured to dynamically reconstruct the order of questions or conversation paths to collect the identified missing information in priority.

[0045] Additionally, the predictive analysis unit (240) may be configured to evaluate the logical consistency between collected response data based on the extracted general-purpose legal issue metadata, and to score the reliability of the evidence based on the form or existence of the collected evidence. The predictive analysis unit (240) may calculate the probability of winning the case as a quantitative figure by comparing and matching the logical consistency score and the reliability of the evidence score with a past case database, and furthermore, may be configured to separately evaluate the actual possibility of recovering the debt after winning by reflecting the analysis results regarding the opposing party's enforceable property items and willingness to repay.

[0046] According to one embodiment of the present invention, the past case database may be configured to store case data from multiple countries by normalizing it according to the universal feature schema. Accordingly, even if cases occur in different legal systems, comparison and analysis based on common structural feature values ​​are possible, and the predictive analysis unit (240) may be configured to perform data integration analysis between countries. In addition, new case data may be continuously accumulated and reflected as training data according to the same schema.

[0047] According to one embodiment of the present invention, a user state adaptive AI-based legal consultation and prediction system (10) can be configured to go beyond simply changing the content of the response to user input and to dynamically modify the control structure of the conversation interface itself based on a user state vector. That is, the system can implement a Human-Computer Interaction structure that adapts to the individual user state by reconstructing the entire conversation flow, including the tone of the response, sentence structure, question order, and question difficulty, in real time.

[0048] Specifically, the multidimensional user state vector generated by the user state analysis unit (210) can be utilized not merely as a result of simple emotion classification, but as a set of quantitative parameters for conversation control, and the adaptive conversation interface unit (220) can select or combine and apply one of a plurality of interface control modes, such as an empathy-first mode, an information-centric mode, or a question-first mode, according to the user state vector. Accordingly, different conversation flows and expression structures can be generated depending on the user state for the same legal matter. In addition, the adaptive interface structure can be linked with an essential metadata template derived from the universal legal feature extraction unit (230) and configured to reconstruct the conversation path according to the level of completeness of the information obtained so far.

[0049] According to one embodiment of the present invention, the input interface unit (110) may be configured to provide a plurality of input channels so that a user can input their problem situation in a natural manner. In one embodiment, the input interface unit (110) may include a microphone interface for collecting voice input, a keyboard or touch input interface for collecting text input, and a file upload interface for collecting attachment input. The input collected through the input interface unit (110) is not limited to a one-time query at a single point in time, but may include streaming inputs collected continuously during the conversation, and accordingly, changes in the user's emotional state or cognitive level may be reflected in real time during the conversation.

[0050] In one embodiment, the input interface unit (110) can collect the user's utterance in real-time or near-real-time and transmit it to the computing device (200), and may be composed of multimodal input data including both voice signals (pitch, energy, speech rate, intonation change) and utterance content (speech-to-text conversion result) during the utterance. Additionally, the input interface unit (110) may be configured to accumulate and transmit the updated input along with time information even when the user's text input is continuously updated, such as when the user repeatedly modifies or deletes while inputting a sentence. Accordingly, the user state analysis unit (210) does not simply determine the state based on the initial input, but repeatedly updates the user state vector according to changes in new utterances and text input during the conversation process, and the adaptive conversation interface unit (220) can dynamically change the question difficulty, question order, and expression method in response to the updated user state immediately.

[0051] Additionally, the input interface unit (110) may be configured to collect various forms of data that can support the facts of an event as input. In one embodiment, the data may be text messages or messenger conversation history, emails, remittance or payment history, contracts or agreements, audio files, images or screenshots, etc., and the data may be treated as a general concept of “data that can support the existence of facts” rather than a document format limited to the legal terminology of a specific country. For example, if a user attaches a “messenger conversation capture” while stating that “I lent money but they are not repaying it,” the input interface unit (110) may collect the captured image and transmit it to a computing device (200), and the computing device (200) may be configured to reflect the existence and type of the attached data as a feature value that can be referenced by the general legal feature extraction unit (230) and the predictive analysis unit (240) thereafter.

[0052] In one embodiment, the input interface unit (110) may further include an input assistance function that divides the user's input into event units or assigns tags. For example, when a user describes multiple events simultaneously, the input interface unit (110) may provide an input classification UI that the user can select, such as “Event 1 (provision of money),” “Event 2 (promise of repayment),” and “Event 3 (non-fulfillment),” and this input classification information may be utilized by the universal legal feature extraction unit (230) to construct an event structure. Additionally, it may be configured so that the user can select an input mode such as “brief explanation” or “detailed explanation” according to their level of understanding, and this input mode information may be used as an auxiliary input of the cognitive level evaluation block (213) or as a control parameter of the adaptive conversational interface unit (220).

[0053] Meanwhile, the input interface unit (110) may be configured to collect input including multilingual or dialect / accent, taking into account that the user's input language or expression style may vary by country. In one embodiment, the input interface unit (110) may transmit the user's input language, regional speech characteristics, or time zone information together, which may be utilized in a subsequent step for sentiment analysis by the user state analysis unit (210) or for adjusting the expression style of the adaptive conversation interface unit (220). This configuration may contribute to the present invention providing a universal counseling experience that is not dependent on the legal system of a specific country, as well as on a specific language or cultural sphere.

[0054] Additionally, the input interface unit (110) may be installed in various user terminal environments, such as wearable devices, vehicle infotainment devices, smart speakers, or corporate consulting terminals, and in this case, voice, text, or attached data input may be transmitted to the computing device (200) to perform the same analysis pipeline. In another variation, the input interface unit (110) may further include a function to detect network delay, missing input, or noise situations to induce retransmission or input supplementation in order to ensure the quality of real-time input.

[0055] According to one embodiment of the present invention, the output interface unit (120) may be configured to provide a consultation response, analysis result, and prediction indicator generated by the computing device (200) to a user. The output interface unit (120) may include at least one of a text output interface, a voice output interface, and a visual indicator display interface, and may be configured to dynamically adjust the output format in accordance with the user state.

[0056] In one embodiment, the output interface unit (120) may display a text response generated by the adaptive conversational interface unit (220) on a screen or provide it in the form of speech through text-to-speech. In this case, the intonation, speech rate, voice intensity, or accent of the voice output may be adjusted based on a user state vector generated by the user state analysis unit (210). For example, if the user's emotional state is analyzed as anger or anxiety, the output interface unit (120) may provide a voice response with an empathetic tone by applying a relatively gentle intonation and a stable speech rate. On the other hand, if it is determined to be a rational analysis-centered state, a more neutral and structured explanation-centered voice output may be provided.

[0057] Additionally, the output interface unit (120) may be configured not only to output a simple response sentence but also to distinguish and visually display multiple quantitative indicators calculated by the prediction analysis unit (240). In one embodiment, the output interface unit (120) may display the legal probability of winning the case and the score of the actual possibility of recovering the debt as separate indicators. For example, the probability of winning may be provided in the form of a percentage or a graph, while the possibility of recovery may be provided in the form of a separate score, stage classification, or risk grade. Accordingly, the user can distinguish between the possibility of legal judgment and the possibility of actual resolution and understand them comprehensively.

[0058] According to one embodiment of the present invention, the output interface unit (120) may be configured to reflect and display the updated analysis results in real-time or near-real-time when the prediction results are updated as the conversation progresses. For example, if a user inputs additional evidence or supplements essential metadata such as monetary amounts or repayment deadlines, the probability of winning or recovery probability score calculated by the prediction analysis unit (240) may change, and the output interface unit (120) may be configured to immediately reflect these changes on the screen. In addition, by displaying a comparison of the indicators before and after the change, the impact of providing additional information on the prediction results can be intuitively conveyed to the user.

[0059] In one embodiment, the output interface unit (120) may be configured to provide not only simple result figures but also summary information on the main basis for judgment. For example, it may provide guidance messages such as “Possibility of winning may increase if evidence is reinforced,” “Need to verify enforceable assets,” and “Discrepancy in time information between statements exists.” Such guidance information may be linked with an adaptive conversation interface unit (220) and utilized to reconstruct the conversation flow in a way that induces additional questions or supplementation of information.

[0060] Additionally, the output interface section (120) may be configured to provide results using universal expressions that do not directly rely on the legal terminology of a specific country. For example, instead of procedural names of a specific country, explanations may be provided using legal system-neutral conceptual terms such as “existence of legal enforcement authority,” “grounds for compulsory execution,” or “existence of a document with legal effect.” This configuration may contribute to implementing a consultation interface that is commonly applicable to the legal systems of multiple countries.

[0061] In a variation of the present invention, the output interface unit (120) may have a UI structure that can be changed according to a mobile terminal, web-based interface, voice-only device, or enterprise consultation system environment, and may provide results in the form of a graph, chart, step-by-step progress bar, or interactive card. In addition, it may be configured to improve user understanding through visual aids (color, icon, highlighting, etc.).

[0062] According to one embodiment of the present invention, a user terminal device (100) may include a communication unit (130) for transmitting and receiving data with a computing device (200), and the computing device (200) may also include a communication unit (250) for communicating with the user terminal device (100). The communication unit (130, 250) may be configured to perform bidirectional data exchange through a wired or wireless communication network and may operate in an Internet Protocol-based communication, mobile communication network, short-range wireless communication, or cloud-based network environment.

[0063] In one embodiment, the communication unit (130, 250) is not limited to a simple request-response structure and may be configured to exchange data in a real-time or quasi-real-time streaming manner. For example, while a user's voice utterance is collected through the input interface unit (110), the voice data may be divided into chunks and sequentially transmitted to a computing device (200), and the computing device (200) may perform user state analysis based on this and immediately reflect the results to update conversation control parameters. Accordingly, changes in emotional state or cognitive level may be reflected in the interface control in real-time even during a conversation.

[0064] Additionally, the communication unit (130, 250) may be configured to include not only simple data transmission but also the transmission and reception of control signals. For example, the computing device (200) may transmit control information related to changing the response tone, adjusting the difficulty of the question, or switching the conversation mode to the user terminal device (100) based on the analysis results of the user state analysis unit (210) and the adaptive conversation interface unit (220). The user terminal device (100) may receive the control information and dynamically change the voice synthesis parameters, display format, or user interface structure of the output interface unit (120).

[0065] According to one embodiment of the present invention, the communication unit (130, 250) may be configured to immediately transmit the updated result to the user terminal device (100) when the probability of winning or the actual possibility of debt recovery calculated by the prediction analysis unit (240) changes during the conversation. Accordingly, whenever the user inputs additional information, the prediction indicator is dynamically updated, and the change can be reflected in the user interface in real time. In addition, the communication unit (130, 250) may apply an encrypted communication protocol to ensure the security and integrity of user data and may be configured to include an authentication procedure. This is intended to strengthen data protection in a service environment containing sensitive information, such as legal consultation.

[0066] In a variation of the present invention, the communication unit (250) of the computing device (200) may be configured to communicate simultaneously with a plurality of user terminal devices (100) and may be configured to provide scalable services through a distributed server structure in a cloud environment. Additionally, it may be configured to communicate with an external data server for linkage with case databases of multiple countries.

[0067] According to one embodiment of the present invention, the computing device (200) may be implemented as a server device. For example, the computing device (200) may be configured in the form of an application server, an AI computing server, or a distributed processing server deployed in a data center or cloud environment, and may be configured to process requests received from a plurality of user terminal devices (100).

[0068] In one embodiment, the computing device (200) may include one or more central processing units (CPUs), graphics processing units (GPUs), neural network accelerators, or combinations thereof for performing high-performance computations, and may be configured to perform computationally intensive functions, such as user state analysis, general legal feature extraction, and predictive analysis, on the server side. In particular, AI models including large-scale learning parameters, such as sentiment analysis models, natural language processing models, and historical case matching models, may be managed and executed in a server environment.

[0069] Additionally, the computing device (200) may be implemented as a cloud-based scalable structure and configured to dynamically expand computing resources as the number of users increases. For example, the user state analysis unit (210), the adaptive conversational interface unit (220), the general legal feature extraction unit (230), and the predictive analysis unit (240) may be executed on a single physical server, but in other embodiments, they may be distributed across multiple server nodes and operate as a microservices structure.

[0070] According to one embodiment of the present invention, the computing device (200) is not limited to data from a specific country and can be linked with an integrated database that stores case data from multiple countries by normalizing it according to a standardized universal feature schema. Accordingly, the computing device (200) may have a global expansion structure capable of performing event structure comparison between countries or multinational predictive analysis. However, the present invention is not necessarily limited to a remote server environment, and in other embodiments, it may be executed within the same device as the user terminal device (100), or implemented in an on-premise system or a dedicated server environment within an enterprise. That is, the computing device (200) may be implemented as at least one of a centralized server, a cloud server, a distributed server, or a local computing device.

[0071] According to one embodiment of the present invention, a computing device (200) may be configured to include one or more processors and a memory in which program code executed by said processors is stored, and said program code may be configured to perform the functions of a user state analysis unit (210), an adaptive conversational interface unit (220), a general legal feature extraction unit (230), and a predictive analysis unit (240).

[0072] The user state analysis unit (210) may be configured to analyze the emotional state and cognitive level in real time based on voice information and text information collected from the user terminal device (100), and to generate a user state vector including the results. The adaptive conversational interface unit (220) may be configured to dynamically adjust the tone, speech characteristics, and difficulty of the response based on the user state vector, and in particular, to generate additional questions by referring to whether the universal legal feature metadata generated by the universal legal feature extraction unit (230) is missing. The universal legal feature extraction unit (230) may be configured to generate universal legal feature metadata based on structural event elements that are not dependent on the legal system of a specific country, and to determine whether the required metadata for each event type is satisfied. The predictive analysis unit (240) may be configured to calculate the legal probability of winning the case and the actual probability of recovering the debt as independent quantitative indicators based on the universal legal feature metadata.

[0073] According to one embodiment of the present invention, an adaptive conversational interface unit (220) may be configured to compare an emotion dimension value included in a user state vector generated by a user state analysis unit (210) with a preset first threshold value. If the emotion dimension value is greater than or equal to the first threshold value, the adaptive conversational interface unit (220) may be configured to switch the tone, speech rate, or voice intonation of the response to a first control mode. On the other hand, if the emotion dimension value is less than the first threshold value, the response may be configured to provide a response in a standardized second control mode.

[0074] According to one embodiment of the present invention, the emotion dimension value included in the user state vector may be a probability value or a score value calculated for each of the multiple emotion categories (anger, anxiety, neutral, stability, etc.). In one embodiment, the emotion dimension value may be normalized to a continuous real value between 0 and 1, and may represent a confidence probability for a specific emotion category. The first threshold value is a reference value for determining whether the emotion dimension value is at a level that affects conversation control, and may be a preset value such as, for example, 0.6 or 0.7. For example, if the anger dimension value is calculated as 0.75 and the first threshold value is set to 0.6, the emotion dimension value is determined to be greater than or equal to the first threshold value, and an empathy-centered control mode may be activated. On the other hand, if the anger dimension value is calculated as 0.3, it is determined to be less than the first threshold value, and a standardized neutral control mode may be maintained. In addition, the first threshold value is not limited to a fixed value and can be dynamically adjusted according to user characteristics, conversation context, input data quality, or analysis reliability. For example, if the reliability of emotion estimation is low, the first threshold value may be configured to be raised to prevent excessive mode switching. Accordingly, when the emotion dimension value is greater than or equal to the first threshold value, it means a state in which the intensity of a specific emotion category included in the user state vector exceeds the control mode switching criterion, and when it is less than the first threshold value, it means a state in which the emotion intensity does not reach the control mode switching criterion.

[0075] According to one embodiment of the present invention, an adaptive conversational interface unit (220) may be configured to compare a cognitive level dimension value included in a user state vector with a pre-set second threshold value. If the cognitive level dimension value is less than or equal to the second threshold value, the adaptive conversational interface unit (220) may be configured to activate a multi-stage question mode that sequentially decomposes a plurality of legal judgment elements and presents them. On the other hand, if the cognitive level dimension value exceeds the second threshold value, it may be configured to activate an integrated question mode that includes a plurality of judgment elements.

[0076] According to one embodiment of the present invention, the cognitive level dimension value included in the user state vector may be a quantitative score calculated based on the user's vocabulary level, sentence structure complexity, frequency of use of technical terms, utterance length, or semantic structure diversity. In one embodiment, the cognitive level dimension value may be expressed as a score between 0 and 100, or as a normalized score between 0 and 1. The second threshold value may be set as a reference value for selecting a question difficulty adjustment mode. For example, if the cognitive level score is 40 or less out of 100, it is determined to be below the second threshold value, and the multi-stage question mode may be activated. Conversely, if the cognitive level score exceeds 40, the integrated question mode may be activated. For example, if the user uses a short sentence structure centered on everyday language, such as “I lent them money, but they still haven't paid it back,” the cognitive level score may be calculated as 35, and if the second threshold value is set to 40, the multi-stage question mode may be activated. Accordingly, “Have you set a repayment deadline?” Complex questions can be broken down and presented step by step, such as “Is there any material to prove that content?” On the other hand, if the user uses technical terms and complex sentences, such as “The repayment deadline has been specified, and there is an electronically signed document to prove this,” the cognitive level score may be calculated as 70 points; in this case, the integrated question mode is activated, and multiple judgment elements can be presented as a single question. Furthermore, the second threshold value is not limited to a fixed value but can be dynamically adjusted based on the user’s past conversation history, age group estimation, input speed, or the presence of repetitive errors. Accordingly, a cognitive level dimension value below the second threshold value signifies a state where the burden of understanding is likely to increase when complex information is presented, while a value exceeding the second threshold value signifies a state where comprehension is highly probable even when multiple judgment elements are presented in an integrated manner.

[0077] According to one embodiment of the present invention, a universal legal feature extraction unit (230) may be configured to have a set of predefined essential metadata items for each case type. The universal legal feature extraction unit (230) may be configured to map the case structure extracted through an input normalization block (231), an entity and role extraction block (232), and a relationship and event extraction block (233) to the essential metadata items using a universal feature schema mapping block (234).

[0078] According to one embodiment of the present invention, a metadata completeness evaluation block (235) may be configured to identify unmet items among the required metadata items. A feature value packaging and output block (236) may be configured to generate an additional information request signal corresponding to the unmet items and transmit it to an adaptive conversational interface unit (220). Accordingly, the adaptive conversational interface unit (220) may be configured to dynamically reconfigure the question generation priority based on the additional information request signal.

[0079] According to one embodiment of the present invention, the predictive analysis unit (240) may be configured to include a logical consistency evaluation block (242) and an evidence reliability scoring block (243). The logical consistency evaluation block (242) may be configured to calculate a logical consistency score by determining whether there is a discrepancy in time information or event information between collected response data. The evidence reliability scoring block (243) may be configured to calculate an evidence reliability score based on the type and number of evidence materials.

[0080] According to one embodiment of the present invention, a predictive analysis unit (240) may be configured to form a feature vector comprising general-purpose legal feature metadata aligned in an input feature reception and normalization block (241), a logical consistency score, and an evidence reliability score. A past case matching block (244) may be configured to match the feature vector with a pre-trained past case database. A winning probability calculation block (245) may be configured to calculate the legal winning probability of a case as a first quantitative indicator in the form of a percentage based on the matching results.

[0081] According to one embodiment of the present invention, the prediction analysis unit (240) may be configured to include an execution possibility factor extraction block (246) and a debt recovery possibility calculation block (247). The debt recovery possibility calculation block (247) may be configured to calculate the actual debt recovery possibility after winning as a second quantitative indicator by applying a weight to each item and summing them up for a plurality of items representing execution possibility factors. The first quantitative indicator calculated in the winning probability calculation block (245) and the second quantitative indicator calculated in the debt recovery possibility calculation block (247) may be configured to be calculated independently of each other.

[0082] According to one embodiment of the present invention, the user state analysis unit (210) may be configured not to determine the user state only at the beginning of the conversation, but to repeatedly update the user state vector based on additional voice information or text information received during the conversation. Specifically, if there is continuous speech or additional input collected through the input interface unit (110), the emotion analysis block (212) and the cognitive level evaluation block (213) may perform re-analysis by reflecting the new input, and the state integration and vector generation block (214) may generate an updated multidimensional user state vector. Additionally, the adaptive conversation interface unit (220) may be configured to receive the updated user state vector in real-time or near-real-time and to re-select or switch the existing interface control mode through the state interpretation block (221) and the response strategy determination block (222). For example, at the beginning of the conversation, the emotional dimension value is high and the empathy-centered control mode is activated; however, if the emotional dimension value subsequently stabilizes and the cognitive level dimension value increases, the adaptive conversation interface unit (220) may be configured to switch to an analysis-centered control mode or an information-centered control mode. Accordingly, the present invention can implement an iterative adaptive (Human-Computer Interaction) structure that does not determine the user state fixedly at a single point in time, but continuously updates the user state vector according to the conversation flow and dynamically reconstructs the interface control structure in response.

[0083] FIG. 2 is a block diagram showing the detailed configuration of a user state analysis unit (210) according to one embodiment of the present invention.

[0084] Referring to FIG. 2, the user state analysis unit (210) may include an input data preprocessing block (211), an emotion analysis block (212), a cognitive level evaluation block (213), and a state integration and vector generation block (214). Hereinafter, to explain an example of the operation of each block, a situation in which a user speaks in an agitated tone, saying, "I lent you money, but you still haven't paid it back. You keep delaying it, and I am so angry," will be described as an example.

[0085] According to one embodiment of the present invention, an input data preprocessing block (211) may be configured to collect voice data or text data received from a user terminal device (100) and to perform normalization and structuring for subsequent analysis. For example, the input data preprocessing block (211) may extract acoustic features such as pitch, energy, and spectral characteristics from voice data or perform voice-to-text conversion, and for text data, may perform tokenization, morphological analysis, syntactic analysis, sentence segmentation, or stop word removal. In addition, the input data preprocessing block (211) may perform background noise removal or abnormal input filtering.

[0086] According to one embodiment of the present invention, an input data preprocessing block (211) may receive voice data from the utterance and extract acoustic features. For example, acoustic characteristics such as an increase in average pitch, an increase in voice energy, and a change in the speed of speech may be detected. Additionally, voice-to-text conversion may be performed to generate text data containing key words such as "money," "borrow," "not pay back," and "angry." The generated text data may be structured through tokenization and parsing.

[0087] According to one embodiment of the present invention, the emotion analysis block (212) may be configured to estimate the psychological state of a user based on preprocessed input data. In one embodiment, the emotion analysis block (212) may calculate score values ​​or probability values ​​for a plurality of emotion categories using acoustic features or text sentiment analysis results, and generate an emotion vector including these values. The emotion vector may subsequently be utilized as a quantitative control parameter for controlling a conversation interface. That is, the emotion analysis block (212) may estimate the emotional state of a user based on preprocessed acoustic features and text data. For example, by analyzing the acoustic characteristics of the utterance and the sentiment word ("I am angry"), score values ​​for anger or anxiety emotions may be calculated. As a result, an emotion vector with a high anger dimension value and a low neutral dimension value may be generated.

[0088] According to one embodiment of the present invention, a cognitive level evaluation block (213) may be configured to estimate a user's level of understanding or vocabulary level based on preprocessed text data. For example, the cognitive level evaluation block (213) may calculate a cognitive level score by analyzing sentence length, vocabulary difficulty, frequency of use of technical terms, syntactic complexity, etc. The cognitive level score may be used as a reference value for adjusting question difficulty or selecting an explanation method. That is, the cognitive level evaluation block (213) may estimate a user's cognitive level by analyzing the difficulty of vocabulary included in the text data, the complexity of sentence structure, etc. For example, if legal technical terms are not used and expressions centered on everyday language are used, a cognitive score of an intermediate or relatively low level may be calculated. The cognitive level score may subsequently be used as a reference value for adjusting question difficulty.

[0089] According to the present invention, the emotion analysis block (212) and the cognitive level evaluation block (213) may operate independently of each other, and each analysis result may be transmitted to the state integration and vector generation block (214). The state integration and vector generation block (214) may be configured to combine a plurality of analysis results, including an emotion vector and a cognitive level score, to generate a multidimensional user state vector representing the user's current state. In one embodiment, the multidimensional user state vector may include a plurality of emotion dimension values, cognitive level dimension values, and confidence values ​​for each estimation result. The generated user state vector may be transmitted to an adaptive conversational interface unit (220) and used as a control signal to dynamically adjust the tone, expression method, question composition, and difficulty of the response.

[0090] That is, the state integration and vector generation block (214) can generate a multidimensional user state vector by combining the emotion vector transmitted from the emotion analysis block (212) and the cognitive level score transmitted from the cognitive level evaluation block (213). For example, the user state vector may be composed of multiple dimension values ​​including an anger dimension value, an anxiety dimension value, a cognitive level score, and confidence for each estimate. The generated user state vector can be transmitted to an adaptive conversational interface unit (220) and used to control the response tone and question composition method.

[0091] Additionally, the input data preprocessing block (211), the sentiment analysis block (212), the cognitive level evaluation block (213), and the state integration and vector generation block (214) do not refer to physically separated hardware, but can be implemented as program code executed by one or more processors included in the computing device (200), and said program code can be stored in the memory of said computing device (200).

[0092] According to one embodiment of the present invention, the user state analysis unit (210) may be configured not to determine the user state only at the beginning of the conversation, but to repeatedly perform the emotion analysis block (212) and the cognitive level evaluation block (213) according to additional utterances received during the conversation. Accordingly, the state integration and vector generation block (214) may continuously update the multidimensional user state vector according to the conversation flow, and the updated user state vector may be transmitted to the adaptive conversation interface unit (220) in real-time or near-real-time and reflected in the interface control.

[0093] Additionally, the state integration and vector generation block (214) may be configured to generate a multidimensional user state vector by reflecting the reliability values ​​included in the sentiment analysis results and the cognitive level evaluation results as weights. For example, if the reliability of the sentiment estimation is low due to low voice data quality or short utterance length, the weights of the corresponding sentiment dimension values ​​may be adjusted to produce an integration vector. Accordingly, the user state vector may be generated not as a simple combination result, but as a dynamically weighted integration result that reflects the analysis reliability and input characteristics.

[0094] FIG. 3 is a block diagram showing the detailed configuration of an adaptive conversational interface unit (220) according to one embodiment of the present invention.

[0095] According to one embodiment of the present invention, a state interpretation block (221) may be configured to receive a user state vector generated in FIG. 2 as input, interpret a plurality of emotion dimension values ​​and cognitive level dimension values ​​included in the vector, and generate conversation control parameters based thereon. Specifically, the state interpretation block (221) may classify the user's current psychological state by comparing the magnitude of each emotion dimension value included in the emotion vector or by comparing it with a predefined threshold value. For example, if the anger dimension value appears higher than a threshold value, a control signal to switch to an empathetic response mode may be generated. Additionally, if the cognitive level score is below a certain threshold, a control parameter to activate an explanation reinforcement mode or a step-by-step question mode may be generated.

[0096] According to one embodiment of the present invention, the state interpretation block (221) may not simply determine the level of emotion or cognition individually, but may produce an interpretation result for determining the direction of a conversation strategy by comprehensively considering a plurality of dimension values. In one embodiment, the state interpretation block (221) may receive a user state vector as input and generate a control signal for selecting one of a plurality of interface control modes, such as an empathy-first mode, an information-centric mode, a question-first mode, or a neutral analysis mode.

[0097] The following is an example of a case where a user speaks in an agitated tone, saying, "I lent you money, but you still haven't paid it back. You keep delaying it, and I am so angry."

[0098] If the anger dimension value is high in the user state vector generated in FIG. 2 and the cognitive level score is calculated to be below the middle, the state interpretation block (221) can generate control parameters to provide an explanation in relatively easy expressions, along with a control signal to activate the empathy priority mode. Accordingly, in subsequent steps, the response can be controlled to start with an expression reflecting the user's emotional state, such as "You may feel very frustrated and angry because you have not received the money you lent," rather than simply listing legal requirements. At the same time, questions related to legal judgment can be controlled to be presented in a stepwise manner rather than being presented in a complex manner.

[0099] According to one embodiment of the present invention, the response strategy determination block (222) may be configured to receive conversation control parameters transmitted from the state interpretation block (221) as input and to select or configure a conversation strategy suitable for the user's current state. Specifically, the response strategy determination block (222) may select at least one of a plurality of conversation strategies, such as an empathy-first strategy, an information-centered strategy, a question-centered strategy, or an analysis-centered strategy, or generate a customized response strategy by combining a plurality of strategy elements. The conversation strategy may subsequently be transmitted to the expression method adjustment block (223) and the question configuration and difficulty adjustment block (224) to be reflected in the specific response generation process.

[0100] In one embodiment, the response strategy determination block (222) may select a strategy that prioritizes the inclusion of empathetic expressions while minimizing the use of complex legal terms when the emotional dimension value included in the user state vector is high and the cognitive level score is calculated to be below a certain threshold. Conversely, when the emotional dimension value is low and the cognitive level score is high, a more structured legal analysis-centered strategy may be selected.

[0101] As previously exemplified, if a user speaks in an agitated tone, saying, "I lent money, but they are not paying it back," and the empathy priority mode is activated in the state interpretation block (221), the response strategy determination block (222) may select a strategy that places an expression of empathy at the beginning of the response and then guides the legal judgment elements step by step. For example, rather than immediately explaining legal requirements, the strategy may be structured to include sentences to soothe the user's emotions first, and then the response flow may be designed to sequentially verify factual relationships such as "whether there is a promissory note" and "whether a repayment deadline has been set."

[0102] According to one embodiment of the present invention, the state interpretation block (221) and the response strategy determination block (222) are not limited to the control mode selected in the initial stage of the conversation, and may be configured to re-select or switch the interface control mode accordingly when the user state vector is updated during the conversation. For example, even if an empathy-first strategy is applied because the anger dimension value is high at the beginning of the conversation, if the user's emotion dimension value stabilizes and the cognitive level score rises thereafter, the response strategy determination block (222) may control to switch to an information-centered strategy or an analysis-centered strategy. Accordingly, the present invention can implement a human-computer interaction structure that dynamically adapts throughout the conversation flow.

[0103] According to one embodiment of the present invention, the interface control mode selected by the state interpretation block (221) and the response strategy determination block (222) is configured not merely to change the expression of the response phrase, but to be directly reflected in the question generation logic, information collection priority, feature value update trigger conditions transmitted to the predictive analysis unit (240), and the control structure of the conversation flow itself. Accordingly, the adaptive conversation interface unit (220) can operate as a dynamic interface control system that reconstructs the control structure of the conversation in real time according to changes in the user state.

[0104] According to one embodiment of the present invention, the expression mode adjustment block (223) may be configured to receive a conversational strategy selected in the response strategy determination block (222) as input and to specifically adjust the expression format of the response to be provided to the actual user. Specifically, the expression mode adjustment block (223) may generate control parameters for adjusting the tone of the response, sentence length, vocabulary selection, level of detail of the explanation, and intonation or speech style of the voice output.

[0105] In one embodiment, the expression method adjustment block (223) may rearrange the sentence structure to include empathetic expressions preferentially or substitute vocabulary to use euphemistic expressions when the user's emotional dimension value is high. Additionally, when the cognitive level score is calculated to be low, it may convert legal technical terms into everyday expressions or reconstruct the sentence into a short and simple structure.

[0106] As previously exemplified, if a user speaks in an agitated tone saying, "I lent money, but they still haven't paid it back," and the response strategy is determined to be an empathy-first strategy, the expression method adjustment block (223) can control the first sentence of the response to be composed of a sentence containing an expression of empathy, such as "You may be very frustrated and upset about the current situation." Additionally, instead of a sentence centered on technical terms such as "We need to review whether the right to claim the return of the loan can be exercised," it can be restructured into an easy-to-understand expression such as "Let's look together into whether there is a way to get the money back."

[0107] In one embodiment, the expression mode adjustment block (223) may refer to speech characteristic information reflecting the user's speech rate, intonation pattern, voice intensity, and regional language characteristics, as well as the emotion dimension value transmitted from the user state analysis unit (210). When the emotional state is determined to be anxiety or anger, the expression mode adjustment block (223) may control the generation of a voice response with an empathetic tone by applying intonation similar to the user's speech pattern or pronunciation and accent corresponding to regional characteristics. On the other hand, when the state is determined to require rational analysis, it may switch to an objective tone by applying standardized pronunciation and neutral intonation.

[0108] According to one embodiment of the present invention, the question configuration and difficulty adjustment block (224) may be configured to configure a question for collecting additional information from a user and to adjust the difficulty level in conjunction with the response strategy determination block (222) and the expression method adjustment block (223). Specifically, the question configuration and difficulty adjustment block (224) may identify missing items among the information collected so far by referring to the essential metadata items provided by the general-purpose legal feature extraction unit (230). Subsequently, a question may be generated to secure the missing items first, and the complexity of the question may be adjusted according to the user's cognitive level score.

[0109] For example, in the preceding embodiment, if the user utters, "I lent money but they are not paying it back," the question composition and difficulty adjustment block (224) can determine that essential feature values, such as the fact of money transfer, whether a repayment deadline exists, or whether a promissory note or evidence exists, have not yet been confirmed. If the cognitive level score is calculated to be relatively low, a simple question regarding a single issue, such as "Did you write a promissory note?", can be presented step by step. On the other hand, if the cognitive level score is calculated to be high, a question containing multiple elements can be composed into a single sentence, such as "It is necessary to confirm whether a repayment deadline has been specified and whether there is evidence to prove this."

[0110] Additionally, the question composition and difficulty adjustment block (224) can control the use of euphemistic expressions instead of direct and pressure-inducing questions when it is determined that the user's emotional state is agitated. For example, instead of "Is there any evidence?", the expression can be adjusted to "Could you check if there is any material remaining that can prove the situation at the time?"

[0111] Additionally, the question composition and difficulty adjustment block (224) may be configured to apply a multi-step questioning method in which multiple legal judgment elements that may be included in a single complex question are broken down step by step and presented sequentially when the user's cognitive level score is determined to be below a pre-set threshold. For example, a complex question such as "It is necessary to confirm whether a repayment deadline has been specified and whether there is evidence to prove it" may be broken down into a first question, "Have you set a separate repayment deadline?" and a second question, "Do you have any evidence to prove that content?" and presented sequentially when the cognitive level is determined to be low.

[0112] Additionally, the question composition and difficulty adjustment block (224) may include vocabulary substitution logic that automatically converts legal technical terms into everyday expressions, for example, converting the term "default" into an expression such as "a state of not paying back promised money." In this way, the block can simultaneously improve the accuracy of information collection and user understanding by dynamically adjusting the structure and vocabulary of the question to suit the user's cognitive level.

[0113] According to one embodiment of the present invention, a response generation block (225) may be configured to generate a final response to be provided to a user based on control parameters transmitted from a representation mode adjustment block (223) and a question composition and difficulty adjustment block (224) and a configured sentence structure. Specifically, the response generation block (225) may generate a text response using a natural language generation model, a rule-based template, or a combination thereof.

[0114] In one embodiment, the response generation block (225) may generate a legal explanation using a pre-trained language model, but may be controlled to reflect the tone and vocabulary level determined in the expression mode adjustment block (223). Additionally, the response may be structured to include a step-by-step question configured in the question composition and difficulty adjustment block (224).

[0115] As previously exemplified, when a user speaks in an agitated tone, saying, "I lent money, but they still haven't paid it back," the response generation block (225) can first generate an introductory sentence containing empathetic expressions, then briefly summarize the main issues of the incident, and then present matters requiring verification in the form of questions. For example, a response with a structure like [Table 1] can be generated.

[0116] - “It must be very frustrating not to receive the money you lent back.” (Expression of empathy) - “It appears that the current situation requires reviewing the procedures to recover the lent money.” (Case summary) - “Did you happen to draw up a promissory note?” (Step-by-step question 1) - “We also need to confirm whether a repayment deadline was set separately.” (Step-by-step question 2)

[0117] The response generated in this way can be transmitted to a user terminal device (100), and if voice output is required, it can be provided through a voice synthesis process. In one embodiment, an intonation or speech style corresponding to the emotion dimension value may be reflected even during voice synthesis.

[0118] According to one embodiment of the present invention, the response generation block (225) is not limited to simply outputting legal information, but can implement a dynamic conversational interface that adapts to the user state by integrating a strategy, expression method, and question structure configured based on the user state vector to generate a final response.

[0119] According to the present invention, the state interpretation block (221), the response strategy determination block (222), the expression method adjustment block (223), the question composition and difficulty adjustment block (224), and the response generation block (225) do not refer to physically separated hardware, but can be implemented as program code executed by one or more processors included in the computing device (200), and said program code can be stored in the memory of the computing device (200).

[0120] FIG. 4 is a block diagram showing the detailed configuration of a general-purpose legal feature extraction unit (230) according to one embodiment of the present invention.

[0121] Referring to FIG. 4, the universal legal feature extraction unit (230) may include an input normalization block (231), an entity and role extraction block (232), a relationship and event extraction block (233), a universal feature schema mapping block (234), a metadata completeness evaluation block (235), and a feature value packaging and output block (236).

[0122] Additionally, the general-purpose legal feature extraction unit (230) may be configured to store and manage case data from multiple countries according to the same general-purpose feature schema. Accordingly, the historical case database is not limited to precedents of a specific country, and can integrate and analyze cases that occurred in different legal systems under a common structure. This provides a technical basis for the predictive analysis unit (240) to learn patterns between countries or to perform comparative analysis based on the structural similarity of similar cases.

[0123] First, the input normalization block (231) may be configured to receive conversation text, speech-to-text conversion results, conversation history information, or time information transmitted from the user terminal device (100), and to perform a normalization process for subsequent feature extraction. Specifically, the input normalization block (231) may perform synonym normalization, tense normalization, indicator interpretation, sentence segmentation, and correction of incomplete sentences. For example, different expressions such as "lent," "rented," and "lent money" can be normalized into the same semantic category, and indicators such as "that person" and "that friend" can be interpreted as specific entities based on the preceding conversation context.

[0124] As previously exemplified, when a user utters, "I lent money, but they still haven't paid it back," the input normalization block (231) can normalize "lent money" as a money transfer event and "not paying back" as a state of non-fulfillment of obligation. Additionally, the temporal order of events can be arranged by considering the time of utterance and the context of the previous conversation.

[0125] Data refined through the input normalization block (231) can be transferred to the entity and role extraction block (232) and the relationship and event extraction block (233), and the two blocks can be configured to operate in parallel independently of each other. First, the entity and role extraction block (232) can be configured to identify key entities related to an event from the normalized text and estimate the role of each entity. For example, the entity and role extraction block (232) can recognize entities such as users, counterparties, third parties, money, contracts, promissory notes, etc. as entities, and through contextual analysis, can classify the roles of the entities as creditors, debtors, agents, etc.

[0126] In the previous example, if the user utters, "I lent them money, but they still haven't paid it back," the entity and role extraction block (232) can identify "me (the user)" as the entity that provided the money and "that person (the counterparty)" as the entity that received the money. Additionally, "money" can be recognized as a money entity, and it can be inferred at the role level that a creditor-debtor relationship has been formed between the user and the counterparty in the event.

[0127] Meanwhile, the relationship and event extraction block (233) may be configured to extract relationships between entities and events that occurred from the normalized text. Here, events may be classified into universally defined event types, such as money transfer, occurrence of obligation, setting of deadline, expiration of deadline, non-performance of obligation, etc., without relying on the legal provisions or offenses of a specific country. For example, the expression "lent money" in the utterance may be extracted as a money transfer event and a debt occurrence event, and the expression "still not paying back" may be extracted as a non-performance of obligation event. The relationship and event extraction block (233) may arrange these events in chronological order and derive a series of event structures such as money transfer -> occurrence of repayment obligation -> non-performance of repayment.

[0128] According to one embodiment of the present invention, a universal feature schema mapping block (234) may be configured to map entity information and event information transmitted from an entity and role extraction block (232) and a relationship and event extraction block (233) to a predefined universal feature schema. Here, the universal feature schema may be defined as a set of structural feature values ​​that are commonly applicable across various legal systems, without directly relying on the legal provisions, charges, or types of litigation of a specific country. For example, the universal feature schema may include categories such as those in [Table 2].

[0129] - Party Role: Creditor, debtor, third party, etc. - Asset or Object: Money, goods, services, etc. - Event Type: Movement of money, occurrence of obligation, setting of deadline, expiration of deadline, non-performance of obligation, etc. - Evidence Presence: Promissory note, bank transfer records, message logs, etc. - Temporal Attribute: Transaction date and time, repayment deadline, etc.

[0130] The universal feature schema mapping block (234) can convert the extracted entity and event information into structured metadata by mapping it to each field of the schema.

[0131] According to one embodiment of the present invention, the universal feature schema may be defined as a neutral metadata structure for normalizing case data from multiple countries into a common structure, rather than a classification system that directly corresponds to the legal system of a specific country. For example, even if an event has the same legal effect in different countries, the legal terminology or procedural names of those countries may differ. Accordingly, the universal feature schema mapping block (234) may be configured to normalize case data for each country into a standard format that allows for comparison between countries by converting it into structural feature values ​​such as party roles, case types, whether obligations occurred, whether obligations were fulfilled, asset types, and time information.

[0132] According to one embodiment of the present invention, the universal feature schema may be defined based on structural event elements such as monetary movement, occurrence of obligation, non-performance of obligation, assertion of rights, and the existence of grounds for enforcement, without relying on specific legal names such as “claim for return of loan” or “claim for damages.” Accordingly, the present invention can implement an analysis structure that is not affected by changes in legal names or institutional differences in specific countries.

[0133] As previously exemplified, if a user utters "I lent money but they still haven't paid it back," and the user and the counterparty are identified as creditor and debtor roles respectively in the entity and role extraction block (232), and a money transfer and default event are extracted in the relationship and event extraction block (233), the general feature schema mapping block (234) can normalize this into a general feature value structure such as [Table 3].

[0134] -Party_Role_1 = Creditor-Party_Role_2 = Debtor-Event_1 = Money_Transfer-Event_2 = Obligation_Default-Object_Type = Monetary Asset

[0135] The feature values ​​generated in this manner are not directly classified into legal terms such as "claim for return of loan" or "default" of a specific country, but instead can be expressed as a general event and role structure. According to one embodiment of the present invention, a metadata completeness evaluation block (235) may be configured to receive structured general legal issue metadata generated from a general feature schema mapping block (234) as input and evaluate whether the required items are satisfied by comparing it with a predefined general event type-specific required metadata template. Specifically, the metadata completeness evaluation block (235) may refer to a predefined set of required metadata items corresponding to each general event type. For example, for money transfer and default event types, the amount of money, date and time of transaction, repayment deadline, roles of parties, and existence of evidence may be defined as required items. The metadata completeness evaluation block (235) may determine whether the general legal issue metadata extracted to date includes all of the required items.

[0136] In one embodiment, when a user utters, "I lent money but they still haven't paid it back," even though the money transfer event and the default event have been structured through the universal feature schema mapping block (234), the amount of money or the repayment deadline information may not yet be specified. In this case, the metadata completeness evaluation block (235) recognizes that the items for the amount of money and the repayment deadline are required items in the essential metadata template and can identify in real time that the items are missing. Additionally, the metadata completeness evaluation block (235) may be configured to identify not only the missing essential items but also items that are uncertain or have low reliability. For example, if the user states that they made a promise only verbally, a low reliability flag may be set for the item regarding the existence of evidence. That is, missing information or item information requiring supplementation identified by the metadata completeness evaluation block (235) can be transmitted to an adaptive conversational interface unit (220), and the adaptive conversational interface unit (220) can dynamically reconfigure the path of the conversation to generate a question to collect the missing information first.

[0137] According to one embodiment of the present invention, the feature value packaging and output block (236) may be configured to integrate structured universal legal issue metadata and missing information identification results transmitted from the universal feature schema mapping block (234) and the metadata completeness evaluation block (235) to generate a standardized feature packet that can be utilized for subsequent analysis. Specifically, the feature value packaging and output block (236) may combine each entity information, event type, time information, existence of evidence, fulfillment of required items, missing item flags, and reliability information into a single structured data set. The structured data set may be arranged according to a predefined data structure so that it can be directly utilized by the predictive analysis unit (240).

[0138] In one embodiment, when a user utters "I lent money but they still haven't paid it back," the feature value packaging and output block (236) can generate a packet containing general-purpose feature value information such as [Table 4].

[0139] - Party role information (creditor / debtor) - Information on monetary transfer events - Information on default events - Existence of monetary amount or repayment deadline - Existence of evidence and confidence level - Flag indicating whether required metadata is met

[0140] According to one embodiment of the present invention, a feature value packet may be transmitted to a predictive analysis unit (240), and the predictive analysis unit (240) may perform logical consistency evaluation, evidence reliability scoring, quantification of the likelihood of winning, and analysis of the likelihood of actual debt recovery based on general legal issue metadata included in the packet. Additionally, the feature value packaging and output block (236) may be configured to update and re-output the existing feature value packet when new information is collected during the conversation.

[0141] FIG. 5 is a block diagram showing the detailed configuration of a prediction analysis unit (240) according to one embodiment of the present invention.

[0142] According to one embodiment of the present invention, the input feature receiving and normalization block (241) may be configured to receive feature value packets transmitted from the universal legal issue extraction unit (230) and to align and normalize them into a form suitable for predictive analysis. Specifically, the input feature receiving and normalization block (241) may receive as input a structured set of feature values ​​consisting of universal case types, party role information, order of occurrence of events, existence of evidence, satisfaction of essential metadata, confidence flags, etc.

[0143] According to one embodiment of the present invention, the input feature receiving and normalization block (241) may perform preprocessing such as sorting each feature value according to a predefined data structure, converting it into a numeric value, or correcting missing items before being input into a prediction model. For example, in the embodiment described above, when a user utters "I lent money but they still haven't paid it back," the input feature receiving and normalization block (241) may receive feature information such as creditor and debtor role information, money transfer events, default events, whether a monetary amount exists, repayment deadline information, whether evidence exists, and confidence level.

[0144] According to one embodiment of the present invention, the input feature receiving and normalization block (241) may quantify, encode categorical values, or reconfigure to include default values ​​or uncertainty indications for missing items in order to utilize this information for predictive analysis. Additionally, the input feature receiving and normalization block (241) may be configured to update and reorder the existing set of feature values ​​when additional information is collected during the conversation process. Accordingly, the predictive analysis may have a dynamic analysis structure that is not a fixed evaluation at a single point in time, but is progressively supplemented as the conversation progresses.

[0145] Universal legal issue metadata aligned through the input feature reception and normalization block (241) can be transmitted in parallel to the logical consistency evaluation block (242) and the evidence reliability scoring block (243), and the two blocks can be configured to operate independently of each other.

[0146] First, the logical consistency evaluation block (242) may be configured to determine whether there is a contradiction between response data collected from the user. Specifically, the logical consistency evaluation block (242) can determine whether a logical conflict exists by comparing multiple feature values ​​such as the time of the event, the amount of money, the roles of the parties, and the repayment deadline. For example, if a user stated at the beginning of the conversation that they "lent it a month ago," but later stated that they "lent it six months ago," a discrepancy between the time information can be detected.

[0147] Additionally, the logical consistency evaluation block (242) can evaluate whether the causal relationship or temporal order of events is natural. For example, if the structure is such that the non-performance of obligations occurred before the event of the money transfer occurred, it can be judged as a logical error. The block can calculate a logical consistency score based on the results of such contradiction detection, and the score can be reflected in the subsequent calculation of the probability of winning.

[0148] Meanwhile, the evidence reliability scoring block (243) may be configured to score the reliability of the evidence based on the form and characteristics of the collected evidence. For example, a notarized promissory note, a financial institution's account transfer record, and an electronically signed contract document may be assigned a relatively high reliability score, while a verbal promise or a simple messenger conversation record may be assigned a relatively low reliability score.

[0149] Additionally, the evidence reliability scoring block (243) can calculate a comprehensive reliability score by considering whether multiple pieces of evidence are mutually complementary, whether different materials consistently prove the same fact, etc. For example, if account transfer records and message conversation content indicate the same amount and date, the reliability score can be calculated by weighting.

[0150] In the preceding embodiment, if the user states, "I lent money but they still haven't paid it back," and in an additional conversation answers, "There is no promissory note, only a message record," the logical consistency evaluation block (242) can calculate a consistency score by checking for any inconsistencies between time and amount information, and the evidence reliability scoring block (243) can calculate a reliability score of a certain level based on the existence of the message record.

[0151] The past case matching block (244) may be configured to match with a pre-trained past case database by combining the universal legal issue metadata aligned in the input feature reception and normalization block (241), the logical consistency score calculated in the logical consistency evaluation block (242), and the evidence reliability score calculated in the evidence reliability scoring block (243). Specifically, the past case matching block (244) may generate a feature vector including universal case types, party role structures, monetary amount ranges, evidence types, case progress, and consistency scores. The feature vector may be compared with similar case data accumulated in the past to produce a similarity score or pattern matching result.

[0152] The past case database can be constructed based on universal case types and structured feature values ​​without directly relying on the legal provisions of a specific country. Accordingly, the past case matching block (244) can compare cases having the same universal structure regardless of the legal system of each country.

[0153] In one embodiment, when a user states that "I lent money but they still haven't paid it back" and responds that a message record exists as evidence, the past case matching block (244) can calculate the similarity with past similar case cases based on the structure of the money transfer and non-performance of obligation case, the type of evidence based on the message record, the consistency score level, etc. For example, a set of similar cases can be derived by referring to the winning tendency in cases where only a message record exists or the pattern of judgment results according to the range of the amount of money.

[0154] According to one embodiment of the present invention, the past case matching block (244) is not limited to case law data of a specific country and may be configured to refer to an integrated case database in which case data collected from multiple countries is normalized according to a common universal feature schema. Accordingly, cases having the same structural case characteristics may be classified as similar cases regardless of the country, and the past case matching block (244) may be configured to perform a comparison based on structural similarity. Such a structure enables universal predictive analysis that does not rely on terminological or procedural differences between legal systems.

[0155] According to one embodiment of the present invention, the probability of winning block (245) may be configured to quantify and calculate the probability of winning in terms of legal aspects of the case based on similar case matching results transmitted from the past case matching block (244). Specifically, the probability of winning block (245) may calculate a predicted score for the probability of winning by synthesizing general legal issue metadata, logical consistency score, evidence reliability score, and similar case similarity information. The predicted score may be expressed as a probability value and, in one embodiment, may be calculated in the form of a percentage (%). In addition, the probability of winning block (245) may be configured to calculate a comprehensive score by comprehensively considering the case structure, quality of evidence, consistency of statements, range of monetary amounts, and past judgment patterns, rather than relying on a simple case type classification. For example, if a case involving the transfer of money and non-performance of obligations is clearly structured, has a high evidence reliability score, and has a logical consistency score above a certain standard, a relatively high probability of winning score may be calculated by reflecting the judgment trends of past similar cases.

[0156] In the preceding embodiment, if the user states, "I lent money but they still haven't paid it back," and a message record exists and no contradictions are found between the statements, the winning probability calculation block (245) can calculate a winning probability score in the form of a percentage, such as 72%, by reflecting the matching result with a similar case database. Conversely, if there is a lack of evidence or inconsistencies between the statements, a low probability score may be calculated. Additionally, the winning probability calculation block (245) can generate a confidence level or auxiliary indicator of the prediction result, and the prediction score can be managed separately from the debt recovery probability calculation block (247) to provide a distinction between legal winning probability and actual recovery probability.

[0157] The executableability element extraction block (246) may be configured to identify elements related to actual executableability after winning, based on general legal issue metadata and additional information collected during the conversation process. Specifically, the executableability element extraction block (246) may independently extract or evaluate items such as whether the user has an executable title, items of executable property of the counterparty (debtor), and items of the counterparty's willingness to repay analyzed in the conversation. For example, if the user states that they already possess a final judgment or a notarial deed, the item regarding the existence of an executable title may be evaluated positively. Additionally, if the counterparty states that they have a certain income or possess real estate or financial assets, items regarding executable property may be identified as existing. Conversely, if the counterparty states that they have no property or are out of contact, the executableability element may be evaluated restrictively.

[0158] Additionally, the executable possibility element extraction block (246) can extract the other party's willingness to repay as a quantitative or categorical value by analyzing expressions such as "I will repay soon" or "I will repay at least a portion first" in the conversation. This willingness to repay item can be used as an evaluation element for actual recovery possibility, distinct from legal probability of winning.

[0159] The debt recovery possibility calculation block (247) may be configured to independently evaluate each item derived from the execution possibility factor extraction block (246) and to synthesize them to calculate a quantitative score of the possibility of actually recovering the debt after winning the case.

[0160] In one embodiment, the debt recovery probability calculation block (247) can calculate a recovery probability score by summing the existence of an execution title, the number and type of executable property items, and the level of willingness to repay based on weights. For example, even if the probability of winning is high, if there are no executable properties at all, the recovery probability score may be calculated as low.

[0161] In the preceding embodiment, if the user possesses a message record and the probability of winning is calculated to be relatively high, but the counterparty currently has no income and their assets are not confirmed, the debt recovery probability calculation block (247) may calculate a low recovery probability score separately from the probability of winning. On the other hand, if the counterparty has a stable income and has expressed an intention to make partial repayment, a relatively high recovery probability score may be calculated.

[0162] According to one embodiment of the present invention, the win probability calculation block (245) and the debt recovery probability calculation block (247) may be configured to recalculate existing prediction results when new feature values ​​are added or modified during the conversation process. For example, if a user inputs the existence of additional evidence or supplements information on the monetary amount or repayment deadline, the feature values ​​are updated through the input feature reception and normalization block (241), and accordingly, the logical consistency score or the evidence reliability score may be changed. These changes may be reflected in the win probability and debt recovery probability scores in real-time or near-real-time. Accordingly, the prediction results may be provided as dynamic analysis results that are progressively refined according to the conversation flow, rather than as a single point-in-time evaluation.

[0163] According to one embodiment of the present invention, the probability of winning block (245) is configured to perform analysis by focusing on the legal judgment possibility of the case, and the possibility of recovering the debt block (247) is configured to independently evaluate factors affecting the actual possibility of execution after winning. Accordingly, even for the same case, if there is a lack of executable assets even if the probability of winning is high legally, the recovery possibility score may be calculated low; conversely, if the probability of winning is at an intermediate level but the possibility of execution is high, the recovery possibility score may be calculated relatively high. By separating and quantifying the legal judgment and the possibility of actual resolution in this way, the present invention can provide a multidimensional decision support structure that is distinguished from a simple winning prediction system.

[0164] According to one embodiment of the present invention, the result packaging and output block (248) may be configured to receive the legal probability of winning calculated in the probability of winning calculation block (245) and the actual debt recovery probability score calculated in the debt recovery probability calculation block (247), and to package them in a form to be provided to a user. Specifically, the result packaging and output block (248) may be configured to output the probability of winning and the debt recovery probability score by aligning them as separate indicators. The probability of winning may be provided as a predicted score in the form of a percentage regarding the legal judgment possibility of the case, and the debt recovery probability score may be provided as a separate quantitative indicator reflecting items such as the execution title, executable property, and willingness to repay.

[0165] In one embodiment, the result packaging and output block (248) may provide results to the user terminal device (100) in a format such as “Legal probability of winning: 72%”, “Substantial probability of recovering debt: Low (35%)”. In this way, by presenting the legal judgment and the substantial probability of recovery separately, the user can comprehensively judge not only the simple probability of winning but also the possibility of solving the actual problem after winning. Additionally, the result packaging and output block (248) may be configured to provide summary information on the main basis for judgment or elements requiring supplementation along with the predicted results. For example, it may include guidance messages such as “Probability of winning may increase with reinforcement of evidence” or “Need to verify enforceable assets”. The guidance information may be linked with the adaptive conversational interface unit (220) and utilized to induce follow-up questions or the collection of additional information.

[0166] According to one embodiment of the present invention, the result packaging and output block (248) may be configured to provide a summary of key influencing factors along with the probability of winning and the debt recovery probability score. For example, if a specific type of evidence contributed to an increase in the probability of winning, or if a lack of information on enforceable assets affected a decrease in recovery probability, the corresponding factors may be output together as explanatory information. Accordingly, the user can understand the basis of the prediction results and determine the direction for supplementing additional information.

[0167] According to one embodiment of the present invention, the result packaging and output block (248) may be configured to calculate information on the influence or contribution of key feature values ​​that contributed to the calculation of the scores, along with the win probability score and the bond recovery probability score. For example, the influence of the presence of a specific type of evidence on the increase in the win probability, or the extent to which a lack of information on enforceable assets contributed to the decrease in recovery probability, may be provided in a quantitative or graded form. Accordingly, the user can understand the basis of the prediction results and determine the direction for supplementing additional information.

[0168] According to one embodiment of the present invention, when a new feature value is input or an existing feature value is modified during the conversation process, the prediction analysis unit (240) may be configured to recalculate the winning probability score and the bond recovery probability score. Accordingly, the prediction result may be provided not as a fixed analysis result at a single point in time, but as a dynamic prediction result that is gradually refined during the interaction process with the user.

[0169] The scope of the present invention is not limited to the embodiments described above but may be implemented in various forms of embodiments within the scope of the appended claims. It is deemed that the scope of the claims of the present invention includes various modifications that are possible by anyone with ordinary knowledge in the technical field to which the invention pertains, without departing from the essence of the invention claimed in the claims.

Claims

1. A user state-adaptive AI legal consultation and prediction system that interacts based on user voice and text data, implemented by a computing device comprising one or more processors and memory storing program code executed by said one or more processors, A user state analysis unit that generates a user state vector by analyzing the user's emotional state and cognitive level in real time from voice information and text information collected from the user's utterance; An adaptive conversational interface unit configured to dynamically adjust the tone of the response and the difficulty of the question based on the above user state vector; A universal legal feature extraction unit configured to extract universal legal feature metadata not subject to the legal system of a specific country from the collected voice and text information above, and to determine whether essential metadata for each case type is satisfied; and It includes a predictive analysis unit configured to calculate the legal probability of winning the case and the actual probability of recovering the debt as independent quantitative indicators based on the extracted metadata above; The above adaptive conversational interface unit is, Configured to generate additional questions based on whether the above-mentioned universal legal feature metadata is missing, User state-adaptive AI legal consultation and prediction system.

2. In Paragraph 1, The above adaptive conversational interface unit is, When the emotion dimension value included in the user state vector is greater than or equal to a preset first threshold, the tone, speech rate, or voice intonation of the response is configured to be switched to a first control mode, and when the emotion dimension value is less than the first threshold, the response is configured to be provided in a standardized second control mode. User state-adaptive AI legal consultation and prediction system.

3. In Paragraph 1, The above adaptive conversational interface unit is, When the cognitive level dimension value included in the above user state vector is below a preset second threshold, a multi-stage question mode is configured to sequentially present multiple legal judgment elements by breaking them down step by step, and when the above cognitive level dimension value exceeds the second threshold, an integrated question mode including multiple judgment elements is configured to be activated. User state-adaptive AI legal consultation and prediction system.

4. In Paragraph 1, The above-mentioned universal legal feature extraction unit is, A system configured to have a set of essential metadata items for each predefined event type, and to extract relationships between entities and event structures from the collected voice information or text information and map them to the essential metadata items. User state-adaptive AI legal consultation and prediction system.

5. In Paragraph 4, The above-mentioned universal legal feature extraction unit is, It is configured to identify unmet items among the above-mentioned essential metadata items and to generate additional information request signals corresponding to the unmet items, and the adaptive conversational interface unit is configured to dynamically reconfigure the question generation priority based on the additional information request signals. User state-adaptive AI legal consultation and prediction system.

6. In Paragraph 1, The above prediction analysis unit is, A logical consistency evaluation block that determines whether there is a discrepancy in time information or event information between the collected response data and calculates a logical consistency score; and Including an evidence reliability scoring block that calculates an evidence reliability score based on the type and number of the collected evidence materials. User state-adaptive AI legal consultation and prediction system.

7. In Paragraph 6, The above prediction analysis unit is, The above-mentioned general-purpose legal feature metadata, the above-mentioned logical consistency score, and the above-mentioned evidence reliability score are configured as feature vectors, and the above-mentioned feature vectors are matched with a pre-trained historical case database to calculate the legal probability of winning a case as a first quantitative indicator in the form of a percentage, User state-adaptive AI legal consultation and prediction system.

8. In Paragraph 1, The above prediction analysis unit is, A debt recovery probability calculation block configured to calculate the actual debt recovery probability after winning as a second quantitative indicator by applying weights to each item and summing them up for multiple items representing executableability factors; The above first quantitative indicator and the above second quantitative indicator are calculated independently of each other. User state-adaptive AI legal consultation and prediction system.

9. In Paragraph 1, The above user state analysis unit is, It is configured to repeatedly update the user state vector based on additional voice information or text information collected during the conversation, and The above adaptive conversational interface unit is, Configured to re-select the interface control mode based on the updated user state vector, User state-adaptive AI legal consultation and prediction system.

10. A user state-adaptive AI legal consultation and prediction method that interacts based on user voice and text data, performed by a computer device, A vector generation step of generating a user state vector including a plurality of emotion dimension values ​​and cognitive level dimension values ​​by analyzing the user's emotional state and cognitive level in real time from voice information and text information collected from the user; A step of dynamically adjusting the tone of the response and the difficulty of the question by comparing the emotion dimension value and the cognitive level dimension value included in the above user state vector with a preset threshold; A step of extracting universal legal characteristic metadata that is not subject to the legal system of a specific country from the above voice information and text information, and determining whether essential metadata items for each case type are satisfied to identify missing items; A step of generating additional questions based on the above missing items and reconstructing the conversation path; A step of calculating a logical consistency score and an evidence reliability score based on the above-mentioned general-purpose legal feature metadata; A step of configuring the above-mentioned general-purpose legal feature metadata, the above-mentioned logical consistency score, and the above-mentioned evidence reliability score into a feature vector, and calculating the legal probability of winning the case as a first quantitative indicator by matching it with a pre-trained historical case database; and It includes a step of evaluating enforceability factors to calculate the actual possibility of debt recovery after winning as a second quantitative indicator; and From the step of generating the user state vector to the step of calculating the first quantitative indicator and the second quantitative indicator, the steps are repeatedly performed according to input information additionally collected during the conversation and are updated in real-time or near-real-time, User State Adaptive AI Legal Consultation and Prediction Method 11. In Paragraph 10, The above vector generation step is, A preprocessing step further comprising, prior to performing natural language processing or sentiment analysis, distinguishing between fixed noise and non-fixed noise included in the voice information and removing background noise by performing adaptive filtering by applying different filtering parameters according to the type of noise; User State Adaptive AI Legal Consultation and Prediction Method

Images