Reverse guiding type intelligent interaction method and system for realizing dynamic information evaluation

Abstract

The invention discloses a reverse guidance type intelligent interaction method and system for realizing dynamic information assessment. The system comprises an information sufficiency assessor, an information sufficiency assessment device and a dynamic information assessment device, wherein the information sufficiency assessor is used for dynamically and quantitatively assessing and identifying key information gaps on the basis of an information entropy theory; the reverse questioning strategy generator is used for generating an optimal reverse guidance questioning sequence for supplementing a key information gap based on a Markov decision process model; the hierarchical problem decomposition engine is used for recursively decomposing a complex problem into a plurality of sub-problems by adopting a semantic tree decomposition technology, and fusing solution results of the sub-problems by utilizing an evidence theory fusion mechanism to form a problem solution; the personalized solution synthesizer is used for analyzing the cognitive level, learning style and attention characteristics of the user portrait, performing matching optimization on a problem solution and the user portrait, and calculating and measuring the semantic matching degree of the problem solution and the user portrait; and applying reinforcement learning to continuously optimize an output strategy according to user feedback to generate a personalized solution.

Description

Technical Field

[0001] The present invention belongs to the technical field of human-computer interaction, and in particular relates to a reverse-guided intelligent interaction method and system for realizing dynamic information evaluation. Background Art

[0002] With the continuous advancement of technology, large language models have exposed numerous limitations. Regarding information completeness assessment, they lack quantitative evaluation mechanisms, the ability to identify key information gaps, and quantitative analysis methods for information uncertainty. Their questioning strategies are not intelligent enough, relying on fixed templates that fail to consider maximizing information value and dynamically adjusting to user characteristics. Their ability to handle complex questions is insufficient, lacking a systematic problem decomposition mechanism, limiting information sharing among sub-questions, and lacking theoretical support for answer integration. Their level of personalization is insufficient, with standardized output solutions failing to fully consider user characteristics and dynamic optimization feedback. A core technical contradiction lies in the mismatch between the open generation capabilities of large language models and the limited information collection mechanisms of traditional dialogue systems. Traditional dialogue systems rely on manually pre-set rules to ensure information completeness, at the expense of interactive flexibility. While modern large language models allow for free interaction, they lack a systematic mechanism for acquiring necessary information, leading to questionable answer reliability. Furthermore, no intermediate technical solution exists to balance these two aspects, making it impossible to achieve quantitative assessment of information sufficiency, optimal questioning strategies guided by information value, and dynamic decision-making based on information completeness.

[0003] Furthermore, existing systems lack a theoretical framework that organically integrates information theory, decision theory, and the capabilities of large language models. For example, they lack uncertainty quantification models based on information entropy, optimization methods that model questioning strategies as Markov decision processes, and a unified representation system for the user-information-question relationship. Consequently, they are unable to build a complete closed-loop "evaluation-guidance-verification" mechanism, severely restricting the application of AI systems in demanding scenarios. Summary of the Invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a reverse-guided intelligent interaction method and system for dynamic information evaluation. The technical problem to be solved by the present invention is achieved through the following technical solutions: In a first aspect, an embodiment of the present invention provides a reverse-guided intelligent interactive system for implementing dynamic information evaluation, the system comprising an information sufficiency evaluator, a reverse questioning strategy generator, a hierarchical problem decomposition engine, and a personalized solution synthesizer; wherein, The information sufficiency evaluator is used to dynamically and quantitatively evaluate user input information based on information entropy theory and in combination with task context information to identify key information gaps; The reverse questioning strategy generator is configured to calculate the expected information gain and user acceptance of each candidate question based on a Markov decision process model if a key information gap is identified, and generate an optimal reverse guidance question sequence based on the expected information gain and user acceptance to guide the user to provide supplementary information to fill the key information gap; The hierarchical problem decomposition engine is used to judge the user-input question supplemented by the reverse questioning strategy generator through a dynamic problem complexity judgment mechanism. When the user-input question is judged to be complex, the semantic tree decomposition technology is used to recursively decompose the complex problem into multiple independently processable sub-problems. The solution results of each sub-problem are integrated using the evidence theory fusion mechanism to form a problem solution. The sub-problems share information among themselves through a shared memory pool framework. The personalized solution synthesizer is used to analyze the cognitive level, learning style, and attention characteristics of the user profile, match and optimize the problem solution with the user profile, calculate and measure the semantic fit between the problem solution and the user profile, and introduce the semantic fit as a reward function into the reinforcement learning framework, so as to use the reinforcement learning framework to continuously optimize the output strategy based on user feedback to generate personalized solutions that are highly matched with the user profile in terms of content depth, formal style, and presentation method.

[0005] In a second aspect, an embodiment of the present invention provides a reverse-guided intelligent interaction method for implementing dynamic information evaluation, the method comprising: Based on information entropy theory and combined with task context information, user input information is dynamically and quantitatively evaluated to identify key information gaps; If a critical information gap is identified, the expected information gain and user acceptance of each candidate question are calculated based on the Markov decision process model, and an optimal reverse guidance question sequence is generated based on the expected information gain and the user acceptance to guide the user to provide supplementary information to fill the critical information gap; The user-input question, supplemented by the reverse questioning strategy generator, is judged through a dynamic problem complexity judgment mechanism. When the user-input question is judged to be complex, the semantic tree decomposition technology is used to recursively decompose the complex question into multiple independently processable sub-problems. The evidence theory fusion mechanism is used to fuse the answers to the sub-problems to form a problem solution. The sub-problems share information among themselves through a shared memory pool framework. Analyze the cognitive level, learning style, and attention characteristics of the user portrait, optimize the matching of the problem solution with the user portrait, calculate and measure the semantic fit between the problem solution and the user portrait, and introduce the semantic fit as a reward function into the reinforcement learning framework. Use the reinforcement learning framework to continuously optimize the output strategy based on user feedback to generate personalized solutions that are highly matched with the user portrait in terms of content depth, formal style, and presentation method.

[0006] Beneficial effects of the present invention: The reverse-guided intelligent interactive system proposed in this invention, which implements dynamic information evaluation, takes the information sufficiency evaluator as the starting point and relies on the reverse questioning strategy generator, the hierarchical problem decomposition engine, and the personalized solution synthesizer to build a dynamic, adaptive, and highly robust intelligent question-answering system cognitive architecture. This architecture integrates information evaluation, strategy generation, problem solving, and personalized output functions to achieve full-link closed-loop control from user input evaluation to high-matching output. This solution not only achieves the leap from passive response to active guidance in intelligent question-answering, but also constructs a cognitive-driven information recognition, problem reconstruction, and content synthesis closed-loop system in complex scenarios, providing a practical and high-performance solution path for interactive artificial intelligence applications that meet professional, high-complexity, and personalized needs.

[0007] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 1 is a schematic structural diagram of a reverse-guidance intelligent interactive system for implementing dynamic information evaluation provided by an embodiment of the present invention; Figure 2 1 is a schematic diagram of an implementation flow of an information sufficiency evaluator provided by an embodiment of the present invention; Figure 3 Schematic diagram of a Markov decision process model in a reverse questioning strategy generator provided by an embodiment of the present invention; Figure 4 This is a schematic diagram of a semantic tree decomposition process using semantic tree decomposition technology in a hierarchical question decomposition engine provided by an embodiment of the present invention; Figure 5 Schematic diagram of a personalized solution generation process in a personalized solution synthesizer provided by an embodiment of the present invention; Figure 6 A schematic diagram of a hierarchical intelligent interaction architecture including the reverse-guided intelligent interaction system for implementing dynamic information evaluation proposed by the present invention; Figure 7 1 is a schematic diagram of a multimodal information processing flow of an input layer in a hierarchical intelligent interaction framework provided by an embodiment of the present invention; Figure 8This is a flow chart of a reverse-guided intelligent interaction method for implementing dynamic information evaluation provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0009] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0010] First, see Figure 1 The embodiment of the present invention provides a reverse-guided intelligent interactive system for implementing dynamic information evaluation, the system comprising an information sufficiency evaluator, a reverse questioning strategy generator, a hierarchical problem decomposition engine, and a personalized solution synthesizer; wherein, An information sufficiency evaluator, which dynamically and quantitatively evaluates user input information based on information entropy theory and combined with task context information to identify key information gaps; A reverse questioning strategy generator is used to calculate the expected information gain and user acceptance of each candidate question based on the Markov decision process model if a key information gap is identified, and to generate an optimal reverse guidance question sequence based on the expected information gain and user acceptance to guide the user to provide supplementary information to fill the key information gap; The hierarchical problem decomposition engine uses a dynamic problem complexity judgment mechanism to judge user-input questions supplemented by the reverse questioning strategy generator. When the user-input question is judged to be complex, semantic tree decomposition technology is used to recursively decompose the complex problem into multiple independently processable sub-problems. The solution results of each sub-problem are integrated using the evidence theory fusion mechanism to form a problem solution. The sub-problems share information among themselves through a shared memory pool framework. The personalized solution synthesizer is used to analyze the cognitive level, learning style, and attention characteristics of the user profile, optimize the matching of problem solutions with the user profile, calculate and measure the degree of semantic fit between the problem solution and the user profile, and introduce the semantic fit as a reward function into the reinforcement learning framework. The reinforcement learning framework is used to continuously optimize the output strategy based on user feedback to generate personalized solutions that are highly matched with the user profile in terms of content depth, formal style, and presentation method.

[0011] Next, each part of the reverse guidance type intelligent interactive system for realizing dynamic information evaluation proposed by the present invention is introduced in detail.

[0012] Information Sufficiency Evaluator: The information sufficiency evaluator in the embodiment of the present invention is based on information entropy theory and combined with task context information to dynamically quantify and evaluate user input information to identify key information gaps, including: based on multi-dimensional information vector representation technology, converting user input information into a high-dimensional semantic vector; calculating the probability of each information point being covered in the high-dimensional semantic vector, and calculating the information entropy of the high-dimensional semantic vector based on all probabilities; introducing the task criticality coefficient, the problem domain complexity factor and the user's professional ability weight in combination with the task context information; calculating the information completeness score based on the domain knowledge graph and information entropy, as well as the task criticality coefficient, the problem domain complexity factor and the user's professional ability weight; based on the task criticality coefficient, the problem domain complexity factor and the user's professional ability weight, dynamically adjusting the evaluation threshold to determine whether the information completeness score is less than the dynamically adjusted evaluation threshold. If it is less than, there is a key information gap in the user input information. More specifically: As the first step of this system, the information sufficiency evaluator is responsible for the dynamic evaluation of the quality of user input information. Its core goal is to accurately identify the key content gaps in user input through quantitative means and provide a basic basis for the formulation of reverse guidance strategies. Figure 2 The following is the workflow diagram of the information sufficiency evaluator: Based on the information entropy theory, an information uncertainty measurement model is constructed, and the uncertainty of semantic information is used as the key indicator for evaluating information completeness. First, the system performs multimodal fusion and semantic extraction on the user input information, and uses the multidimensional information vector representation technology to convert the user input information into a high-dimensional semantic vector. , and related concept vectors in the domain knowledge graph Matching is done through semantic similarity function The fit between the user input information and the domain knowledge graph is calculated, and the semantic similarity function serves as the basic input for information coverage. The multi-dimensional information vector representation technology adopts a hybrid representation method that combines subword encoding with context embedding. First, the user input information is segmented to generate a sequence of subword units. Multi-layer semantic features at the character level, word level, and sentence level are extracted through a pre-trained language model (such as BERT) to construct a high-dimensional semantic vector containing grammatical structure, semantic dependency, and context association. For multimodal input (such as images and speech), the visual features (such as image semantic features extracted by CNN), speech features (such as the acoustic vector of the Mel-spectrogram encoded by Transformer) and text semantic vectors are mapped to a unified semantic space through a cross-modal attention mechanism to form a multi-dimensional fusion vector containing text semantics, modal features, and scene information. This technology achieves accurate semantic modeling of user input information through multi-dimensional feature fusion, which is different from the traditional single-modal vector representation method.

[0013] At the same time, in order to further characterize the uncertainty of user input information, the information entropy model is constructed as follows: ; in, Indicates that the user input information corresponds to the first The probability of an information point being covered is obtained by using a neural network model to discriminate the semantic distribution of user input. Indicates the number of information points in a high-dimensional semantic vector. This entropy value is used to measure the system's degree of certainty about the user's input information. The higher the entropy, the more key content the system has not yet grasped in the user's input information, and the lower the completeness. In order to further improve the dynamic adaptability of the evaluation, the system introduces the task criticality coefficient , Problem Domain Complexity Factor and user professional ability weight , dynamically adjust the evaluation threshold, and finally form the information completeness score function: ; in, represents the information completeness score, represents information entropy, It is the theoretical maximum information entropy calculated based on the domain knowledge graph and used for normalization. When the information completeness score C is lower than the dynamically adjusted evaluation threshold τ, the system determines that there is a significant gap in the current information and initiates a supplementary request to the reverse question strategy generator. In actual deployment, the system uses historical interaction data to dynamically update 、 、 , adapting to different user groups and task types, ensuring that evaluation results are both universal and considerate of individual differences. This information quantification evaluation mechanism opens up the logical chain between information input and strategic decision-making, enabling the system to respond efficiently while sensing insufficient input, laying a precise data foundation for subsequent reverse guidance, complex problem decomposition, and personalized output.

[0014] Reverse Question Strategy Generator: In the embodiment of the present invention, the expected information gain and user acceptance of each candidate question are calculated based on the Markov decision process model, and the optimal reverse guidance question sequence is generated according to the expected information gain and user acceptance, which is intended to guide the user to provide supplementary information to fill the key information gap, including: modeling the MDP state, the MDP state including the state space, the action space, the state transition probability, the reward function and the discount factor; wherein the state space is the splicing result of the user input information and the user portrait, the action space is the set of all candidate questions, the state transition probability is used to measure the response probability of the user feedback after the question is asked, the reward function is used to measure the expected information gain brought by the question, the discount factor is used to measure the expected information gain brought by the question, and the reward function is used to measure the expected information gain brought by the question. Factors are used to control the impact of long-term returns; based on the MDP state, an optimal questioning strategy function is constructed with the goal of maximizing the long-term cumulative information value; the first information entropy in the current state space is calculated, and the second information entropy in the expected state space after the question is calculated based on the optimal questioning strategy function, and the information entropy decrease value is calculated based on the first information entropy and the second information entropy; a comprehensive score function for each action in the action space in the current state space is constructed based on the information entropy decrease value and the user acceptance weight; the optimal value of the comprehensive score function is solved through an iterative algorithm, and combined with a personalized question template library, an optimal reverse guidance question sequence is generated to guide users to fill key information gaps. More specifically: After the information sufficiency evaluator identifies the key information gaps, the reverse questioning strategy generator constructs a Markov decision process model to globally plan the questioning behavior, with the goal of minimizing the system uncertainty caused by the missing information. Figure 3 The following is a Markov decision process diagram of the reverse question strategy generator: The MDP state of this model is a five-tuple . Where: state space S The semantic state of the splicing result of user input information and user portrait; action space A is a set of all candidate questions, which is a set of 3 to 5 candidate questions generated for each key information gap based on the domain knowledge graph; state transition probability It is used to measure the response probability of user feedback after asking a question. It is estimated based on the user feedback model trained based on historical interaction data. The formula is: ,in, For the Transformer-based user response prediction model, Softmax seeks the probability distribution of each category in the multi-classification problem, converting the model output into an interpretable probability form. The input is the state and questioning actions The concatenated vector of The probability distribution of is optimized through historical interaction log training; the reward function It is used to measure the expected information gain brought by asking questions. The formula is expressed as: ,in, In state The information entropy under In state Execute the question action The expected information entropy after To perform the question action The user acceptance score is calculated by the cosine similarity between the interaction preference vector and the question style in the user portrait; The construction relies on the principle of maximizing information value; discount factor Used to control the impact of long-term returns. The system aims to maximize the long-term cumulative information value and construct the optimal questioning strategy. , so that the maximum cumulative reward that can be obtained by taking a certain action in any state, that is, the long-term expected return, is formally expressed as a value function: ; in, Indicates that the status Follow the questioning strategy The corresponding maximum cumulative reward, Expressing hope, Indicates that the status Take the following action to ask questions The corresponding expected information gain is, Indicates that the questioning strategy will be followed Status As the initial state.

[0015] Specifically calculate the information entropy reduction value brought by each candidate question relative to the current semantic state. The information entropy reduction value can be defined as: ; in, Indicates status Take the action of asking questions The corresponding information entropy is Expected state after asking the question Take the action of asking questions Corresponding information entropy. In order to integrate the user acceptance factor, the system expresses the comprehensive score of each action as: ; in, Indicates status Take the following action to ask questions The comprehensive score of , represents the state space, , represents the action space, Indicates status Take the following action to ask questions The information entropy decreases, It represents the user acceptance weight, which is dynamically adjusted based on the user portrait to reflect the user's tolerance for the current question style and complexity. Finally, the system solves the optimal strategy through policy iteration or value iteration algorithm. , and combined with a library of personalized question templates to construct natural language expressions, this enables guided questioning with semantic accuracy, style matching, and manageable burden. The question sequence is not only optimal in terms of information value, but also ensures comfortable and responsive interactions through user modeling. The system architecture effectively connects the information sufficiency evaluator and the reverse questioning strategy generator, enabling systematic planning of uncertainty-minimizing paths.

[0016] Hierarchical Problem Decomposition Engine: In this embodiment of the present invention, a user-entered question is first judged using a dynamic problem complexity determination mechanism. If the user-entered question is determined to be complex, semantic tree decomposition technology is used to recursively decompose the complex question into multiple independently processable sub-questions. A shared memory pool architecture is used to enable information sharing between sub-questions, improving processing efficiency. Using an evidence theory fusion mechanism, the answers to each sub-question are integrated to form a complete problem solution, ensuring the integrity of information and the rationality of the dependencies between sub-questions during the complex problem processing process. More specifically: Before entering the hierarchical question decomposition engine, the system first determines the complexity of the user's input question using the dual criteria of information completeness score C and pre-decomposition depth of the semantic tree. The user's input question can be pre-complemented with the optimal reverse-guided question sequence output by the reverse question strategy generator. If the user's input information completeness score C is lower than a preset threshold, and the initial depth of the semantic tree derived from syntactic analysis exceeds a preset level (e.g., three layers), or contains more than two logical connectives (e.g., "and," "or," and "if"), the question is considered complex and the hierarchical decomposition process is triggered. This automatically identifies the complexity of the question, ensuring that the engine is activated only when necessary, improving system efficiency.

[0017] As the central component connecting information evaluation and output generation, the hierarchical problem decomposition engine is oriented towards problems with complex structures and high semantic nesting. It uses a semantic tree recursive decomposition mechanism to transform the original problem into multiple sub-problem nodes with clear boundaries and logical relationships. Figure 4The figure shows a schematic diagram of semantic tree decomposition of a hierarchical problem decomposition engine: During the semantic tree construction process, the embodiment of the present invention uses semantic tree decomposition technology to recursively decompose complex problems into multiple sub-problems that can be processed independently, including: based on syntactic analysis and dependency extraction technology, constructing a root node to represent the overall problem semantics, leaf nodes to represent the smallest semantic unit, and each intermediate node to correspond to the semantic combination of subtasks, so as to recursively decompose the supplemented complex problem into multiple sub-problems that can be processed independently. In order to achieve efficient information sharing between sub-problems, the system introduces a shared memory pool architecture, which uniformly stores user input information, intermediate results of completed sub-problems, and key context information in a high-dimensional semantic cache unit, and realizes semantic reference and context completion across sub-tasks through a query mapping mechanism. Among them, The shared memory pool uses a key-value pair storage structure, where the key is the unique identifier of the semantic unit (such as the missing semantic unit ID or sub-question number), and the value is a triple containing the semantic vector, timestamp, and task relevance. ,in, is the semantic vector, is the timestamp, is the task relevance. Information sharing across sub-problems is achieved through the following mechanisms: Contextual retrieval: When a sub-problem requires additional information, the current semantic vector Calculate all entries in the memory pool Cosine similarity of , before extraction high similarity entries as candidate context vectors.

[0018] Dynamic completion: The candidate context vector is combined with the current semantic vector through the attention mechanism Fusion, the formula is: , .

[0019] Consistency maintenance: A gated recurrent unit (GRU) is introduced to update the memory pool entries. When a new sub-problem result is written, outdated information is filtered through the forget gate to ensure that the context is consistent with the current task semantics.

[0020] After the sub-problems are solved, the system needs to perform multi-source information fusion without destroying the original semantic structure. In the embodiment of the present invention, the evidence theory fusion mechanism is used to fuse the answer results of each sub-problem, including: constructing a trust distribution function; calculating the trust level of the answer results of any two sub-problems according to the trust distribution function, and then calculating the corresponding fused trust level based on the two trust levels obtained by calculation; judging whether the fused trust level is less than a preset threshold. If it is less than, triggering the reverse question strategy generator to regenerate the optimal reverse guidance question sequence. If it is greater than or equal to, outputting the fusion result.

[0021] Adopting a fusion model based on evidence theory, the answer to each sub-problem is set as a proposition. , and uses its uncertainty as the basis for conflict measurement to construct the trust allocation function , indicating the i The degree of trust in the proposition set by each sub-question. The final fusion result of the system is given by Dempster's synthesis rule: ; in, Indicates that the trust degree corresponding to sub-problem B is calculated using the trust distribution function. Indicates that the trust degree corresponding to the sub-problem C is calculated using the trust distribution function. Indicates that A is the intersection of subproblems B and C, Indicates that the trust level corresponding to the intersection A is calculated using the trust distribution function. represents the conflict coefficient, , Representing subproblems B There is no intersection with subproblem C. Among them, the conflict coefficient It characterizes the degree of consistency between the results of each sub-problem. The smaller the conflict, the stronger the synergy between the sub-solutions, and the higher the credibility of the fusion result. Through this fusion process, the system not only ensures the logical consistency of each sub-problem in the semantic dimension, but also improves the robustness of handling uncertain results, so that the final overall solution is optimal in terms of completeness, rationality and task adaptability. At the same time, the reverse questioning strategy generator forms a closed feedback loop with the previous information sufficiency evaluator and the reverse questioning strategy generator. When the confidence level of the sub-problem fusion result is lower than the system's preset threshold, it can reversely trigger the reverse questioning strategy generator to ask targeted supplementary questions, forming a stable and efficient path for solving complex problems.

[0022] Individual solution synthesizer: In an embodiment of the present invention, the cognitive level, learning style, and attention characteristics of the user portrait are analyzed, the problem solution is matched and optimized with the user portrait, the semantic fit between the problem solution and the user portrait is calculated and measured, and the output strategy is continuously optimized based on user feedback using a reinforcement learning framework to generate a personalized solution that is highly matched with the user portrait in terms of content depth, formal style, and presentation. This includes: modeling the user portrait as a multidimensional feature vector including cognitive level, learning style, and attention characteristics; mapping the problem solution to a semantic output space, and constructing a semantic fit function based on the multidimensional feature vector and the problem solution mapped to the semantic output space; introducing the semantic fit function as a reward function into the reinforcement learning framework, and iteratively optimizing the output strategy through a policy optimization gradient algorithm to generate a personalized solution that is highly matched with the user characteristics in terms of content depth, formal style, and presentation. More specifically: As the core module of the system output, the personalized solution synthesizer integrates the processing results of the previous components with the multi-dimensional user portrait on the basis of ensuring information completeness, and realizes the personalized adaptation of the output content in terms of cognitive depth, expression form and style tendency. The system models the user portrait as a multi-dimensional feature vector ,in Indicates the cognitive level, modeled based on the depth of questions and response speed in the user's historical interactions. Represent learning styles and model visual, structural, and textual presentations through preference identification. Represents attention features, extracted by combining the user's stay time during the interaction process and gaze tracking data. Cognitive level : Based on the quantification of question depth and response speed in user historical interactions, the formula is: ,in, D is the average level of professional concepts involved in historical questions (calculated by semantic tree depth, for example, basic concepts = 1 level, complex concepts = 2 levels, cross-domain concepts = 3 levels), is the average response time (seconds), and is the weight coefficient , the default is =0.6, =0.4. A larger value indicates a higher level of cognition.

[0023] Learning Style :Through preference identification, it is divided into three categories of labels: visualization (such as chart requirements), structured (such as point-by-point summary), and textual (such as long article analysis), corresponding to , For visual labels, For structured tags, A textual label.

[0024] Attention characteristics : Quantified based on the dwell time and gaze tracking data (such as the gaze point distribution collected by the eye tracker) during the interaction process, the formula is: in, For users in The dwell time of each information point (seconds), For the The importance weight of each information point (calculated by the concept association in the domain knowledge graph), Represents the number of information points in the high-dimensional semantic vector. The larger the value, the higher the concentration.

[0025] like Figure 5 The figure shows the fitness calculation flow chart of the personalized solution synthesizer. During the solution generation process, the system integrates the sub-problems corresponding to the solution results. Mapped to the semantic output space, Indicates the solution to the problem n The answer results are combined with the user portrait for matching optimization and the overall semantic fit function is calculated. , which measures the semantic fit between the solution and the user's features. The semantic fit function constructed in the embodiment of the present invention is expressed as follows: ; in, represents a multidimensional feature vector, Indicates a solution to the problem, Indicates the number of answer results in the problem solution, Indicates the i The weight of each answer in the overall problem solution is dynamically assigned based on the criticality of the sub-problem tasks. express Middle i The answer results, Represents a multidimensional feature vector In the i The user feature transformation value under the solution result, Indicates solution and The semantic similarity between them is achieved by calculating the cosine similarity in the vector space through the semantic encoder. Among them, the user feature transformation function used to calculate the user feature transformation value is expressed as follows: ; in, Cognitive level The semantic embedding vector of For learning styles The one-hot encoded vector of Attention feature The normalized vector of Represents a vector concatenation operation.

[0026] The system further introduces the semantic fit as a reward function into the reinforcement learning framework, and uses the policy gradient algorithm to output the strategy Perform iterative optimization with the goal of maximizing long-term expected user satisfaction: ; in, represents the long-term expected user satisfaction value, represents the optimization strategy parameters, Indicates the output strategy Induced sub-problem fusion results The mathematical expectation under the distribution. This process is driven by user feedback. The system continuously collects response accuracy, reading behavior and subjective evaluation during interaction, and optimizes the strategy parameters through back propagation. , making future output more aligned with user preferences and understanding paths, thereby achieving comprehensive personalized adaptation of content depth, presentation format, and presentation rhythm. This personalized solution synthesizer is not only the final implementation of information processing results, but also the driving engine for the self-evolution of the system's cognitive network. Together with the information sufficiency evaluator, the reverse questioning strategy generator, and the hierarchical problem decomposition engine, it forms a closed-loop feedback mechanism, continuously improving the system's response quality and user satisfaction in human-computer interaction.

[0027] From the above, it can be seen that when the system is running, the embodiment of the present invention first uses the information sufficiency evaluator to dynamically evaluate the user input information and identify key information gaps. If the information is insufficient, the reverse question strategy generator generates accurate reverse questions based on the principle of maximizing information value to guide the user to supplement the information. In complex problem situations, the processing efficiency is improved through the hierarchical problem decomposition engine. When the information completeness reaches the preset threshold, the personalized solution synthesizer generates a personalized solution that is highly matched with the user's cognitive style, professional level and usage scenario. This two-way interactive verification mode significantly improves the reliability of the intelligent question-answering system, optimizes the interaction efficiency, establishes a self-evolving cognitive network, and enables the system to have information completeness verification and question strategy optimization capabilities.

[0028] Furthermore, the reverse guidance intelligent interaction system for dynamic information evaluation proposed in the embodiment of the present invention can be used in a hierarchical intelligent interaction architecture to achieve seamless connection between information evaluation, reverse guidance and solution generation. For example, the hierarchical intelligent interaction framework is divided into the following Figure 6In the six-layer architecture shown, each layer works together through standardized interfaces to form a complete closed loop of information acquisition and verification. Specifically: The input layer is responsible for user interface interaction and multimodal information preprocessing, including receiving multimodal input (text, voice, images, etc.), standardizing input information, and maintaining user interaction status. Through efficient preprocessing mechanisms, it converts different forms of input information into a format that the system can uniformly process, laying the foundation for subsequent interaction processes.

[0029] The input layer serves as the entry point for information access and format conversion, and is responsible for uniformly encoding and structuring multimodal information, providing consistent support for information evaluation, strategy planning, and output generation. Figure 7 The following is a flowchart of multimodal information processing at the input layer. The system first uses the modality recognition module to classify the user input and then uses the input channel label to classify the user input. , assigned to the corresponding preprocessing path. In the text channel In the speech channel, the system uses subword encoding and semantic context enhancement mechanism to build semantic vectors; In the image channel, key speech fragments are extracted based on the Mel-spectrogram and attention mechanism; In the video channel, semantic feature maps are extracted through multi-scale convolutional neural networks, and the content labels and scene information are obtained by combining the target detection model. In the process, the video input is frame sampled and keyframes are identified, and the spatiotemporal joint features are extracted using a three-dimensional convolutional neural network or a spatiotemporal attention mechanism. At the same time, the audio track is separated, and the speech semantic features are extracted using a Mel-spectrogram and a Transformer encoder. Finally, the video features are fused with the speech and text modalities into a unified semantic space through a cross-modal attention mechanism. In order to achieve the fusion of different modalities in a unified vector space, the system introduces a cross-modal alignment network to map the features of each modality to a common semantic embedding space, and finally construct a multimodal fusion vector. : ; in, Indicates the The representation of each modality after feature extraction, For the The normalized encoding function of the modes, It is The fusion weight of each modality is dynamically allocated according to the modality confidence and context importance. Represents the number of modalities. This vector serves as the input to the subsequent information evaluation layer of the system, not only preserving the modality-specific characteristics but also achieving a unified expression at the semantic level. At the same time, the input layer is also responsible for maintaining the user interaction state, including input rhythm, modality switching behavior, and historical context continuity, to construct a dynamic interaction context tensor. , which contains the temporal structure and historical trajectory features of the current multimodal input, and provides user behavior trend support for the dialogue management layer. Continuously updated in the time dimension through a gating mechanism: ; in, express t The dynamic interaction context tensor at each moment, express t -1 moment dynamic interaction context tensor, is the weight matrix, is the bias term, and σ is the Sigmoid activation function. This update mechanism not only ensures the temporal continuity of the state vector but also strengthens the semantic parsing capability of the current input in the context of historical behavior. The high-quality unified vector representation and state tracking results generated by the input layer serve as the basic input for the information evaluation layer to calculate semantic completeness and identify key information gaps. It also provides the necessary behavioral context and multimodal fusion support for subsequent personalized guidance strategies, ensuring the accuracy and stability of the entire interactive system during the information perception phase.

[0030] The dialogue management layer controls the overall dialogue flow and state transitions based on a state machine, maintains the context of multiple dialogue rounds, and dynamically adjusts interaction strategies based on dialogue progress and user feedback. This ensures the coherence and fluency of dialogues, enabling the system to respond appropriately to different dialogue scenarios.

[0031] The dialogue management layer architecture is the core execution unit of the intelligent interactive closed loop. Its overall functional logic revolves around information understanding, strategy planning and dynamic response, ensuring that the system can maintain efficient information control and personalized interaction capabilities in an open environment. The dialogue management layer builds a multi-round dialogue process based on finite state automata. Each state node Represents the specific interaction scenario in the conversation, state transfer function Receive the user's question action As input and drives the evolution of the dialogue state. To enhance the coherence of the dialogue in a multi-round environment, the system maintains a context tracking vector , is dynamically updated through the context fusion function, where the fusion process adopts a gated recursive mechanism: ; in, expresst The context tracking vector at the moment, express t -1 time context tracking vector, 、 is the weight matrix, represents the bias term, Represents the Tanh activation function. This structure makes the current round input In history Based on this, the fusion update is completed to realize the dynamic capture of state information and semantic trajectory, thereby supporting personalized state transfer and response planning.

[0032] Information Assessment Layer: This layer employs an information sufficiency evaluator, which uses uncertainty quantification methods based on information entropy and multidimensional information vector representation and analysis techniques to assess information sufficiency and identify key information gaps. These gaps are prioritized, providing accurate information assessment results to the strategy generation layer, enabling the generation of targeted reverse guidance strategies.

[0033] In the information evaluation layer, the system measures the uncertainty of the semantic vector distribution of the current input based on entropy reasoning, and constructs the information distribution probability vector using the projection results of the multi-dimensional embedding vector in the semantic space of the knowledge graph. , and calculate the information entropy of the input, which measures the consistency between the input content and the expected semantic space. The larger the entropy value, the more dispersed the input and the more incomplete the structure. In the identification of key information gaps, the system weights and integrates the contribution and missing degree of each semantic unit to the overall task goal, constructs a criticality scoring matrix, and prioritizes the gaps to generate a gap vector for the strategy generation layer to call For more detailed implementation, please refer to the description of the information sufficiency evaluator part, which will not be repeated here.

[0034] Strategy Generation Layer: This layer uses a reverse questioning strategy generator, which generates the optimal reverse guidance strategy based on MDP question sequence planning and information value maximization algorithms. It generates personalized question templates based on user profiles, guiding users to supplement key information to maximize information acquisition efficiency.

[0035] The strategy generation layer plans the questioning behavior based on the Markov decision process. The current state Combined gap vector and user profile embedding vector , through the policy function Generate optimal reverse questions In selecting the optimal question path, the system uses the information gain function as a value benchmark to maximize the uncertainty compression effect of each round of interaction on the system's cognitive map, ensuring the targeted and efficient nature of the questions. For more detailed implementation, please refer to the description of the reverse question strategy generator section and will not be repeated here.

[0036] Problem Processing Layer: This layer utilizes a hierarchical problem decomposition engine. It employs semantic tree-based problem decomposition techniques to decompose complex problems and improves processing efficiency through parallel sub-problem processing. Using a fusion mechanism based on evidence theory, the results of the sub-problems are integrated to form a complete and accurate problem solution.

[0037] After the problem processing layer receives the original question or the question from the strategy generation layer, it uses the semantic tree builder to perform semantic structured decomposition on the task and obtain a set of sub-questions. While multithreading and concurrently processing subtasks, the system constructs a trust allocation function based on evidence theory to evaluate the credible intervals of the answers to each subproblem, integrating them into a comprehensive solution to the problem and ensuring the rationality of integrating multidimensional information sources under uncertain conditions. For more detailed implementation, please refer to the description of the hierarchical problem decomposition engine and will not be repeated here.

[0038] Output layer: This layer uses a personalized solution synthesizer. It performs dynamic style conversion based on user profile matching results, generates multimodal output, and provides users with personalized solutions. At the same time, it collects user feedback to provide data support for system optimization and self-evolution.

[0039] The output layer is based on the problem solution generated by the hierarchical problem decomposition engine and user portraits Perform personalized content re-encoding. The system introduces a style adaptation network to migrate the original content to the target user style domain and construct an output response. , is the style transformation mapping function. Meanwhile, the output layer constructs a reward function using user feedback labels and reinforcement learning to evaluate and update the overall interaction path, driving the system strategy to evolve in real user interactions, ultimately achieving cognitively driven optimization of the intelligent question-answering closed loop.

[0040] In the entire hierarchical intelligent interaction architecture, the system dynamically decides whether to trigger reverse questioning or enter the solution generation phase based on the information sufficiency assessment results, achieving efficient coordination of information acquisition and processing. Specifically, the system proposed in this invention has the following beneficial effects: First, the efficiency and accuracy of information acquisition are significantly enhanced. The information sufficiency evaluator models information entropy and performs semantic vector analysis on user input, not only quantitatively determining the structural integrity and key content of the input but also providing high-precision gap information support for the reverse questioning strategy generator. Based on this, the strategy generator, using the MDP and information value maximization algorithms, generates an optimal question-guided sequence, ensuring that supplementary information achieves maximum cognitive gain at minimal interaction cost, fundamentally improving the system's information control capabilities.

[0041] Secondly, it effectively supports the structured decomposition and efficient solution of complex problems. Through a semantic tree-driven recursive problem decomposition mechanism, the system automatically maps unstructured complex problems into subtask units that can be computed in parallel, and uses a shared memory pool architecture to maintain contextual consistency between sub-problems. The subsequent fusion stage integrates and optimizes uncertain solutions based on evidence theory, ensuring stable and reliable solution output even in scenarios with fragmented information or incomplete reasoning chains.

[0042] Third, it enables highly personalized content generation and output adaptation. Based on the cognitive dimensions, interaction preferences, and content comprehension models in user profiles, the system dynamically matches output style and interaction methods. It then recodes the content using a style mapping function, ensuring that the output fully aligns with the target user in terms of language style, information granularity, and expression, achieving the transition from "correct answers" to "good answers." The reinforcement learning module incorporates user feedback as a signal for policy updates, enabling the system to continuously self-optimize and establish a cognitive interaction network with evolutionary characteristics.

[0043] Finally, a stable and controllable closed-loop dialogue and task-oriented interaction model is constructed. Through a state-machine-driven multi-round dialogue management mechanism, combined with context vector modeling technology, the system is capable of managing semantic continuity, maintaining goal-oriented consistency and semantic cohesion integrity across multiple rounds of complex interactions. Each functional module operates collaboratively within a clear logical chain, significantly improving the system's interpretability, controllability, and deployment robustness.

[0044] In summary, the reverse-guided intelligent interactive system for dynamic information evaluation proposed in the embodiment of the present invention takes the information sufficiency evaluator as the starting point, and relies on the reverse questioning strategy generator, the hierarchical problem decomposition engine and the personalized solution synthesizer to build a dynamic, adaptive and highly robust intelligent question-answering system cognitive architecture. This architecture integrates information evaluation, strategy generation, problem solving and personalized output functions to achieve full-link closed-loop control from user input evaluation to high-matching output. This solution not only achieves the leap from passive response to active guidance in intelligent question-answering, but also constructs a cognitive-driven information recognition, problem reconstruction and content synthesis closed-loop system in complex scenarios, providing a practical and high-performance solution path for interactive artificial intelligence applications that meet professional, high-complexity and personalized needs.

[0045] Second, see Figure 8 The embodiment of the present invention provides a reverse-guided intelligent interaction method for implementing dynamic information evaluation, the method comprising: S10. Based on information entropy theory and combined with task context information, dynamically quantify and evaluate user input information to identify key information gaps. S20. If a critical information gap is identified, the expected information gain and user acceptance of each candidate question are calculated based on the Markov decision process model, and an optimal reverse guidance question sequence is generated based on the expected information gain and user acceptance to guide the user to provide supplementary information to fill the critical information gap; S30. The question input by the user after being supplemented by the reverse questioning strategy generator is judged by a dynamic question complexity judgment mechanism. When the question input by the user is judged to be a complex question, the semantic tree decomposition technology is used to recursively decompose the complex question into multiple independently processable sub-questions. The solution results of each sub-question are integrated using the evidence theory fusion mechanism to form a problem solution. The sub-questions share information among themselves through a shared memory pool framework. S40. Analyze the cognitive level, learning style, and attention characteristics of the user profile, optimize the matching of the problem solution with the user profile, calculate and measure the semantic fit between the problem solution and the user profile, and introduce the semantic fit as a reward function into the reinforcement learning framework. Use the reinforcement learning framework to continuously optimize the output strategy based on user feedback to generate personalized solutions that are highly matched with the user profile in terms of content depth, formal style, and presentation method.

[0046] As for the method embodiment of the second aspect, since it is basically similar to the system embodiment of the first aspect, the description is relatively simple, and the relevant parts can be referred to the partial description of the system embodiment of the first aspect.

[0047] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0048] Although the present invention is described herein in conjunction with various embodiments, those skilled in the art may understand and implement other variations of the disclosed embodiments by reviewing the specification and accompanying drawings in the process of implementing the claimed invention. In the specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple components or steps. The fact that certain measures are described in different embodiments does not mean that these measures cannot be combined to produce good results.

[0049] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A reverse-guidance intelligent interactive system for dynamic information evaluation, characterized in that: The system includes an information sufficiency evaluator, a reverse questioning strategy generator, a hierarchical problem decomposition engine, and a personalized solution synthesizer; wherein, The information sufficiency evaluator is used to dynamically and quantitatively evaluate user input information based on information entropy theory and in combination with task context information to identify key information gaps; The reverse questioning strategy generator is configured to calculate the expected information gain and user acceptance of each candidate question based on a Markov decision process model if a key information gap is identified, and generate an optimal reverse guidance question sequence based on the expected information gain and user acceptance to guide the user to provide supplementary information to fill the key information gap; The hierarchical problem decomposition engine is used to judge the user-input question supplemented by the reverse questioning strategy generator through a dynamic problem complexity judgment mechanism. When the user-input question is judged to be complex, the semantic tree decomposition technology is used to recursively decompose the complex problem into multiple independently processable sub-problems. The solution results of each sub-problem are integrated using the evidence theory fusion mechanism to form a problem solution. The sub-problems share information among themselves through a shared memory pool framework. The personalized solution synthesizer is used to analyze the cognitive level, learning style, and attention characteristics of the user profile, optimize the matching of the problem solution with the user profile, calculate and measure the semantic fit between the problem solution and the user profile, and introduce the semantic fit as a reward function into the reinforcement learning framework, so as to use the reinforcement learning framework to continuously optimize the output strategy based on user feedback to generate personalized solutions that are highly matched with the user profile in terms of content depth, formal style, and presentation method.

2. The reverse-guided intelligent interactive system for realizing dynamic information evaluation according to claim 1, characterized in that: The information sufficiency evaluator is based on information entropy theory and combines task context information to dynamically and quantitatively evaluate user input information to identify key information gaps, including: Based on multi-dimensional information vector representation technology, user input information is converted into high-dimensional semantic vectors; Calculating the probability of each information point in the high-dimensional semantic vector being covered, and calculating the information entropy of the high-dimensional semantic vector based on all the probabilities; Combined with task context information, the task criticality coefficient, problem domain complexity factor and user professional ability weight are introduced; Calculate the information completeness score based on the domain knowledge graph and the information entropy, as well as the task criticality coefficient, the problem domain complexity factor and the user professional ability weight; Based on the task criticality coefficient, the problem domain complexity factor and the user professional ability weight, the evaluation threshold is dynamically adjusted to determine whether the information completeness score is less than the dynamically adjusted evaluation threshold. If so, there is a key information gap in the user input information.

3. The reverse-guided intelligent interactive system for realizing dynamic information evaluation according to claim 2, characterized in that: The formula for calculating the information completeness score is: ； in, represents the information completeness score, represents information entropy, , Represents the first The probability that an information point is covered, Indicates the number of information points in the high-dimensional semantic vector, represents the theoretical maximum information entropy calculated based on the domain knowledge graph, represents the mission criticality coefficient, represents the complexity factor of the problem domain, Indicates the user's professional ability weight.

4. The reverse-guided intelligent interactive system for realizing dynamic information evaluation according to claim 1, characterized in that: The reverse questioning strategy generator calculates the expected information gain and user acceptance of each candidate question based on a Markov decision process model, and generates an optimal reverse guidance question sequence based on the expected information gain and user acceptance, aiming to guide the user to provide supplementary information to fill the key information gap, including: Modeling an MDP state, which includes a state space, an action space, a state transition probability, a reward function, and a discount factor. The state space is the concatenation of user input information and user profiles, the action space is the set of all candidate questions, the state transition probability is used to measure the probability of user feedback after asking a question, the reward function is used to measure the expected information gain brought by the question, and the discount factor is used to control the impact of long-term benefits. Based on the MDP state, an optimal questioning strategy function is constructed with the goal of maximizing the long-term cumulative information value; Calculating a first information entropy in a current state space, and calculating a second information entropy in an expected state space after asking the question according to the optimal questioning strategy function, and calculating an information entropy decrease value according to the first information entropy and the second information entropy; Constructing a comprehensive score function for each action in the action space under the current state space according to the information entropy reduction value and the user acceptance weight; The optimal value of the comprehensive score function is solved by an iterative algorithm, and combined with a personalized question template library, an optimal reverse guidance question sequence is generated to guide the user to fill the key information gap.

5. The reverse-guided intelligent interactive system for realizing dynamic information evaluation according to claim 4, characterized in that: The constructed comprehensive score function is expressed as follows: ； in, Indicates status Take the following action to ask questions The comprehensive score of , represents the state space, , represents the action space, Indicates status Take the following action to ask questions The information entropy decreases, Represents the user acceptance weight.

6. The reverse-guided intelligent interactive system for realizing dynamic information evaluation according to claim 1, characterized in that: The hierarchical problem decomposition engine uses semantic tree decomposition technology to recursively decompose the complex problem into multiple independently processable sub-problems, including: Based on syntactic analysis and dependency extraction technology, a root node is constructed to represent the overall problem semantics, a leaf node represents the smallest semantic unit, and each intermediate node corresponds to the semantic combination of subtasks, so as to recursively decompose the complex problem into multiple sub-problems that can be processed independently.

7. The reverse-guided intelligent interactive system for realizing dynamic information evaluation according to claim 1, characterized in that: The hierarchical problem decomposition engine uses the evidence theory fusion mechanism to fuse the solution results of each sub-problem, including: Construct a trust allocation function; Calculate the trust level of the corresponding answer of any two sub-questions according to the trust allocation function, and then calculate the corresponding fused trust level based on the two trust levels obtained by calculation; Determine whether the trust level after fusion is less than the preset threshold. If it is less than, trigger the reverse question strategy generator to regenerate the optimal reverse guidance question sequence. If it is greater than or equal to, output the fusion result.

8. The reverse-guided intelligent interactive system for realizing dynamic information evaluation according to claim 1, characterized in that: The personalized solution synthesizer analyzes the user profile's cognitive level, learning style, and attention characteristics, optimizes the matching between the problem solution and the user profile, calculates and measures the semantic fit between the problem solution and the user profile, and introduces the semantic fit as a reward function into the reinforcement learning framework. The reinforcement learning framework is then used to continuously optimize the output strategy based on user feedback to generate personalized solutions that are highly matched to the user profile in terms of content depth, formal style, and presentation method, including: Model user profiles as multidimensional feature vectors including cognitive level, learning style, and attention characteristics; Mapping the problem solution to a semantic output space, and constructing a semantic fit function based on the multidimensional feature vector and the problem solution mapped to the semantic output space; The semantic fit function is introduced into the reinforcement learning framework as a reward function, and the output strategy is iteratively optimized through the policy optimization gradient algorithm to generate a personalized solution that is highly matched with user characteristics in terms of content depth, formal style and presentation method.

9. The reverse-guided intelligent interactive system for realizing dynamic information evaluation according to claim 8, characterized in that: The constructed semantic fit function is expressed as follows: ； in, represents a multidimensional feature vector, Indicates a solution to the problem, Indicates the number of answer results in the problem solution, Indicates the i The weight of each answer in the overall problem solution, express Middle i The answer results, Represents a multidimensional feature vector In the i The user feature transformation value under the solution result, Indicates solution and The semantic similarity between .

10. A reverse guided intelligent interaction method for realizing dynamic information evaluation, characterized in that: The method comprises: Based on information entropy theory and combined with task context information, user input information is dynamically and quantitatively evaluated to identify key information gaps; If a critical information gap is identified, the expected information gain and user acceptance of each candidate question are calculated based on the Markov decision process model, and an optimal reverse guidance question sequence is generated based on the expected information gain and the user acceptance to guide the user to provide supplementary information to fill the critical information gap; The user-input question, supplemented by the reverse questioning strategy generator, is judged through a dynamic problem complexity judgment mechanism. When the user-input question is judged to be complex, the semantic tree decomposition technology is used to recursively decompose the complex question into multiple independently processable sub-problems. The evidence theory fusion mechanism is used to fuse the answers to the sub-problems to form a problem solution. The sub-problems share information among themselves through a shared memory pool framework. Analyze the cognitive level, learning style, and attention characteristics of the user portrait, optimize the matching of the problem solution with the user portrait, calculate and measure the semantic fit between the problem solution and the user portrait, and introduce the semantic fit as a reward function into the reinforcement learning framework. Use the reinforcement learning framework to continuously optimize the output strategy based on user feedback to generate personalized solutions that are highly matched with the user portrait in terms of content depth, formal style, and presentation method.

Images