AI intelligent interrogation teaching system and method

Abstract

The invention relates to the technical field of judicial education, and discloses an AI intelligent interrogation teaching system and method, and the system comprises a virtual interrogation scene generation module which is used for generating the language, expression, limb movement and environment change of a suspect in a virtual interrogation scene; the multi-modal analysis module is used for capturing and fusing voice, expression and text features of the suspect and generating a current emotional state vector; the strategy reasoning module is used for generating an optimal questioning strategy based on the questioning content of the trainee and the emotional state of the suspect; and the optimization suggestion generation module is used for evaluating the questioning performance of the trainees in real time and generating optimization suggestions. Through the real-time strategy optimization module, the questioning effect can be dynamically evaluated, the optimal strategy suggestion can be generated, the reinforcement learning model is combined with multi-modal emotion analysis, intelligent adjustment of the questioning strategy is achieved, students can put forward questions with more pertinence and effectiveness, and the interrogation efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of judicial education technology, and in particular to an AI intelligent interrogation teaching system and method. Background Art

[0002] As an important means of obtaining key information of a case in judicial practice, the effect of interrogation directly affects the process and quality of case detection. For a long time, interrogation teaching has been an important part of judicial training, providing support for case investigation by cultivating the interrogators' questioning ability, psychological judgment ability and logical reasoning ability. At present, the main methods of interrogation teaching include classroom lectures, case analysis and role simulation, which play an important role in the basic teaching of interrogation skills. However, with the increasing complexity of cases and the increasing requirements for interrogation quality and efficiency in the judicial environment, the traditional teaching model still has room for further development in improving the systematic, practical and personalized nature of interrogation skills.

[0003] First, interrogation teaching has increasingly higher requirements for the authenticity of scenario simulation. In actual interrogations, the psychological state, emotional response, and language behavior of suspects are often complex and changeable, requiring interrogators to have the ability to make quick judgments and respond flexibly. In traditional interrogation teaching methods, simulation scenarios are often based on fixed cases or artificial role-playing. This method is difficult to fully simulate the behavior patterns of suspects in real environments, such as sudden mood swings, logical confrontations, or resistance to emotional offensive strategies. In addition, how to adjust questioning strategies based on suspect reactions in dynamic scenarios is also a key issue that needs to be addressed in current interrogation skills teaching.

[0004] Secondly, the choice of questioning strategy during the interrogation process is highly context-dependent. Questioning must not only meet the needs of case information mining, but also be adjusted according to the psychological characteristics and emotional state of the suspect. For example, suspects in different case types (such as economic crimes and violent crimes) may show completely different behavior patterns, and the logic, guidance and emotional effects of questioning will also change accordingly. Existing teaching methods focus more on the teaching of general skills, while the optimization of strategies for specific situations often relies on the students' personal experience accumulation, which makes students with different backgrounds and experience levels have certain differences in mastering complex questioning skills.

[0005] In addition, interrogation teaching places higher demands on students' real-time feedback and personalized support. In modern teaching concepts, students' ability growth needs to rely on immediate evaluation and guidance. In interrogation teaching, how to combine students' questioning logic, suspects' reactions, and the progress of the case to dynamically provide targeted feedback and optimization suggestions is an important direction worthy of further exploration. At the same time, different students have different learning needs. For example, beginners may need more basic skills training, while students with certain interrogation experience may need more practical exercises in complex scenarios. This personalized demand needs to be met by dynamically adjusting teaching content and paths. Summary of the invention

[0006] In view of the deficiencies of the prior art, the present invention provides an AI intelligent interrogation teaching system, which solves the problems of lack of real dynamic scene simulation in interrogation teaching, difficulty in real-time optimization of questioning strategies, and inability to dynamically adjust students' personalized learning paths.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: AI intelligent interrogation teaching system and method, including:

[0008] A virtual interrogation scene generation module, which is used to generate the suspect's language, facial expressions, body movements, and environmental changes in the virtual interrogation scene;

[0009] Multimodal analysis module, which is used to capture and fuse the suspect's voice, expression, and text features to generate the current emotional state vector;

[0010] A strategy reasoning module, which is used to generate the best questioning strategy based on the content of the student's question and the suspect's emotional state;

[0011] Optimization suggestion generation module, which is used to evaluate students' question performance in real time and generate optimization suggestions;

[0012] The teaching evaluation module is used to evaluate students' performance and dynamically adjust the teaching path.

[0013] Preferably, the virtual interrogation scene generation module uses a generative AI model to generate the suspect's verbal response, while inputting case background, student questions, and the suspect's current emotional state, and uses a conditional generative adversarial network to generate the suspect's facial expressions, body movements, and voice intonation;

[0014] The generative AI model simulates random perturbations in the environment of the interrogation scene through Markov random fields.

[0015] Preferably, the multimodal analysis module processes data including the following steps:

[0016] a. Use speech-to-text models to extract speech content and analyze speech rate, intonation, and volume;

[0017] b. Use convolutional neural networks to analyze the suspect’s facial expressions and generate expression embeddings;

[0018] c. Use the natural language processing model to perform semantic analysis on the students’ questions and the suspects’ responses and generate semantic embeddings;

[0019] d. Based on the multimodal attention mechanism, the speech, expression and text features are integrated to generate the current emotional state vector.

[0020] Preferably, in step d, the multimodal attention mechanism fuses features according to the following formula:

[0021]

[0022] Among them, E s is the facial expression feature embedding, V t is the speech feature embedding, H t is the text semantic feature embedding, Q, K, V are feature vector matrices, d k The dimension of attention.

[0023] Preferably, the strategy reasoning module generates a questioning strategy through the following steps:

[0024] A. Define the state space, including the suspect's emotional state and the student's questioning history;

[0025] B. Define the action space, including the possible questioning strategies;

[0026] C. Evaluate the effectiveness of the questioning strategy based on the reward function;

[0027] The reward function formula is as follows:

[0028] R t =w1·r Emotion +w2·r InfoGain +w3·r Logic

[0029] Among them, w1, w2, w3 are weight factors, r Emotion , r InfoGain , r Logic Represent the scores of emotional impact, information gain and logic respectively.

[0030] Preferably, the strategy reasoning module calculates the fitness of the questioning strategy by the following formula:

[0031] Score(S,Z t )=GNN(V,E)

[0032] Among them, S is the questioning strategy, Z tis the current emotional state vector, V and E are the embedded representations of the nodes and edges of the knowledge graph, respectively.

[0033] Preferably, the optimization suggestion generating module includes:

[0034] Generate optimization suggestions based on current questioning strategy and suspect status;

[0035] Generate optimization suggestion text using natural language generation technology;

[0036] Provides graphical feedback, including sentiment curves, question scores, and optimization directions.

[0037] Preferably, the teaching evaluation module calculates the student ability score by the following formula:

[0038]

[0039] Among them, F i is the student's score on the i-th indicator, w i is the weight of the corresponding indicator.

[0040] Preferably, the teaching evaluation module plans a personalized learning path through the following formula:

[0041]

[0042] Among them, L t is the learning path of the current time step, w t is the time step weight, R t is the profit value of the current time step.

[0043] The AI ​​intelligent interrogation teaching method includes the following steps:

[0044] S1. Virtual scene generation: Generate a virtual interrogation scene based on the case background and the students’ questions, including the suspect’s language, facial expressions, intonation, and body movements;

[0045] S2, multimodal data processing: capture the suspect's voice, expression, and text features, and generate the current emotional state vector through multimodal fusion;

[0046] S3, strategy reasoning: calculate the best questioning strategy based on reinforcement learning and knowledge graph reasoning modules;

[0047] S4. Real-time feedback generation: real-time evaluation of students’ questioning performance and generation of optimization suggestions;

[0048] S5. Dynamic teaching evaluation: record students’ performance and dynamically adjust teaching paths to achieve personalized learning.

[0049] The present invention provides an AI intelligent interrogation teaching system and method. It has the following beneficial effects:

[0050] 1. Through the real-time strategy optimization module, the system can dynamically evaluate the questioning effect and generate the best strategy suggestions according to the content of the students' questions and the emotional state of the suspect. The reinforcement learning model is combined with multimodal sentiment analysis to realize the intelligent adjustment of the questioning strategy, so that students can ask more targeted and effective questions according to the psychological state of the suspect and the case situation, thereby improving the efficiency of interrogation.

[0051] 2. The present invention is based on reinforcement learning and knowledge graph reasoning, and can dynamically optimize the students' questioning strategies in real-time scenarios. No matter how the suspect's emotions change, the system can optimize the questioning path through the reward function and generate the optimal questioning strategy that suits the current situation. This dynamic adjustment capability helps students learn to respond flexibly in complex interrogation situations and avoid the limitations of a single strategy.

[0052] 3. The present invention realizes a closed-loop mechanism of "training-feedback-optimization" through real-time evaluation and feedback of students' questioning performance. The system combines the emotional fluctuation curve and strategy score to provide students with intuitive optimization suggestions, enabling them to quickly adjust their questioning logic and strategy, improve training efficiency, and shorten the learning cycle.

[0053] 4. Through real-time emotion analysis and strategy optimization, the system of the present invention can help students learn to effectively guide the emotions of suspects during interrogation and master complex interrogation strategies such as emotional attack and logical deterrence. The cultivation of this ability not only improves the students' interrogation skills, but also enhances their psychological response and control capabilities in real scenarios.

[0054] 5. The present invention ensures the intelligence and standardization of the training process by combining reinforcement learning, knowledge graph reasoning and multimodal data analysis. Trainees can obtain unified strategy optimization support during the training process and continuously improve their questioning ability through intelligent suggestions, providing judicial institutions with a replicable high-quality interrogation teaching solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a system logic diagram of the present invention;

[0056] Figure 2 It is a schematic diagram of a module of the present invention;

[0057] Figure 3 It is a flow chart of the steps of the present invention. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the specification of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] Embodiment 1:

[0060] Please see attached Figure 1 -Attached Figure 2 , an embodiment of the present invention provides an AI intelligent interrogation teaching system, including:

[0061] A virtual interrogation scene generation module, which is used to generate the suspect's language, facial expressions, body movements, and environmental changes in the virtual interrogation scene;

[0062] Multimodal analysis module, which is used to capture and fuse the suspect's voice, expression, and text features to generate the current emotional state vector;

[0063] A strategy reasoning module, which is used to generate the best questioning strategy based on the content of the student's question and the suspect's emotional state;

[0064] Optimization suggestion generation module, which is used to evaluate students' question performance in real time and generate optimization suggestions;

[0065] Teaching evaluation module, which is used to evaluate students' performance and dynamically adjust the teaching path;

[0066] The virtual interrogation scene generation module uses a generative AI model to generate the suspect's verbal responses. The input includes the case background, the student's questions, and the suspect's current emotional state. The conditional generative adversarial network is used to generate the suspect's facial expressions, body movements, and voice intonation.

[0067] The generative AI model simulates random perturbations in the environment of the interrogation scene through Markov random fields;

[0068] The multimodal analysis module processes data in the following steps:

[0069] a. Use speech-to-text models to extract speech content and analyze speech rate, intonation, and volume;

[0070] b. Use convolutional neural networks to analyze the suspect’s facial expressions and generate expression embeddings;

[0071] c. Use the natural language processing model to perform semantic analysis on the students’ questions and the suspects’ responses and generate semantic embeddings;

[0072] d. Based on the multimodal attention mechanism, the speech, expression and text features are integrated to generate the current emotional state vector;

[0073] In step d, the multimodal attention mechanism fuses features according to the following formula:

[0074]

[0075] Among them, E s is the facial expression feature embedding, V t is the speech feature embedding, H t is the text semantic feature embedding, Q, K, V are feature vector matrices, d k is the attention dimension;

[0076] The policy reasoning module generates a questioning strategy through the following steps:

[0077] A. Define the state space, including the suspect's emotional state and the student's questioning history;

[0078] B. Define the action space, including the possible questioning strategies;

[0079] C. Evaluate the effectiveness of the questioning strategy based on the reward function;

[0080] The reward function formula is as follows:

[0081] R t =w1·r Emotion +w2·r InfoGain +w3·r Logic

[0082] Among them, w1, w2, w3 are weight factors, r Emotion , r InfoGain , r Logic They represent the scores of emotional impact, information gain, and logic, respectively;

[0083] The strategy reasoning module calculates the fitness of the question strategy using the following formula:

[0084] Score(S,Z t )=GNN(V,E)

[0085] Among them, S is the questioning strategy, Z t is the current emotional state vector, V and E are the embedded representations of nodes and edges in the knowledge graph respectively;

[0086] The optimization suggestion generation module includes:

[0087] Generate optimization suggestions based on current questioning strategy and suspect status;

[0088] Generate optimization suggestion text using natural language generation technology;

[0089] Provide graphical feedback, including sentiment curves, question scores, and optimization directions;

[0090] The teaching evaluation module calculates the student ability score through the following formula:

[0091]

[0092] Among them, F i is the student's score on the i-th indicator, w i is the weight of the corresponding indicator;

[0093] The teaching evaluation module plans personalized learning paths through the following formula:

[0094]

[0095] Among them, L t is the learning path of the current time step, w t is the time step weight, R t is the further benefit value at the current time.

[0096] Specifically, the system consists of a virtual interrogation scene generation module, a multimodal emotion and semantic analysis module, a strategy reasoning module, an optimization suggestion generation module, and a teaching evaluation and path planning module. Through the collaborative work of these five modules, the system realizes dynamic scene generation, real-time emotion and semantic analysis, intelligent reasoning and optimization suggestion generation of questioning strategies for virtual interrogation training, and realizes personalized learning path planning through dynamic teaching evaluation. The specific implementation methods are as follows:

[0097] The system first constructs a multimodal dynamic interrogation scene through a virtual interrogation scene generation module, including the multimodal expression of the suspect's language, expression, tone of voice and body movements, and the dynamic changes of the scene environment. The suspect's language is generated by a generative AI model. The input data includes the case background, the content of the students' questions, and the suspect's current emotional state, and the output is a logically consistent and scenario-adapted language response. In order to improve the authenticity of the suspect's performance, the system uses a conditional generative adversarial network to generate expressions, intonations and body movements. The generator outputs dynamic behavior characteristics by inputting the suspect's current emotional state and language content, and the discriminator optimizes the authenticity of the generated behavior. The loss function between the generator and the discriminator uses the standard GAN formula to ensure that the generated behavior meets the requirements of the interrogation scene. In addition, in order to simulate the real interrogation environment, the dynamic changes in the scene are modeled through Markov random fields to control the random perturbations of scene elements, making interrogation training more realistic.

[0098] During the virtual interrogation process, the multimodal emotion and semantic analysis module captures and analyzes the suspect’s voice, expression, and text features in real time to generate the current emotional state vector Z tThe speech features are extracted by the speech-to-text model, and parameters such as speech rate, intonation, and volume are analyzed, and the emotional intensity is calculated through support vector regression; the expression features are analyzed by deep convolutional neural networks on video frames to extract expression embeddings; the text features are analyzed by the BERT model for semantic embedding to generate semantic vectors H t The above features are fused through the multimodal attention mechanism to finally generate the emotional state vector Z t ,This vector contains the comprehensive psychological state of the suspect in the current situation, which can provide an important basis for the subsequent questioning strategy reasoning.,The fusion process adopts a computational method based on the attention mechanism to enhance the modeling ability of the association between semantics and emotions.

[0099] The strategy reasoning module uses the student's question history and the suspect's emotional state as input, and generates the optimal questioning strategy through knowledge graph reasoning and reinforcement learning model. The knowledge graph consists of interrogation strategies, suspect psychological characteristics and case-related factors, and uses graph embedding method to calculate the adaptability of the current questioning strategy to the scenario. The reinforcement learning part uses deep Q learning to optimize the question path, state and question logic.

[0100] Based on the strategy reasoning module, the optimization suggestion generation module evaluates the students' questioning performance in real time and generates optimization suggestions. This module uses natural language generation technology to generate specific suggestion texts, including the improvement direction of questions and the adjustment method of language expression. At the same time, the system displays optimization suggestions through graphical feedback, including the suspect's emotional fluctuation curve, questioning strategy score and improvement direction, to help students quickly understand the optimization goals. The real-time nature of this module ensures that students can get effective feedback in real time during the training process and improve learning effects.

[0101] The teaching evaluation and path planning module is used to record the performance of students in interrogation training and dynamically adjust the learning path according to the performance. The system uses a comprehensive ability scoring mechanism to quantitatively evaluate the students' questioning logic, case advancement effect, and emotional control ability. Based on the scoring results, the system uses a path optimization algorithm to generate a personalized learning path for the students.

[0102] Through the collaborative work of the above modules, the AI ​​intelligent interrogation teaching system of the present invention realizes the full-process functions from virtual scene generation, real-time multimodal analysis, questioning strategy reasoning and optimization to dynamic teaching evaluation. Its open architecture enables it to support the training needs of different case backgrounds and student ability levels, while providing a highly targeted and real-time feedback mechanism. The above technical details fully disclose the system implementation process and its principles, and provide innovative solutions for intelligent applications in the field of judicial education.

[0103] Embodiment 2:

[0104] Please see attached Figure 3 , AI intelligent interrogation teaching method, according to any one of claims 1-9 of the AI ​​intelligent interrogation teaching system, characterized in that it includes the following steps:

[0105] S1. Virtual scene generation: Generate a virtual interrogation scene based on the case background and the students’ questions, including the suspect’s language, facial expressions, intonation, and body movements;

[0106] S2, multimodal data processing: capture the suspect's voice, expression, and text features, and generate the current emotional state vector through multimodal fusion;

[0107] S3, strategy reasoning: calculate the best questioning strategy based on reinforcement learning and knowledge graph reasoning modules;

[0108] S4. Real-time feedback generation: real-time evaluation of students’ questioning performance and generation of optimization suggestions;

[0109] S5. Dynamic teaching evaluation: record students’ performance and dynamically adjust teaching paths to achieve personalized learning.

[0110] Specifically, in step S1, virtual scene generation is the first step of this method, which is used to create a highly simulated virtual interrogation environment for trainees to interact with simulated suspects. In this step, the system combines generative AI technology to generate the suspect's language response, and uses conditional generative adversarial networks to generate dynamic behaviors such as expressions, intonations, and body movements. The input for suspect language generation includes case background information, the content of the trainee's questions, and the suspect's current emotional state. The generative model ensures the logic and situational adaptability of the generated content through the contextual attention mechanism. In addition, to enhance the immersion of the virtual scene, the system generates the suspect's multimodal dynamic behavior, including facial expressions, voice intonation, and body movements, through an adversarial network. The generator generates dynamic behavior features by inputting the suspect's emotional state and language content; the discriminator learns to discriminate the authenticity of the generated behavior, iteratively optimizes the output of the generator, and makes the generated behavior consistent with human natural performance. The dynamic environmental changes in the scene are modeled through Markov random fields to ensure the authenticity and randomness of the scene, so that trainees can conduct interrogation training in an environment close to the real world.

[0111] In step S2, during the interaction with the suspect, the system captures and processes the suspect’s voice, expression, and text features in real time, and generates the suspect’s emotional state vector Z through multimodal data fusion. t Specifically, the voice data is transcribed through a speech-to-text model to extract parameters such as speech rate, intonation, and volume; these parameters are further analyzed based on the support vector regression algorithm to obtain the intensity value of the voice emotion. At the same time, the system uses a deep convolutional neural network to analyze the suspect's facial expressions and body movements in the video frame and extract the expression embedding vector Es , and perform semantic analysis on the text content through the natural language processing model to generate text semantic embedding H t Finally, the speech, expression and text features are fused through the multimodal attention mechanism to generate the emotional state vector Z t This vector contains the suspect’s comprehensive psychological state in the current situation, including his emotional fluctuations and language behavior characteristics, providing input data for the reasoning of subsequent questioning strategies.

[0112] In step S3, strategy reasoning is the core step of this method, which is used to determine the suspect’s emotional state Z. t The system generates the optimal questioning strategy based on the questioning history of the students. In this step, the system completes the strategy selection by combining knowledge graph reasoning and reinforcement learning. The knowledge graph consists of interrogation strategies, suspect psychological characteristics, and case information. The graph embedding algorithm is used to calculate the fitness of the questioning strategy in the current situation. At the same time, the reinforcement learning model dynamically optimizes the questioning path through deep Q learning. The state space of the model includes the suspect's emotional state Z t and the history of students’ questions, action space A t is all the optional questioning strategies. The reward function comprehensively considers the three dimensions of the suspect's emotional fluctuations (emotional influence), case information advancement (information gain), and question logic. Through the optimization of the reward function, the system can dynamically select the best questioning strategy to achieve a balance between emotional guidance and case advancement.

[0113] In step S4, real-time feedback generation is the process by which the system dynamically evaluates the students' questioning performance and provides optimization suggestions. In this step, the system uses the optimization suggestion generation module to evaluate the students' questioning performance in real time, including the logic of the questions, the impact on the suspect's emotions, and the contribution to the advancement of the case. Based on these evaluation results, the system uses natural language generation technology to generate optimization suggestions, providing students with natural language, specific and actionable improvement directions, such as how to adjust the tone, structure or content of questions. The system also displays the students' questioning effects through graphical feedback, including a curve chart of the suspect's emotional fluctuations, a comparison of questioning strategy scores, and improvement directions, so that students can intuitively understand the advantages and disadvantages of their own questioning. At the same time, the generation and display of real-time feedback ensures that students can adjust their questioning strategies in real time, thereby improving interrogation results.

[0114] In step S5, this step quantitatively evaluates the students' performance through the dynamic teaching evaluation module, and provides students with personalized learning paths in combination with the path planning module. During the evaluation process, the system scores students based on their performance in multiple dimensions such as question logic, case advancement effect, and emotional control ability. Based on the scoring results, the system further analyzes the students' performance characteristics and areas that need improvement, and uses the path optimization algorithm to generate a personalized learning path for the students. The results of path optimization are used to recommend the next step of learning content, including interrogation scenarios for intensive training, simulated suspect types, and directions for skill improvement, to ensure that students can continuously improve their abilities through continuous practice.

[0115] Through the collaborative work of the above five steps, the AI ​​intelligent interrogation teaching method of the present invention realizes a complete process from virtual scene generation, emotion and semantic analysis, questioning strategy reasoning and optimization to dynamic evaluation and learning path planning. This method supports diverse case backgrounds and interrogation scenarios, can respond to the interaction needs of trainees and suspects in real time, and provide trainees with highly personalized and situation-adapted interrogation skills training, while taking into account the logic, emotional guidance and case advancement goals of the interrogation process. The operability and replicability of the method are ensured by the disclosure of technical details.

[0116] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. AI intelligent interrogation teaching system, characterized by: include: A virtual interrogation scene generation module, which is used to generate the suspect's language, facial expressions, body movements, and environmental changes in the virtual interrogation scene; Multimodal analysis module, which is used to capture and fuse the suspect's voice, expression, and text features to generate the current emotional state vector; A strategy reasoning module, which is used to generate the best questioning strategy based on the content of the student's question and the suspect's emotional state; Optimization suggestion generation module, which is used to evaluate students' question performance in real time and generate optimization suggestions; The teaching evaluation module is used to evaluate students' performance and dynamically adjust the teaching path.

2. The AI ​​intelligent interrogation teaching system according to claim 1 is characterized in that: The virtual interrogation scene generation module uses a generative AI model to generate the suspect's verbal response, while inputting case background, student questions, and the suspect's current emotional state, and uses a conditional generative adversarial network to generate the suspect's facial expressions, body movements, and voice intonation; The generative AI model simulates random perturbations in the environment of the interrogation scene through Markov random fields.

3. The AI ​​intelligent interrogation teaching system according to claim 1 is characterized in that: The multimodal analysis module processes data including the following steps: a. Use speech-to-text models to extract speech content and analyze speech rate, intonation, and volume; b. Use convolutional neural networks to analyze the suspect’s facial expressions and generate expression embeddings; c. Use the natural language processing model to perform semantic analysis on the students’ questions and the suspects’ responses and generate semantic embeddings; d. Based on the multimodal attention mechanism, the speech, expression and text features are integrated to generate the current emotional state vector.

4. The AI ​​intelligent interrogation teaching system according to claim 3 is characterized in that: In step d, the multimodal attention mechanism fuses features according to the following formula: Among them, E s is the facial expression feature embedding, V t is the speech feature embedding, H t is the text semantic feature embedding, Q, K, V are feature vector matrices, d k The dimension of attention.

5. The AI ​​intelligent interrogation teaching system according to claim 1 is characterized in that: The strategy reasoning module generates a question strategy through the following steps: A. Define the state space, including the suspect's emotional state and the student's questioning history; B. Define the action space, including the possible questioning strategies; C. Evaluate the effectiveness of the questioning strategy based on the reward function; The reward function formula is as follows: R t =w1·r Emotion +w2·r InfoGain +w3·r Logic Among them, w1, w2, w3 are weight factors, r Emotion , r InfoGain , r Logic Represent the scores of emotional impact, information gain and logic respectively.

6. The AI ​​intelligent interrogation teaching system according to claim 1 is characterized in that: The strategy reasoning module calculates the fitness of the questioning strategy using the following formula: Score(S,Z t )=GNN(V,E) Among them, S is the questioning strategy, Z t is the current emotional state vector, V and E are the embedded representations of the nodes and edges of the knowledge graph, respectively.

7. The AI ​​intelligent interrogation teaching system according to claim 1 is characterized in that: The optimization suggestion generation module includes: Generate optimization suggestions based on current questioning strategy and suspect status; Generate optimization suggestion text using natural language generation technology; Provides graphical feedback, including sentiment curves, question scores, and optimization directions.

8. The AI ​​intelligent interrogation teaching system according to claim 1 is characterized in that: The teaching evaluation module calculates the student ability score through the following formula: Among them, F i is the student's score on the i-th indicator, w i is the weight of the corresponding indicator.

9. The AI ​​intelligent interrogation teaching system according to claim 1 is characterized in that: The teaching evaluation module plans the personalized learning path through the following formula: Among them, L t is the learning path of the current time step, w t is the time step weight, R t is the profit value of the current time step.

10. An AI intelligent interrogation teaching method, according to the AI ​​intelligent interrogation teaching system according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1. Virtual scene generation: Generate a virtual interrogation scene based on the case background and the students’ questions, including the suspect’s language, facial expressions, intonation, and body movements; S2, multimodal data processing: capture the suspect's voice, expression, and text features, and generate the current emotional state vector through multimodal fusion; S3, strategy reasoning: calculate the best questioning strategy based on reinforcement learning and knowledge graph reasoning modules; S4. Real-time feedback generation: real-time evaluation of students’ questioning performance and generation of optimization suggestions; S5. Dynamic teaching evaluation: record students’ performance and dynamically adjust teaching paths to achieve personalized learning.

Images