A method and system for requirement analysis and dynamic interrogation based on cognitive reasoning

Abstract

The application discloses a demand analysis and dynamic follow-up question method and system based on cognitive reasoning, comprising the following steps: A1, collecting student multi-round question texts and preprocessing to obtain preprocessed word segmentation sequences; A2, sequentially calculating hidden state features, semantic correlation features and semantic coherence features, and extracting basic context state features; A3, calculating enhanced word features; constructing a cognitive evolution attention network, extracting semantic evolution path features, and combining the basic context state features to decode and generate a thinking track language description; A4, calculating semantic openness features; then constructing a dynamic follow-up question threshold adaptive mechanism to generate candidate follow-up questions; A5, integrating the candidate follow-up questions and the thinking track language description, and performing integrity checking and backup operations. The application can solve the problems of the traditional education AI question mode solidification, the lack of logical deduction of the multi-round thinking path of children, and the resulting poor feedback.

Description

Technical Field

[0001] This invention relates to the field of intelligent teaching technology, and in particular to a method and system for needs analysis and dynamic questioning based on cognitive reasoning. Background Technology

[0002] With the development of artificial intelligence technology, intelligent education systems are being increasingly widely used in personalized education for children. Among them, conversational learning assistance technology based on natural language processing, as an important means of realizing human-computer interactive teaching, is gradually being introduced into classroom teaching and self-directed learning scenarios. These systems interact with learners through text-based question-and-answer formats, aiming to simulate the teacher's guided questioning process and help children consolidate knowledge and develop thinking skills.

[0003] Currently, most existing text-based question-answering systems rely primarily on pre-set question templates or intent recognition mechanisms based on single-turn input, generating responses through rule-driven or fixed-path approaches. While these methods are effective with well-structured and clearly expressed input, they have significant limitations in real-world children's education scenarios. Children's language expression abilities are still developing, often exhibiting incomplete sentences, logical leaps, and vague word choice, making it difficult to accurately convey their cognitive state or specific confusion. Furthermore, existing systems lack the ability to dynamically model the evolution of learners' cognitive states during multi-turn dialogues, failing to effectively identify potential comprehension biases or knowledge gaps from fragmented and non-linear expressions.

[0004] Furthermore, current deep learning-based question-answering models mostly focus on single-round semantic matching or local intent classification tasks, emphasizing the accuracy and relevance of answers while neglecting the continuous tracking and guidance of learners' thought processes. These models typically use "question-answer pairs" as the basic processing unit, lacking modeling of the deep semantic connections within the dialogue context, making it difficult to achieve functions such as heuristic questioning, cognitive diagnosis, and reflective guidance. When children fail to actively raise clear questions, if the system only passively responds to their surface statements, the interaction is likely to become merely a formality, failing to deeply stimulate their thinking or expose potential cognitive impairments.

[0005] More importantly, because children often struggle to clearly express their confusion or formulate high-quality questions during the learning process, a system lacking the ability to proactively explore and dynamically adjust its questioning strategies will find it difficult to provide truly personalized tutoring. This vicious cycle of "the more vague the expression, the more off-target the feedback" severely restricts the adaptability and teaching effectiveness of intelligent question-answering systems in real educational environments.

[0006] Therefore, there is an urgent need for an intelligent question-and-answer method that can dynamically understand children's cognitive development path based on multi-turn text interaction and generate guiding and adaptive questions accordingly, so as to break through the bottlenecks of existing technologies in terms of understanding depth, interaction continuity and personalized support, and improve the practical value and teaching effect of intelligent education systems in children's learning scenarios. Summary of the Invention

[0007] In view of this, the present invention aims to provide a method and system for demand analysis and dynamic questioning based on cognitive reasoning, in order to solve the problem of poor feedback caused by the rigid questioning methods of traditional educational AI and the lack of logical deduction of children's multi-round thinking paths.

[0008] This invention also discloses a needs analysis and dynamic inquiry system based on cognitive reasoning, comprising:

[0009] A1: Collect the text of students' questions from multiple rounds and preprocess it to obtain the preprocessed word segmentation sequence;

[0010] A2: Based on the preprocessed word segmentation sequence, calculate the hidden state features, semantic association features, and semantic coherence features in sequence, and extract the basic context state features;

[0011] The calculation method of the semantic coherence feature includes: calculating the cognitive jump decay feature based on the semantic Euclidean distance using a multilayer perceptron, and then performing a weighted summation of the cognitive jump decay feature with the semantic association feature; the calculation method of the cognitive jump decay feature includes: performing a nonlinear transformation on the hidden state features and their difference between the current round and the previous round of dialogue using a multilayer perceptron based on the semantic Euclidean distance.

[0012] A3: Based on the preprocessed word segmentation sequence and semantic coherence features, calculate the enhanced word features, construct a cognitive evolution attention network, extract semantic evolution path features, and combine them with basic context state features to decode and generate a language description of thought trajectory.

[0013] The extraction method of the semantic evolution path features includes: based on the enhanced word features, the enhanced word features of each round are concatenated, multilayer perceptron and self-attention are calculated through the cognitive evolution attention network to obtain the enhanced word attention score features, and then combined with the cognitive jump decay features, and obtained through concatenation, multilayer perceptron and layer normalization processing.

[0014] A4: Based on the language description of the thought process and the preprocessed word segmentation sequence of the current round of questions, calculate the semantic openness feature; then construct a dynamic follow-up question threshold adaptive mechanism to generate candidate follow-up questions;

[0015] A5: Integrate candidate follow-up questions and language descriptions of thought processes, and perform integrity verification and backup operations.

[0016] Furthermore, step A1 also includes:

[0017] A11: Collect the text of students' multiple rounds of questions in the educational interactive interface to obtain the raw question data;

[0018] A12: Use regular expressions to remove garbled characters, punctuation marks, and whitespace characters from the original question data to obtain the cleaned question data;

[0019] A13: The cleaned question data is labeled with the dialogue turn, and Chinese word segmentation tools are used for word segmentation. Then, the word segmentation results are filtered for stop words to obtain the preprocessed word segmentation sequence.

[0020] Furthermore, step A2 also includes:

[0021] A21: Based on the preprocessed word segmentation sequence, the hidden state features are calculated through an embedding layer and a gated recurrent unit. The calculation formula is as follows:

[0022] ;

[0023] in, For the first Hidden state characteristics of turn-based dialogue. For dialogue round index 1, For gated loop unit, For embedding layer, For the first The preprocessed word segmentation sequence of a turn-based dialogue;

[0024] A22: Based on the hidden state features, semantic association features are calculated using a multi-head attention mechanism. The calculation formula is as follows:

[0025] ;

[0026] in, For the first Semantic relevance features of turn-based dialogue This is a multi-head attention mechanism;

[0027] A23: Based on the hidden state features, calculate the semantic Euclidean distance and cognitive jump decay features, and combine them with semantic association features to calculate the semantic coherence features. The calculation formula is as follows:

[0028] ;

[0029] ;

[0030] ;

[0031] in, For the first Wheel and the first Semantic Euclidean distance in turn-based dialogue For dialogue round index 2, , For the L2 norm operation, For the first Cognitive jump decay characteristics of turn-based dialogue It is a multilayer perceptron. For splicing operations, For semantic coherence features, The total number of rounds of dialogue. For the first Semantic relevance features of turn-based dialogue It is the dot product;

[0032] A24: Based on semantic coherence features and hidden state features, extract basic context state features. The calculation formula is as follows:

[0033] ;

[0034] in, Based on the basic context state features, The first Hidden state characteristics of turn-based dialogue.

[0035] It should be further explained that in multi-round question and answer sessions, children's expressions are characterized by strong leaps in thinking, large fluctuations in semantic coherence, and relatively vague expressions. This makes it difficult for traditional methods to accurately capture the dynamic evolution of children's thinking paths. In particular, when the number of dialogue rounds increases, existing technologies cannot effectively distinguish between key thinking nodes and redundant information, resulting in a lack of targeted follow-up questions.

[0036] This invention addresses the aforementioned challenges through deep coupling of semantic coherence features and hidden state features. Specifically, it first quantifies the semantic differences between adjacent dialogue rounds using semantic Euclidean distance. Then, it uses a multilayer perceptron to perform a nonlinear transformation on the hidden state features and their differences in each round of dialogue, generating cognitive jump decay features to dynamically reflect the intensity of children's thought jumps. Subsequently, it performs a dot product weighted summation of the cognitive jump decay features and semantic association features to capture the changes in the strength of semantic coherence between different rounds. Finally, it fuses the semantic coherence features with the hidden state features of all rounds through a concatenation operation, and extracts the basic contextual state features through a multilayer perceptron. Among these features, the cognitive jump decay features serve as a dynamic adjustment term, amplifying the contribution of rounds with strong thought coherence and suppressing noise interference from rounds with excessive jumps. The global concatenation of the hidden state features of each round preserves complete temporal information, ensuring that the basic contextual state features can reflect the current dialogue focus and inherit the historical thought process, providing a timely and continuous representational basis for subsequent reasoning about children's cognitive trajectories, thereby effectively supporting the accurate generation of dynamic follow-up questions.

[0037] Existing technologies typically employ simple concatenation of multi-turn dialogue features or rely solely on single-turn semantic analysis, failing to consider the dynamic decay characteristics of semantic coherence between turns. This leads to generated follow-up questions easily deviating from children's actual thought processes. The innovative advantage of this invention lies in: explicitly modeling the non-uniformity of thought jumps through cognitive jump decay features, enabling semantic coherence features to adaptively adjust the weights of each turn, avoiding path distortion caused by fixed weights in traditional methods; simultaneously, the deep fusion of semantic coherence features and hidden state features allows the basic context state features to possess both local semantic details and global evolutionary trends, significantly improving the ability to analyze children's ambiguous expressions and making follow-up questions more aligned with children's cognitive rhythm.

[0038] Because children's multi-turn conversations often involve sudden topic shifts or logical breaks, this invention utilizes a cognitive jump decay feature driven by semantic Euclidean distance to accurately identify such thought jumps and dynamically adjust feature weights, avoiding misjudging accidental jumps as valid paths. Simultaneously, the dot product fusion mechanism of semantic association features and cognitive jump decay features strengthens historical dependencies when thought flows smoothly and focuses on the current context when jumps are dramatic, better matching the characteristics of children's thinking. This ensures that subsequent follow-up questions both continue valid thought lines and respond promptly to cognitive shifts, solving the feedback bias problem caused by the rigid processing of multi-turn conversations in traditional educational AI, and demonstrating stronger adaptability to this scenario.

[0039] Furthermore, the calculation method of the enhanced word features includes: vectorizing the preprocessed word segmentation sequence through the word embedding layer to obtain the original word embedding features, and performing a dot product between the semantic coherence features and the sinusoidal position code after transformation by a multilayer perceptron, and then adding it to the original word embedding features to obtain the final feature.

[0040] Furthermore, step A3 also includes:

[0041] A31: Based on the preprocessed word segmentation sequence and semantic coherence features, calculate the enhanced word features using the following formula:

[0042]

[0043]

[0044] in, For the first Original word embedding features of turn-based dialogue For the first Enhanced word features in turn-based dialogue. Encoding the position of the sine wave;

[0045] A32: Based on enhanced word features and semantic Euclidean distance, a cognitive evolutionary attention network is constructed to extract semantic evolutionary path features. The calculation formula is as follows:

[0046]

[0047]

[0048]

[0049]

[0050] in, To enhance global word features, These are the enhanced word features for rounds 1, 2, ..., K of the dialogue. To enhance word attention score features, For self-attention mechanism, For the first Evolutionary attention score features in turn-based dialogues Features representing semantic evolution paths For layer normalization;

[0051] A33: Based on semantic evolution path features and basic context state features, decode and generate a language description of the thought trajectory. The calculation formula is as follows:

[0052]

[0053] in, A linguistic description of thought processes. This is for computation of lightweight pre-trained language models.

[0054] It should be further explained that, in response to the difficulties that children's expressions often present, such as frequent semantic jumps, incomplete logical chains, implicit connections between successive stages and non-linear evolution, this invention achieves accurate modeling of semantic evolution paths through cognitive evolutionary attention networks;

[0055] Specifically, firstly, the enhancement word features of each round of dialogue are concatenated and transformed by a multilayer perceptron to form global enhancement word features, preserving complete cross-round semantic information. Then, the attention score features of the enhancement words are calculated through a self-attention mechanism to initially establish word-level cross-round associations. Based on this, cognitive jump decay features are introduced and concatenated with the enhancement word attention score features, and then processed by a multilayer perceptron to generate evolutionary attention score features for each round. This allows attention distribution to be explicitly regulated by the degree of cognitive jump: when a round of dialogue differs significantly from the previous round semantically, the cognitive jump decay features automatically reduce their attention weight, reducing interference from irrelevant information; when semantic coherence is strong, their attention contribution is enhanced, strengthening the continuity of effective paths. Finally, the evolutionary attention score features of each round are concatenated with the global enhancement word features, and processed by a multilayer perceptron and layer normalization to obtain semantic evolution path features. This enables the network to dynamically identify and retain effective evolutionary cues in children's thinking, suppress noise interference, and thus generate a coherent and interpretable description of thought trajectories.

[0056] It should be further explained that existing technologies typically use standard self-attention mechanisms to directly process multi-turn dialogue sequences, without considering the non-uniformity of children's thought processes. This results in an overly even distribution of attention weights or interference from abnormal jumps, making it difficult to distinguish key thought nodes. The innovative advantage of this invention lies in the fact that by introducing cognitive jump decay features to reweight attention scores, dynamic adjustment of the attention mechanism is achieved, making the modeling process of semantic evolution paths cognitively reasonable. At the same time, the round-by-round independent calculation and fusion strategy of evolutionary attention score features not only preserves the evolutionary information at the round granularity but also avoids the problem of excessive smoothing of global attention, significantly improving the ability to characterize children's non-linear thinking paths.

[0057] By leveraging the synergistic effect of cognitive jump decay characteristics and self-attention mechanisms, this invention can identify children's thought jump points and automatically adjust their focus of attention. It continues the semantic chain when thinking is coherent and resets the focus of attention in a timely manner when a jump occurs, thereby more realistically restoring the discontinuous evolutionary process of children's thinking. This ensures that the language description of the subsequent thought trajectory not only conforms to children's expression habits but also has logical traceability, providing a high-quality cognitive reasoning foundation for subsequent dynamic questioning. It effectively overcomes the path distortion problem caused by ignoring cognitive jumps in traditional methods.

[0058] Furthermore, step A4 also includes:

[0059] A41: Based on the linguistic description of the thought trajectory and the preprocessed word segmentation sequence of the current round, extract the thought trajectory features and the current question features respectively, and then calculate the semantic openness features. The calculation formula is as follows:

[0060]

[0061]

[0062]

[0063] in, Characteristics of thought process trajectory For average pooling layer, A linguistic description of thought processes. Based on the current question characteristics, For the first The preprocessed word segmentation sequence of turn-based dialogue. For semantic openness features, For the Sigmoid function;

[0064] A42: Based on semantic openness features and current question features, a dynamic follow-up question threshold adaptive mechanism is constructed to generate candidate follow-up questions. The calculation formula is as follows:

[0065]

[0066]

[0067] in, To dynamically inquire about thresholds, Follow-up questions for candidates Kernel sampling function based on cumulative probability, For lightweight pre-trained language model computation, The number of candidate questions.

[0068] It should be further explained that the semantic openness feature described in this invention has significant technical effects in the dynamic question generation scenario of educational AI. The difficulty of this scenario lies in the highly dynamic nature of children's thinking paths, and the questioning needs to dynamically balance between "guiding exploration" and "focusing on in-depth exploration": when children's thinking is in the divergent stage, open-ended questions need to be generated to encourage expression; when thinking tends to be focused, concise questions need to be generated to promote accurate feedback. Traditional methods are difficult to quantify this degree of semantic openness, resulting in a single questioning mode that cannot adapt to the cognitive pace of children.

[0069] This invention achieves precise control over the openness of follow-up questions through semantic openness features. Specifically, firstly, the thought trajectory description is obtained through word embedding and average pooling to reflect the overall cognitive trend of children. Simultaneously, the preprocessed word segmentation sequence of the current round is used to extract the current question feature through a gated recurrent unit to capture the immediate focus of expression. Then, the two are concatenated and transformed by a multilayer perceptron to generate semantic openness features, comprehensively reflecting the breadth of historical thought paths and the convergence of the current question: when the thought trajectory covers multiple semantic directions and the current question does not explicitly point to a sub-topic, the semantic openness feature value is high, indicating the need for exploratory follow-up questions; conversely, when the thought process is focused and the current question has a clear direction, the semantic openness feature value is low. A low semantic openness feature value indicates the need for focused follow-up questions; this enables the system to possess "cognitive situational awareness" capabilities, providing quantifiable decision-making basis for subsequent dynamic follow-up questions; existing technologies typically use fixed templates or static rules to generate follow-up questions, lacking a dynamic evaluation mechanism for semantic openness, resulting in follow-up questions that are either too open and deviate from the topic, or too closed and suppress expression; the innovative advantage of this invention lies in: explicitly modeling the semantic openness of follow-up questions through semantic openness features, making the generation strategy adjustable; this feature is jointly driven by thought trajectory features and current question features, considering both the global cognitive trend and the local expression state, avoiding misjudgments caused by a single signal, and significantly improving the adaptability and guidance of follow-up questions;

[0070] On the other hand, children's language expression abilities exhibit individual differences and periodic fluctuations. Therefore, when generating follow-up questions, a dynamic balance needs to be struck between semantic diversity and linguistic conciseness. If the generation strategy is too open, it can easily produce complex and lengthy questions that exceed children's comprehension. If it is too conservative, the questions will be repetitive and monotonous, failing to stimulate in-depth thinking. Traditional generation methods use fixed thresholds for sampling, lacking the ability to respond to contextual cognitive states. This invention constructs a dynamic follow-up threshold adaptive mechanism to achieve fine-grained control of the question generation process. Specifically, when the semantic openness feature value is high, it indicates that the child's thinking is in the exploratory stage, and the dynamic follow-up threshold is increased accordingly to expand the sampling vocabulary space and generate more open and guiding questions (such as "What else did you think of?"). When the semantic openness feature value is low, it indicates that the child's expression tends to be focused, and the dynamic follow-up threshold is automatically reduced to limit the sampling range and generate concise and clear questions (such as "What does the 'it' you just mentioned refer to?"). This mechanism directly couples the semantic openness feature with the generation strategy, enabling the language style and semantic span of the follow-up questions to be adjusted in real time according to the child's cognitive state.

[0071] Furthermore, step A5 also includes:

[0072] A51: Collect candidate follow-up questions and thought process descriptions generated in the current round to obtain follow-up question generation records;

[0073] A52: The follow-up questions are formatted and packaged, and then transmitted to the database, where they are associated with and stored with the user profile information to obtain the user's question and answer records;

[0074] A53: Perform integrity verification and backup operations on user question and answer records.

[0075] This invention also discloses a needs analysis and dynamic inquiry system based on cognitive reasoning, comprising:

[0076] Question text acquisition module: Collects and preprocesses students' multiple rounds of question text to obtain the preprocessed word segmentation sequence;

[0077] Feature extraction module: Based on the preprocessed word segmentation sequence, calculate the hidden state features, semantic association features, and semantic coherence features in sequence, and extract the basic context state features;

[0078] The thought trajectory generation module calculates enhanced word features based on the preprocessed word segmentation sequence and semantic coherence features; it constructs a cognitive evolution attention network, extracts semantic evolution path features, and combines them with basic contextual state features to decode and generate a thought trajectory language description.

[0079] Dynamic follow-up question module: Based on the linguistic description of thought trajectory and the preprocessed word segmentation sequence of the current round, the semantic openness feature is calculated; then, a dynamic follow-up question threshold adaptive mechanism is constructed to generate candidate follow-up questions;

[0080] Storage module: Integrates candidate follow-up questions and language descriptions of thought processes, and performs integrity verification and backup operations.

[0081] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0082] (1) In response to the technical problems of poor feedback caused by the rigid questioning methods of traditional educational AI and the lack of logical deduction of children's multi-turn thinking paths, this invention introduces a weighted mechanism of cognitive jump decay feature and semantic coherence feature to accurately depict the thinking jump and semantic evolution law in children's dialogue. Combined with cognitive evolution attention network to dynamically model the enhanced word features, it effectively restores the cognitive trajectory of multi-turn dialogue and generates an interpretable thinking path description. Furthermore, this invention utilizes semantic openness feature and dynamic follow-up question threshold adaptive mechanism to enable follow-up questions to be dynamically adjusted between openness and simplicity according to the child's current cognitive state, avoiding the defects of single question pattern and poor adaptability in traditional methods, and significantly improving the interactive intelligence and personalization level of educational AI.

[0083] (2) This invention addresses the problem of strong leaps in thinking and large fluctuations in semantic coherence in children's multi-turn dialogues by proposing a semantic coherence modeling method based on cognitive leap decay features. The semantic differences between turns are quantified by semantic Euclidean distance, and cognitive leap decay features are generated by multilayer perceptron to dynamically reflect the degree of thinking leaps. Then, the features are weighted by dot product with semantic association features to adaptively adjust the contribution weight of each turn, effectively suppressing leap noise and preserving coherent paths. Finally, the semantic coherence features and the hidden state features of each turn are integrated to construct basic context state features that have both temporal continuity and local focus. Compared with the rigid processing of fixed splicing or single-turn analysis in traditional methods, this invention can accurately capture the evolutionary law of children's nonlinear thinking and significantly improve the ability to analyze fuzzy expressions and the adaptability of question generation.

[0084] (3) In response to the problem of distortion in thinking path modeling, this invention proposes a semantic evolution path modeling method based on cognitive evolutionary attention network to achieve dynamic regulation of attention distribution: suppressing noise interference when semantic jumps are severe and strengthening the continuation of effective paths when coherence is strong; finally, the evolutionary attention score features and global enhanced word features are integrated, and semantic evolution path features are obtained through multilayer perceptron and layer normalization; compared with the defects of traditional self-attention mechanism, which has average weight distribution and is easily affected by jump interference, this invention explicitly guides attention focus through cognitive jump decay features, accurately captures the nonlinear and discontinuous thinking evolution law of children, and significantly improves the coherence and interpretability of thinking trajectory modeling.

[0085] (4) This invention addresses the problem of difficulty in adapting questioning strategies to the dynamic evolution of children's thinking. It proposes a generation and control method based on semantic openness features and a dynamic questioning threshold adaptive mechanism. By integrating the characteristics of thinking trajectory and the current questioning features, semantic openness features are generated through a multilayer perceptron to quantify the openness of children's thinking and achieve dynamic perception of cognitive state: enhancing openness when thinking is divergent and enhancing conciseness when expression is focused. Furthermore, semantic openness features are introduced into the question generation process to construct a dynamic questioning threshold adaptive mechanism, which adjusts the sampling threshold in real time to achieve a dynamic balance between vocabulary diversity and language conciseness in questioning. Compared with the rigid questioning problem caused by relying on fixed templates or static parameters in traditional methods, this invention achieves adaptive switching of questioning modes through semantic openness features and generation strategies. This avoids deviating from the topic and prevents suppression of expression, significantly improving the personalization and intelligence of question generation, and better matching the pace of children's cognitive development and differences in expression ability. Attached Figure Description

[0086] Figure 1 A flowchart illustrating a cognitive reasoning-based demand analysis and dynamic questioning method provided by this invention;

[0087] Figure 2The interface diagram of the cognitive reasoning-based requirements analysis system provided by the present invention is shown. Detailed Implementation

[0088] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0089] Example 1: A method for needs analysis and dynamic probing based on cognitive reasoning, such as... Figure 1 As shown, it includes the following steps:

[0090] A1: Collect the text of students' questions from multiple rounds and preprocess it to obtain the preprocessed word segmentation sequence;

[0091] include:

[0092] A11: Collect the text of students' multiple rounds of questions in the educational interactive interface to obtain the raw question data;

[0093] A12: Use regular expressions to remove garbled characters, punctuation marks, and whitespace characters from the original question data to obtain the cleaned question data;

[0094] A13: The cleaned question data is labeled with the dialogue turn, and Chinese word segmentation tools are used for word segmentation. Then, the word segmentation results are filtered for stop words to obtain the preprocessed word segmentation sequence.

[0095] A2: Based on the preprocessed word segmentation sequence, calculate the hidden state features, semantic association features, and semantic coherence features in sequence, and extract the basic contextual state features; including:

[0096] A21: Based on the preprocessed word segmentation sequence, the hidden state features are calculated through an embedding layer and a gated recurrent unit. The calculation formula is as follows:

[0097] ;

[0098] in, For the first Hidden state characteristics of turn-based dialogue. For dialogue round index 1, For gated loop unit, For embedding layer, For the first The preprocessed word segmentation sequence of a turn-based dialogue;

[0099] A22: Based on the hidden state features, semantic association features are calculated using a multi-head attention mechanism. The calculation formula is as follows:

[0100] ;

[0101] in, For the first Semantic relevance features of turn-based dialogue This is a multi-head attention mechanism;

[0102] A23: Based on the hidden state features, calculate the semantic Euclidean distance and cognitive jump decay features, and combine them with semantic association features to calculate the semantic coherence features. The calculation formula is as follows:

[0103] ;

[0104] ;

[0105] ;

[0106] in, For the first Wheel and the first Semantic Euclidean distance in turn-based dialogue For dialogue round index 2, , For the L2 norm operation, For the first Cognitive jump decay characteristics of turn-based dialogue It is a multilayer perceptron. For splicing operations, For semantic coherence features, The total number of rounds of dialogue. For the first Semantic relevance features of turn-based dialogue It is the dot product;

[0107] A24: Based on semantic coherence features and hidden state features, extract basic context state features. The calculation formula is as follows:

[0108] ;

[0109] in, Based on the basic context state features, The first Hidden state characteristics of turn-based dialogue.

[0110] A3: Based on the preprocessed word segmentation sequence and semantic coherence features, calculate enhanced word features, construct a cognitive evolution attention network, extract semantic evolution path features, and combine them with basic contextual state features to decode and generate a linguistic description of thought trajectory; including:

[0111] A31: Based on the preprocessed word segmentation sequence and semantic coherence features, calculate the enhanced word features using the following formula:

[0112]

[0113]

[0114] in, For the first Original word embedding features of turn-based dialogue For the first Enhanced word features in turn-based dialogue. Encoding the position of the sine wave;

[0115] A32: Based on enhanced word features and semantic Euclidean distance, a cognitive evolutionary attention network is constructed to extract semantic evolutionary path features. The calculation formula is as follows:

[0116]

[0117]

[0118]

[0119]

[0120] in, To enhance global word features, These are the enhanced word features for rounds 1, 2, ..., K of the dialogue. To enhance word attention score features, For self-attention mechanism, For the first Evolutionary attention score features in turn-based dialogues Features representing semantic evolution paths For layer normalization;

[0121] A33: Based on semantic evolution path features and basic context state features, decode and generate a language description of the thought trajectory. The calculation formula is as follows:

[0122]

[0123] in, A linguistic description of thought processes. This is for computation of lightweight pre-trained language models.

[0124] Specifically, addressing the challenges of adapting to different cognitive stages in multi-turn dialogues due to changes in children's thought processes, this invention also provides a decoding method based on cognitive gating fusion to replace the original A33 step. The calculation formula is as follows:

[0125]

[0126]

[0127] in, To recognize gating features, This is the Sigmoid function.

[0128] A4: Based on the linguistic description of the thought process and the preprocessed word segmentation sequence of the current round of questions, calculate the semantic openness feature; then construct a dynamic follow-up question threshold adaptive mechanism to generate candidate follow-up questions; including:

[0129] A41: Based on the linguistic description of the thought trajectory and the preprocessed word segmentation sequence of the current round, extract the thought trajectory features and the current question features respectively, and then calculate the semantic openness features. The calculation formula is as follows:

[0130]

[0131]

[0132]

[0133] in, Characteristics of thought process trajectory For average pooling layer, A linguistic description of thought processes. Based on the current question characteristics, For the first The preprocessed word segmentation sequence of turn-based dialogue. For semantic openness features, For the Sigmoid function;

[0134] A42: Based on semantic openness features and current question features, a dynamic follow-up question threshold adaptive mechanism is constructed to generate candidate follow-up questions. The calculation formula is as follows:

[0135]

[0136]

[0137] in, To dynamically inquire about thresholds, Follow-up questions for candidates Kernel sampling function based on cumulative probability, For lightweight pre-trained language model computation, The number of candidate questions.

[0138] For example, in the field of science classes, the system collected three rounds of dialogues between 7-year-old Xiaoming and his child about plant growth:

[0139] Xiaoming asked, "How do grasses grow from the soil?" This expresses his curiosity about the initial process of plant growth, but he did not clearly distinguish between the seed germination and growth stages.

[0140] Xiaoming suddenly turned to "My little flower wilted yesterday, did it not drink water?", showing a leap in thinking and connecting the classroom question with his personal family experience, but he confused the concepts of "grass" and "flower", showing the typical generalized thinking characteristics of children.

[0141] After being guided by the teacher, Xiaoming further considered, "Is the sun also important? I put it on the windowsill." He began to consider the influence of multiple factors, but there was still ambiguity in the reference ("it" was unclear), which showed the imprecision of children's expression.

[0142] The system first preprocesses the three-round dialogue, extracts hidden state features, and calculates the semantic Euclidean distance; analysis reveals that the semantic Euclidean distance between the second and first rounds is relatively large. =0.85), indicating a significant leap in thinking; while the distance between the third round and the second round is relatively small ( =0.32, indicating that thinking is beginning to focus; based on this, the system calculates the cognitive jump decay characteristic and assigns a lower weight to the second round of dialogue ( =0.28), to avoid its excessive influence on the overall path judgment;

[0143] Subsequently, the system constructed a cognitive evolutionary attention network to process the enhanced word features of the three rounds of dialogue. The analysis showed that "water" and "sun" received higher attention weights in the third round of dialogue (0.72 and 0.68, respectively), while the attention of the word "withered" was suppressed to 0.15 by the cognitive jump decay feature, effectively filtering out the interference of emotional expression.

[0144] Ultimately, the system-generated thought process language description is as follows: "Starting from the basic process of plant growth, the child goes through a leap in thinking that connects classroom knowledge with personal experience, and gradually develops into a cognitive mode that considers the influence of multiple environmental factors. This shows a progressive development from single-factor to multi-factor thinking, but there are characteristics of conceptual generalization and ambiguous reference."

[0145] Based on this thought process, the system calculates a semantic openness feature value of 0.63 (between 0.5 and 0.8, indicating that the child's thinking is in a balance between divergence and focus; the dynamic follow-up questioning threshold adaptive mechanism is set accordingly. =0.75, generating three candidate follow-up questions:

[0146] 1. "What role do you think soil plays in the growth of small flowers?"

[0147] 2. "Besides water and sunlight, what else does a small flower need to grow healthily?"

[0148] 3. "Can you describe the process of a flower growing from seed to bloom?"

[0149] The system ultimately selected question 2 as the follow-up question because it maintained a suitable degree of openness to promote exploration while focusing on the core concepts of the current discussion, which is suitable for children's current cognitive development stage. In practical application, Xiaoming showed great interest in this follow-up question and further answered: "Maybe we also need air, just like we need to eat!", which verified the effectiveness of the dynamic follow-up questioning strategy of the present invention.

[0150] A5: Integrate candidate follow-up questions and verbal descriptions of thought processes, and perform integrity checks and backup operations, including:

[0151] A51: Collect candidate follow-up questions and thought process descriptions generated in the current round to obtain follow-up question generation records;

[0152] A52: The follow-up questions are formatted and packaged, and then transmitted to the database, where they are associated with and stored with the user profile information to obtain the user's question and answer records;

[0153] A53: Perform integrity verification and backup operations on user question and answer records.

[0154] Example 2: This invention also discloses a needs analysis and dynamic questioning system based on cognitive reasoning, such as... Figure 2 As shown, it includes:

[0155] Question text acquisition module: Collects and preprocesses students' multiple rounds of question text to obtain the preprocessed word segmentation sequence;

[0156] Feature extraction module: Based on the preprocessed word segmentation sequence, calculate the hidden state features, semantic association features, and semantic coherence features in sequence, and extract the basic context state features;

[0157] The thought trajectory generation module calculates enhanced word features based on the preprocessed word segmentation sequence and semantic coherence features; it constructs a cognitive evolution attention network, extracts semantic evolution path features, and combines them with basic contextual state features to decode and generate a thought trajectory language description.

[0158] Dynamic follow-up question module: Based on the linguistic description of thought trajectory and the preprocessed word segmentation sequence of the current round, the semantic openness feature is calculated; then, a dynamic follow-up question threshold adaptive mechanism is constructed to generate candidate follow-up questions;

[0159] Storage module: Integrates candidate follow-up questions and language descriptions of thought processes, and performs integrity verification and backup operations;

[0160] It should be noted that the calculation process of hidden state features, semantic association features, and semantic coherence features involved in the feature extraction module belongs to the internal processing of the backend model and is not an independent physical module. Its function is implicitly completed by the thought trajectory generation module calling the deep learning network, so it is not displayed in the interface.

[0161] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0162] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0163] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for requirement analysis and dynamic interrogation based on cognitive reasoning, characterized in that, Includes the following steps: A1: Collect the text of students' questions from multiple rounds and preprocess it to obtain the preprocessed word segmentation sequence; A2: Based on the preprocessed word segmentation sequence, calculate the hidden state features, semantic association features, and semantic coherence features in sequence, and extract the basic context state features; The calculation method of the semantic coherence feature includes: calculating the cognitive jump decay feature based on the semantic Euclidean distance using a multilayer perceptron, and then performing a weighted summation of the cognitive jump decay feature with the semantic association feature; the calculation method of the cognitive jump decay feature includes: performing a nonlinear transformation on the hidden state features and their difference between the current round and the previous round of dialogue using a multilayer perceptron based on the semantic Euclidean distance. A3: Based on the preprocessed word segmentation sequence and semantic coherence features, calculate the enhanced word features, construct a cognitive evolution attention network, extract semantic evolution path features, and combine them with basic context state features to decode and generate a language description of thought trajectory. The extraction method of the semantic evolution path features includes: based on the enhanced word features, the enhanced word features of each round are concatenated, multilayer perceptron and self-attention are calculated through the cognitive evolution attention network to obtain the enhanced word attention score features, and then combined with the cognitive jump decay features, and obtained through concatenation, multilayer perceptron and layer normalization processing. A4: Based on the language description of the thought process and the preprocessed word segmentation sequence of the current round of questions, calculate the semantic openness feature; then construct a dynamic follow-up question threshold adaptive mechanism to generate candidate follow-up questions; A5: Integrate candidate follow-up questions and language descriptions of thought processes, and perform integrity verification and backup operations.

2. The method of claim 1, wherein the cognitive reasoning-based requirement analysis and dynamic follow-up question method is characterized by, Step A1 includes: A11: Collect the text of students' multiple rounds of questions in the educational interactive interface to obtain the raw question data; A12: Use regular expressions to remove garbled characters, punctuation marks, and whitespace characters from the original question data to obtain the cleaned question data; A13: The cleaned question data is labeled with the dialogue turn, and Chinese word segmentation tools are used for word segmentation. Then, the word segmentation results are filtered for stop words to obtain the preprocessed word segmentation sequence.

3. The method of claim 1, wherein the cognitive reasoning-based requirement analysis and dynamic follow-up question method is characterized by, Step A2 includes: A21: Based on the preprocessed word segmentation sequence, the hidden state features are calculated through an embedding layer and a gated recurrent unit. The calculation formula is as follows: ； wherein, is the hidden state feature of the dialog turn index 1, is the embedding layer, is the preprocessed token sequence of the A22: Based on the hidden state features, semantic association features are calculated using a multi-head attention mechanism. The calculation formula is as follows: ； wherein, is a first semantic association feature for the turn, is a multi-head attention mechanism; A23: Based on the hidden state features, calculate the semantic Euclidean distance and cognitive jump decay features, and combine them with semantic association features to calculate the semantic coherence features. The calculation formula is as follows: ； ； ； in, For the first Wheel and the first Semantic Euclidean distance in turn-based dialogue For dialogue round index 2, , For the L2 norm operation, For the first Cognitive jump decay characteristics of turn-based dialogue It is a multilayer perceptron. For splicing operations, For semantic coherence features, The total number of rounds of dialogue. For the first Semantic relevance features of turn-based dialogue It is the dot product; A24: Based on semantic coherence features and hidden state features, extract basic context state features. The calculation formula is as follows: ； wherein, are the base contextual state features, are the first hidden state features of the previous dialog.

4. The method of claim 3, wherein the cognitive reasoning-based requirement analysis and dynamic follow-up question method is characterized by, The calculation method of the enhanced word features includes: vectorizing the preprocessed word segmentation sequence through the word embedding layer to obtain the original word embedding features, and then performing a dot product between the semantic coherence features and the sinusoidal position code after transformation by a multilayer perceptron, and finally adding it to the original word embedding features.

5. The method of claim 4, wherein the cognitive reasoning-based requirement analysis and dynamic follow-up question method is characterized by, Step A3 includes: A31: Based on the preprocessed word segmentation sequence and semantic coherence features, calculate the enhanced word features using the following formula: wherein, is the first round of dialogue's original word embedding features, is the first round of dialogue's enhanced word features, is the sine position encoding; A32: Based on enhanced word features and semantic Euclidean distance, a cognitive evolutionary attention network is constructed, and semantic evolutionary path features are extracted. The calculation formula is as follows: in, To enhance global word features, These are the enhanced word features for rounds 1, 2, ..., K of the dialogue. To enhance word attention score features, For self-attention mechanism, For the first Evolutionary attention score features in turn-based dialogues Features representing semantic evolution paths For layer normalization; A33: Based on semantic evolution path features and basic context state features, decode and generate a language description of the thought trajectory. The calculation formula is as follows: in, A linguistic description of thought processes. This is for computation of lightweight pre-trained language models.

6. The method for needs analysis and dynamic questioning based on cognitive reasoning according to claim 3, characterized in that, The A4 step includes: A41: Based on the linguistic description of the thought trajectory and the preprocessed word segmentation sequence of the current round, extract the thought trajectory features and the current question features respectively, and then calculate the semantic openness features. The calculation formula is as follows: in, Characteristics of thought process trajectory For average pooling layer, A linguistic description of thought processes. Based on the current question characteristics, For the first The preprocessed word segmentation sequence of turn-based dialogue. For semantic openness features, For the Sigmoid function; A42: Based on semantic openness features and current question features, a dynamic follow-up question threshold adaptive mechanism is constructed to generate candidate follow-up questions. The calculation formula is as follows: in, To dynamically inquire about thresholds, Follow-up questions for candidates Kernel sampling function based on cumulative probability, For lightweight pre-trained language model computation, The number of candidate questions.

7. The cognitive reasoning-based needs analysis and dynamic questioning method according to claim 6, characterized in that, Step A5 includes: A51: Collect candidate follow-up questions and thought process descriptions generated in the current round to obtain follow-up question generation records; A52: The follow-up questions are formatted and packaged, and then transmitted to the database, where they are associated with and stored with the user profile information to obtain the user's question and answer records; A53: Perform integrity verification and backup operations on user question and answer records.

8. A needs analysis and dynamic inquiry system based on cognitive reasoning, characterized in that, include: Question text acquisition module: Collects and preprocesses multiple rounds of student question texts to obtain a preprocessed word segmentation sequence; Feature extraction module: Based on the preprocessed word segmentation sequence, calculate the hidden state features, semantic association features, and semantic coherence features in sequence, and extract the basic context state features; The thought trajectory generation module calculates enhanced word features based on the preprocessed word segmentation sequence and semantic coherence features; We construct a cognitive evolution attention network, extract semantic evolution path features, and combine them with basic contextual state features to decode and generate a linguistic description of thought trajectories. Dynamic follow-up questioning module: Calculates semantic openness features based on the linguistic description of thought processes and the preprocessed word segmentation sequence of the current round; Then, a dynamic follow-up threshold adaptive mechanism is constructed to generate candidate follow-up questions; Storage module: Integrates candidate follow-up questions and language descriptions of thought processes, and performs integrity verification and backup operations; To achieve a cognitive reasoning-based demand analysis and dynamic questioning method as described in any one of claims 1-7.

Images