Learning method and device based on two-channel neural network attention mechanism and storage medium

Abstract

The invention discloses a learning method and device based on a two-channel neural network attention mechanism and a storage medium, and the method comprises the steps: constructing logic verification and heuristic guidance training data, constructing a learning guidance type two-channel path language processing model which is located as a repeated understanding learning listener and a guider, and carrying out the learning guidance type two-channel path language processing model. And receiving the retelling content of the learner, analyzing and judging the logical coherence state and the knowledge missing condition of the learner, and generating a heuristic questioning guide completion logic chain. According to the method, the core pain point of lack of high-quality listeners in retelling type understanding learning is effectively solved, natural interaction and rigorous logic review are considered, logic vulnerabilities and knowledge blind areas in the retelling content of the learner can be accurately identified, the learner is effectively guided to deepen knowledge understanding, and the learning efficiency is improved. And a stable and reliable technical support is provided for efficient development of repeated understanding learning.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a learning method, device, and storage medium based on a dual-channel neural network attention mechanism. Background Technology

[0002] In the field of knowledge acquisition, paraphrasing-based learning methods, exemplified by the Feynman Learning Technique, are widely recognized as highly effective because their core principle is "discovering one's own knowledge gaps by clearly paraphrasing concepts to others." The key to this method is that learners, while paraphrasing concepts, need external feedback to identify breaks in their logical chains and knowledge gaps, thereby deepening their understanding and internalizing the knowledge. However, in practice, learners often struggle to find readily available, high-quality "listeners"—ideal listeners need keen logical scrutiny skills to accurately identify logical gaps in the paraphrased content and guide learners to fill in knowledge gaps through heuristic questioning, rather than simply echoing or directly imparting answers.

[0003] With the development of artificial intelligence technology, large language models (LLMs) are gradually being used to play the aforementioned "listener" role. However, existing technologies still have significant technical bottlenecks, making it difficult to meet the core requirements of paraphrasing-based comprehension learning.

[0004] First, existing large-scale language models lack the ability to logically examine the completeness of explanations. Current mainstream large models (such as GPT-5 and Gemini 3.0 Pro) are essentially trained as knowledge completers, using a vast knowledge base to smoothly continue and supplement input content. When learners omit key logical points or make leaps in reasoning when restating concepts, these models automatically fill in the missing logical links and acknowledge the learner's statement. This prevents learners from recognizing their own knowledge gaps, directly weakening the core effect of paraphrasing-based comprehension learning.

[0005] Secondly, existing language modeling paradigms have inherent limitations, leading to the "fluency bias" problem. The core of autoregressive language model training is maximizing the probability of text sequences, excelling at capturing high-probability token co-occurrence patterns. This makes the model naturally inclined to consider texts that are "fluent and conform to common language habits" to be more credible, even if the text has obvious logical flaws. This "emphasis on fluency over logic" characteristic means that elaborate but seemingly plausible paraphrasing can easily mislead the model, preventing it from effectively fulfilling its logical review function.

[0006] Third, in existing technologies, some solutions attempt to make the model act as a "Feynman Learning listener" through prompt engineering. However, this approach cannot enable the model to internalize logical review capabilities as its own inherent ability. In long-context interaction scenarios, the model is easily disturbed by irrelevant information, gradually deviating from its preset role, losing its ability to perceive logical breakpoints, and making it difficult to continuously provide stable and effective feedback to the learner.

[0007] In summary, existing technologies cannot provide a "high-quality listener" for paraphrasing-based comprehension learning methods, and therefore it is necessary to improve them. Summary of the Invention

[0008] In view of the above problems, this application provides a learning method based on a dual-channel neural network attention mechanism to solve the technical problem that the aforementioned artificial intelligence technology cannot provide an auxiliary tool for paraphrasing comprehension learning that combines natural interactivity with rigorous logical review capabilities.

[0009] To achieve the above objectives, this application provides a learning method based on a dual-channel neural network attention mechanism, comprising:

[0010] Pre-construct logical verification and heuristic-guided training data;

[0011] A dual-channel language processing learning guidance model is constructed. The learning guidance model is configured as a listener for paraphrasing comprehension learning and adopts a dual-channel parallel structure of semantic listening and logical auditing. The semantic listening channel is used to capture the semantic fluency and referential relationships of the text context, and the logical auditing channel is used to focus on the core logical skeleton of the text. The learning guidance model includes, but is not limited to, Transformer-like models and other approximate language processing models based on multi-head attention mechanisms.

[0012] The learning guidance model receives the learner's concept restatement, and analyzes and judges the logical coherence and knowledge gaps of the content through semantic listening and logical auditing, generating corresponding heuristic follow-up questions to guide the learner to complete the logical chain.

[0013] Furthermore, the construction of the dual-channel language processing learning guidance model includes: physically grouping the semantic listening channel and the logical auditing channel, wherein the physical grouping includes:

[0014] The multi-head attention heads of the learning guidance model are physically divided into semantic listening channels and logical auditing channels according to a certain ratio;

[0015] The semantic listening channel and the logical auditing channel are capable of parallel computation, which includes:

[0016] The semantic listening channel receives a standard causal mask and outputs semantic features according to the self-attention calculation logic; wherein the standard causal mask is an upper triangular mask matrix, which is used to constrain the semantic listening channel to capture contextual semantics according to the text temporal order, so as to ensure the temporal consistency of semantic understanding.

[0017] The logic audit channel injects a logic mask matrix and outputs logic features through self-attention calculation with logical constraints. These logic features are used to capture the logical dependencies between entities in the text and encode the logical association state of the text. When the learner's concept restatement content has a broken logical chain or loopholes, the logic features present an abnormal signal, providing a core judgment basis for subsequent logical integrity verification and knowledge blind spot location.

[0018] Furthermore, after physically grouping the semantic listening channel and the logical auditing channel, the method also includes a feature splicing and end-to-end task-driven fine-tuning step: splicing the output features of the semantic listening channel and the output features of the logical auditing channel, and mapping them back to the original feature dimension through a linear projection layer;

[0019] By using training data from the critical thinking chain, the concatenated learning guidance model is fine-tuned end-to-end, enabling the learning guidance model to adaptively coordinate the dual-channel feature weights during backpropagation, thus achieving dynamic collaboration between logical judgment and semantic interaction.

[0020] Furthermore, before combining semantic listening and logical auditing to analyze and judge the logical coherence and knowledge gaps of the content, the method further includes the step of: training the learning guidance model in a targeted manner using the training data, so that the learning guidance model has the function of verifying the logical integrity of the concept restatement content output by the learner and locating knowledge blind spots.

[0021] Furthermore, the construction of the logical verification and heuristic-guided training data adopts a hybrid graph-based strategy, specifically including:

[0022] Obtain basic logical connections from the aforementioned open-source common sense graph;

[0023] Semantic segmentation, structured extraction, and natural language reasoning verification are performed on private domain documents to construct a private domain graph from unstructured private domain documents;

[0024] The logical paths in the private domain graph are perturbed programmatically using graph perturbation operators to generate erroneous input samples containing logical loopholes; the erroneous input samples are then reverse-processed using a higher-order teacher model to generate training data containing a critical thinking chain that includes logical verification processes and heuristic probing questions.

[0025] Furthermore, the construction of the logical verification and heuristic-guided training data also includes a closed-loop quality verification step based on graph path reconstruction:

[0026] The learner's response to the heuristic probing questions was simulated using a low-order reasoning model;

[0027] Generative grading of simulated responses is performed using a high-order model to determine whether the responses hit the target logical nodes in the graph. Only when the simulated responses are valid, the grading determines that the target node has been hit, and the heuristic follow-up questions do not reveal the name of the target node, is the training data of the corresponding critical thinking chain included in the final training set.

[0028] Furthermore, before injecting the logical audit channel into the logical mask matrix, the logical mask matrix is ​​constructed. The steps include: parsing the dependency tree structure of the input text using a dependency parser; defining the core logical relation set and the modification relation set; analyzing the dependency paths of any word pair in the text, retaining only word pair connections whose paths are all core logical relations, and masking word pair connections containing modification relations, thereby generating a topological logical mask matrix.

[0029] Furthermore, it also includes a step of asking follow-up questions to complete the logic: after receiving the learner's concept restatement text, the learning guidance model extracts the logical skeleton of the text in real time and generates a corresponding logical mask; the text is encoded in parallel through the dual-channel path, and if a logical feature mismatch signal is detected, the compliant response is suppressed and the heuristic follow-up question generation module is activated to output follow-up questions to guide the learner to complete the logic.

[0030] To address the aforementioned technical problems, this application also provides another technical solution:

[0031] A learning device based on a dual-channel neural network attention mechanism, comprising:

[0032] The storage module is used to store logical verification and heuristic training data;

[0033] A dual-channel language processing learning guidance model is configured as a listener for paraphrasing comprehension learning and adopts a dual-channel parallel structure of semantic listening and logical auditing. The semantic listening channel is used to capture the semantic fluency and referential relationships of the text context, while the logical auditing channel is used to focus on the core logical skeleton of the text. The learning guidance model includes, but is not limited to, Transformer-like models and other approximate language processing models based on multi-head attention mechanisms.

[0034] The application module is used by the learning guidance model to receive the learner's concept restatement content, and to analyze and judge the logical coherence and knowledge gaps of the content through semantic listening and logical auditing, and generate corresponding heuristic follow-up questions to guide the learner to complete the logical chain.

[0035] To address the aforementioned technical problems, this application also provides another technical solution:

[0036] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the learning method based on a dual-channel neural network attention mechanism as described in any of the above technical solutions.

[0037] Unlike existing technologies, the above technical solution pre-constructs training data for logical verification and heuristic guidance, building a dual-channel language processing model positioned as both a listener and guide in paraphrasing comprehension learning. This model employs a parallel structure of semantic listening and logical auditing. The semantic listening channel captures the semantic fluency and referential relationships of the text context, while the logical auditing channel focuses on the core logical framework of the text. After training, it possesses the ability to verify the logical integrity of learners' conceptual paraphrasing and locate knowledge gaps. During the learning process, it receives learners' paraphrasing content, analyzes and judges its logical coherence and knowledge gaps, and generates heuristic follow-up questions to guide the completion of the logical chain. The model includes, but is not limited to, Transformer-like models and other approximate language processing models based on multi-head attention mechanisms. This technical solution effectively solves the core pain point of lacking high-quality listeners in paraphrasing comprehension learning, achieving a balance between natural interaction and rigorous logical review. It can accurately identify logical loopholes and knowledge gaps in learners' paraphrasing content, effectively guiding learners to deepen their knowledge understanding, and providing stable and reliable technical support for efficient paraphrasing comprehension learning.

[0038] The above description of the invention is merely an overview of the technical solution of this application. In order to enable those skilled in the art to better understand the technical solution of this application and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of this application easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of this application. Attached Figure Description

[0039] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on this application.

[0040] In the accompanying drawings of the instruction manual:

[0041] Figure 1 A flowchart illustrating the learning method based on a dual-channel neural network attention mechanism as described in a specific implementation;

[0042] Figure 2 A flowchart illustrating the strategy for constructing training data based on hybrid spectra in a specific implementation method;

[0043] Figure 3 This is a flowchart illustrating the construction process of the private domain graph described in a specific implementation method;

[0044] Figure 4 The block diagram of the learning device based on the dual-channel neural network attention mechanism described in the specific implementation embodiment;

[0045] Figure 5 A schematic diagram of the computer-readable storage medium described in a specific embodiment;

[0046] The reference numerals used in the above figures are explained as follows:

[0047] 400. A learning device based on a dual-channel neural network attention mechanism; 401. A storage module; 402. A learning guidance model; 403. An application module;

[0048] 500. Computer-readable storage medium; Detailed Implementation

[0049] To illustrate the possible application scenarios, technical principles, implementable specific solutions, and achievable objectives and effects of this application in detail, the following description, in conjunction with the listed specific embodiments and accompanying drawings, provides a detailed explanation. The embodiments described herein are merely illustrative of the technical solutions of this application and are therefore intended to limit the scope of protection of this application.

[0050] In this document, the term "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The term "embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment, nor does it specifically limit its independence or connection with other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the technical features mentioned in each embodiment can be combined in any way to form corresponding implementable technical solutions.

[0051] Unless otherwise defined, the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the use of related terms herein is merely for the purpose of describing particular embodiments and is not intended to limit this application.

[0052] In the description of this application, the term "and / or" is used to describe the logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and A and B exist simultaneously. Additionally, the character " / " in this document generally indicates that the preceding and following objects have an "or" logical relationship.

[0053] In this application, terms such as “first” and “second” are used only to distinguish one entity or operation from another, and do not necessarily require or imply any actual quantity, hierarchy or order relationship between these entities or operations.

[0054] Without further limitations, the use of terms such as “comprising,” “including,” “having,” or other similar open-ended expressions in this application is intended to cover non-exclusive inclusion, which does not exclude the presence of additional elements in a process, method, or product that includes the stated elements, such that a process, method, or product that includes a list of elements may include not only those defined elements but also other elements not expressly listed, or elements inherent to such a process, method, or product.

[0055] As understood in the Examination Guidelines, in this application, expressions such as "greater than," "less than," and "exceeding" are understood to exclude the stated number; expressions such as "above," "below," and "within" are understood to include the stated number. Furthermore, in the description of the embodiments in this application, "multiple" means two or more (including two), and similar expressions related to "multiple" are also understood in this way, such as "multiple groups" and "multiple times," unless otherwise explicitly specified.

[0056] In the description of the embodiments of this application, the space-related expressions used, such as "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "vertical," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," indicate the orientation or positional relationship based on the orientation or positional relationship shown in the specific embodiments or drawings. They are only for the purpose of describing the specific embodiments of this application or for the reader's understanding, and do not indicate or imply that the device or component referred to must have a specific position, a specific orientation, or be constructed or operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of this application.

[0057] Unless otherwise expressly specified or limited, the terms "installation," "connection," "linking," "fixing," and "setting," as used in the description of the embodiments of this application, should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral setting; it can be a mechanical connection, an electrical connection, or a communication connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two components or the interaction between two components. For those skilled in the art to which this application pertains, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0058] Please see Figure 1 This embodiment provides a learning method based on a dual-channel neural network attention mechanism. The core of paraphrasing-based learning lies in discovering one's own knowledge gaps by clearly retelling learned knowledge to others; the Feynman Learning Technique is a typical example. This embodiment's learning method based on a dual-channel neural network attention mechanism can be widely applied to various scenarios of paraphrasing-based learning methods such as the Feynman Learning Technique. Specifically, it can include knowledge point retelling and consolidation during student self-study, concept decomposition and explanation in workplace skills training, knowledge deepening and review for professionals in specialized fields (such as medicine and engineering), and logical organization in interdisciplinary learning. This learning method based on a dual-channel neural network attention mechanism is suitable for general knowledge learning and can also meet the learning assistance needs of vertical fields through customized private domain graphs, providing intelligent guidance for learners who lack high-quality listeners.

[0059] like Figure 1 As shown, the learning method based on the dual-channel neural network attention mechanism includes the following steps:

[0060] S101. Pre-construct logical verification and heuristic training data;

[0061] S102. Construct a dual-channel language processing learning guidance model. The learning guidance model is configured as a listener for paraphrasing comprehension learning and adopts a dual-channel parallel structure of semantic listening and logical auditing. The semantic listening channel is used to capture the semantic fluency and referential relationships of the text context, and the logical auditing channel is used to focus on the core logical skeleton of the text. The learning guidance model includes, but is not limited to, Transformer-like models and other approximate language processing models based on multi-head attention mechanisms.

[0062] S103. The learning guidance model receives the learner's concept restatement content, and analyzes and judges the logical coherence and knowledge gaps of the content through semantic listening and logical auditing, and generates corresponding heuristic follow-up questions to guide the learner to complete the logical chain.

[0063] In step S102, the existing self-attention model is modified by adding a logical auditing channel to the existing semantic listening channel, constructing a "semantic listening-logical auditing dual-channel" feature extraction model. This model is defined as the listener role in paraphrasing comprehension learning, thus obtaining the learning guidance model. This learning guidance model, while retaining the model's natural language understanding ability, forces the model to focus on the logical skeleton of the input text through physical division of labor and masking mechanisms, thereby achieving low-level perception of logical breakpoints.

[0064] In step S103, the learning guidance model receives a user's (i.e., learner's) restatement or explanation of a concept, such as: "Plants can magically grow organic matter as long as they get warm sunshine and water." The learning guidance model can then simultaneously perform semantic and logical integrity verification and knowledge blind spot location through semantic listening channel and logical auditing channel.

[0065] The logical verification and heuristic training data are used to train the learning guidance model in a targeted manner. Before combining semantic listening and logical auditing to analyze and judge the logical coherence and knowledge gaps of the content, the method further includes the step of: training the learning guidance model in a targeted manner using the training data, so that the learning guidance model has the function of verifying the logical integrity of the conceptual restatements output by learners and locating knowledge blind spots.

[0066] In this embodiment, logical verification and heuristic training data are first constructed to provide a foundation for model training. This logical verification and heuristic training data can be built based on open-source commonsense graphs (such as ConceptNet) and private domain graphs. Subsequently, a learning guidance model with a dual-channel path including a semantic listening channel and a logical auditing channel is constructed. Through training data, the model learns methods for logical integrity verification and knowledge blind spot location. In practical application, the learner inputs concept restatement content (such as "Because it rained, the ground was wet, so earthworms crawled out"). After receiving the input, the learning guidance model first processes the semantic and logical information through the dual channels, then comprehensively analyzes and judges logical breakpoints (such as the missing "oxygen deficiency" link), and finally generates heuristic follow-up questions (such as "What is the most direct survival challenge for earthworms living in the soil after it is soaked by water?").

[0067] Compared to existing technologies where learners struggle to obtain high-quality feedback with logical review and guidance capabilities in paraphrasing-based comprehension learning, and where existing large-scale language models lack core logical review capabilities, suffer from fluency biases, and rely on insufficient stability due to their reliance on prompt words, these approaches fail to effectively support the core needs of this type of learning.

[0068] This embodiment, through a physically separated parallel structure of semantic listening and logical auditing, enables the learning guidance model to possess both "emotional listening" and "rational review" capabilities. Therefore, it successfully simulates the core capabilities of a high-quality listener, ensuring both natural and smooth interaction with learners and accurate identification of logical loopholes. It effectively guides learners to discover knowledge blind spots and solves the core pain point of lacking stable and efficient auxiliary tools in paraphrasing comprehension learning.

[0069] In step S101, the construction of the logical verification and heuristic training data adopts a hybrid graph-based strategy. Step S101 aims to obtain high-confidence domain knowledge graph data as a benchmark for subsequent logical perturbations. Figure 2 As shown, step S101 specifically includes:

[0070] S201. Obtain basic logical connections from the open-source common sense graph;

[0071] S202. Perform semantic segmentation, structured extraction, and natural language reasoning verification on the private domain documents to construct a private domain graph from the unstructured private domain documents;

[0072] S203. The logical paths in the private domain graph are perturbed in a programmed manner using graph perturbation operators to generate erroneous input samples containing logical vulnerabilities.

[0073] S204. The erroneous input samples are reverse-processed using a higher-order teacher model to generate training data for a critical thinking chain that includes logical verification and heuristic questioning. This training data for the critical thinking chain is the logical verification and heuristic guidance training data mentioned in step S101.

[0074] In this embodiment, an open-source graph (such as ConceptNet) is introduced to provide the model with basic common-sense logic (such as "rain -> wet" and "fire -> hot"). Furthermore, for specific vertical domains (such as IT operations and healthcare), a "semantic segmentation-evidence backtracking" process is adopted to automatically construct a high-precision graph from unstructured private documents. The steps for constructing the private domain graph include: using a vector similarity algorithm to perform semantic segmentation on the unstructured documents, generating text slices that maintain logical coherence;

[0075] Based on predefined relation sets and data class constraints, compliant triples are extracted from text slices; the triples are converted into declarative sentence hypotheses, and combined with the original text slices as premise inputs to the natural language inference model, only implication relation triples with confidence scores higher than a preset threshold are retained.

[0076] like Figure 3 As shown, the following is a specific example of the steps involved in constructing a private domain graph:

[0077] S301. Semantic segmentation based on vector similarity:

[0078] This step utilizes a semantic vector model to intelligently segment private documents, ensuring the integrity of the logical chain.

[0079] The tools used for intelligent segmentation include LangChain SemanticChunker, in conjunction with the Embedding model (text-embedding-3-small).

[0080] The intelligent segmentation implementation process includes:

[0081] The input document is split into sentence sequences by periods. .

[0082] Calculate adjacent sentences S i With S i+1 The vector cosine similarity.

[0083] Set a dynamic threshold (such as the 95th percentile of the similarity distribution) to identify points where the similarity drops sharply as semantic inflection points.

[0084] The document is segmented at turning points to generate text chunks that maintain semantic coherence.

[0085] S302. Structured extraction based on structured constraints, including:

[0086] This step utilizes a large model and a structured parser to extract triples that conform to predefined specifications from text slices.

[0087] The implementation tools used for structured extraction include LangChain Extraction Chain, combined with the Pydantic data validation library.

[0088] The structured extraction implementation process includes:

[0089] Define the graph ontology: Pre-define the set of legal relations. .

[0090] Constructing data class constraints: Use Pydantic to define the output structure, forcing the relation field to belong to R. valid Enumeration type.

[0091] Extraction process: Input text slices into the LLM. If the output generated by the model does not conform to the schema definition (such as generating undefined relations), the parser will automatically intercept and discard it, and finally only retain the compliant triple data.

[0092] S303, NLI-based evidence backtracking verification:

[0093] This step utilizes Natural Language Inference (NLI) technology to verify whether the extracted triples are supported by original textual evidence, thus eliminating model illusions.

[0094] The tools used for evidence backtracking verification include: Cross-Encoder model DeBERTa-v3-large-mnli.

[0095] The NLI-based evidence backtracking verification process includes:

[0096] Hypothesis building: The triples to be verified Templates are used to convert the data into declarative sentences H (Hypothesis). For example, (high temperature, leading to, system crash) is converted into "high temperature leads to system crash".

[0097] Implication reasoning: taking the corresponding slice of the original text as the premise P and the statement as the hypothesis. Input the data into the NLI model for classification.

[0098] Judgment logic: The triple is considered valid and stored in the graph database only when the model output category is Entailment and the confidence level is greater than a preset threshold (e.g., 0.9); if the output is Neutral or Contradiction, it is considered an illusion without evidence and is removed.

[0099] In step S203 above, the blind spot simulation based on the "graph perturbation operator" is to simulate the cognitive deficiencies of trainees. It utilizes graph algorithms to programmatically disrupt correct and complete logic (i.e., the so-called "truth," whether from ConceptNet or a self-built graph), thereby generating the aforementioned cognitive deficiencies. Specifically, the blind spot simulation based on the "graph perturbation operator" includes the following steps:

[0100] Truth path sampling: Randomly sampling a causal chain P in the graph. truth (example: ).

[0101] Applying perturbation operators: Define a set of logical perturbation operators and apply them randomly to generate "false inputs":

[0102] Skip operator: Remove intermediate nodes direct connection (Simulated reasoning jump).

[0103] Drop operator: If rely and Force delete (Simulation conditions are missing).

[0104] Reverse operator: ... Reversed to (Simulated causal reversal).

[0105] Output: Produces pairs of samples (true path, false input).

[0106] The following is a specific demonstration of blind zone simulation based on the "graph perturbation operator":

[0107] Original output (structured):

[0108] Full Truth Path: [Rain] --leads to --> [Waterlogging] --leads to --> [Oxygen Deprivation] --leads to --> [Earthworm Emergence]

[0109] Corrupted Path (after jump operator processing): [Rain] -- leads to --> [Waterlogging] -- leads to --> [Earthworms emerging from the soil]

[0110] The jump operator (removes intermediate nodes) means that the "oxygen-deficient" node is deleted.

[0111] In this embodiment, a general-purpose LLM such as DeepSeek-V3.2 is used to batch translate the Corrupted_Path into spoken language. The learning guide model prompts the user to rewrite the logical chain "[rain]--leads to-->[water accumulation]--leads to-->[earthworms emerging from the soil]" into a middle school student's spoken explanation.

[0112] Output text (Corrupted_Input): "Because it rained, the ground got wet, so the earthworms crawled out."

[0113] In step S204 above, a higher-order teacher model (GPT-5 or equivalent) is used to transform the structured perturbation samples (Corrupted_Input) generated in step S203 into natural language training data containing explicit thought processes. The system prompt for the learning guidance model is: "You are proficient in paraphrasing and other paraphrasing learning methods such as the Feynman Learning Technique. Your task is to compare [reference logic] and [user paraphrasing], identify the user's logical blind spots, and generate a thought chain (Thought) and a heuristic question (Heuristic_Questioning)." The requirements for the learning guidance model include:

[0114] 1. In the thought process, write in detail: the missing node (target node) found through comparison, the logical verification process, and the guidance strategy.

[0115] 2. In the response, you must guide the user to fill in the missing parts themselves by using "rhetorical questions" or "indirect hints".

[0116] 3. It is strictly forbidden to directly disclose the name of missing nodes. For example, if "photosynthesis" is missing, you cannot directly say "Is photosynthesis missing?", but you should ask "How do plants convert light energy into chemical energy?".

[0117] [Reference Logic]: {P_truth_text} (e.g.: Rain -> Soil water accumulation -> (leading to) -> Oxygen deficiency -> (leading to) -> Earthworms emerging from the soil).

[0118] [User paraphrase]: {Corrupted_Input} (e.g., because it rained, the ground was wet, so the earthworms crawled out).

[0119] [Task]: Please generate a thought chain. <thought>And heuristic questioning.

[0120] The data sample generated in step S204 above is shown below:

[0121] [Reference Logic] is: Rain -> Soil water accumulation -> (leading to) -> Oxygen deficiency -> (leading to) -> Earthworms emerging from the soil. [User's restatement] Because it rained, the ground became wet, so the earthworms crawled out.

[0122] Compare the reference logic "rain -> waterlogged soil -> (leading to) -> lack of oxygen -> (leading to) -> earthworms emerging from the soil" with the user's restatement "because it rained, the ground was wet, so the earthworms crawled out".

[0123] Missing Node Analysis: The user's description captured the initial cause (rain) and the final phenomenon (earthworms emerging), mentioning an intermediate state (wet ground / waterlogged soil). However, the causal relationship from "wet ground" to "earthworms emerging" in the user's description is direct and abrupt, omitting the crucial logical bridge—how "waterlogged soil" specifically and necessarily "led" to "earthworms emerging." The reference logic clearly points to "oxygen deficiency" as the intermediate link, which constitutes a complete causal chain: waterlogging causes soil pores to be filled with water, thereby displacing air and creating an oxygen-deficient environment for earthworms. To breathe air, earthworms are forced to emerge from the soil.

[0124] Logical verification process: The user's logical chain is "rain -> wet ground -> earthworms emerge." There's a logical gap between "wet ground" and "earthworms emerge." "Wet ground" itself doesn't necessarily directly drive away earthworms. We need to ask: Is the soil merely "wet," or has some deeper state change occurred? How does this change specifically affect the earthworms' survival or behavior, forcing them to leave the soil? The "oxygen deficiency" node provided by the reference logic answers the question: "What specific changes in conditions unfavorable to earthworm survival are caused by waterlogged soil?"

[0125] Guiding Strategy: Avoid directly stating the concept of "oxygen deficiency." Instead, use indirect hints and rhetorical questions to guide users to consider questions like, "Besides being 'wet,' what other physical or chemical properties of the soil change after it's filled with water?" and "How do these changes specifically challenge the earthworms' survival needs?" The goal is to allow users to deduce the concept of "respiratory obstruction" or similar ideas themselves, thus completing the causal chain.

[0126] Heuristic question: Do you think that "wet ground" directly makes earthworms feel "uncomfortable" and causes them to come out, or does it actually trigger a more fundamental change within the soil, making it impossible for earthworms to "stay" burrowed? Think about it: what are the most direct survival challenges that waterlogged soil poses to the earthworms living within it?

[0127] In this embodiment, constructing the logic verification and heuristic-guided training data further includes a closed-loop quality verification step based on graph path reconstruction; this closed-loop quality verification step based on graph path reconstruction includes:

[0128] The learner's response to the heuristic probing questions was simulated using a low-order reasoning model;

[0129] Generative grading of simulated responses is performed using a high-order model to determine whether the responses hit the target logical nodes in the graph. Only when the simulated responses are valid, the grading determines that the target node has been hit, and the heuristic follow-up questions do not reveal the name of the target node, is the training data of the corresponding critical thinking chain included in the final training set.

[0130] An automated grading mechanism is employed in the closed-loop quality verification step. The student's answer (Corrupted in step S203) is combined with the teacher's heuristic questioning (from step S204). A general small model is used to simulate the student's answer (the small model can be from Qwen3-8b / 4b level) to verify whether the logical correction of the student's answer (Corrupted) after heuristic questioning has logically repaired the broken path of the knowledge graph. This determines whether the teacher's heuristic questioning is effective and ultimately decides whether to retain the training data.

[0131] The schematic implementation process of the closed-loop quality verification step based on graph path reconstruction is as follows:

[0132] Sub-step 1.4.1: Simulate student reasoning:

[0133] The reasoning of learners (e.g., beginners) is simulated using a standard low-order reasoning model (qwen3-8b), with the generated test samples as input.

[0134] The prompt template for low-order reasoning models is as follows:

[0135] [Previous text - User question]: {User_Input} (e.g.: Earthworms came out because it rained);

[0136] [Teacher's Questions - Heuristic Follow-up Questions]: {Heuristic Questioning} (For example: Do you think that "wet ground" directly makes earthworms feel "uncomfortable" and cause them to run out, or does it actually trigger a more fundamental change inside the soil, making it impossible for earthworms to "stay" in the soil? Think about it, what is the most direct survival challenge that waterlogged soil poses to the earthworms living in it?)

[0137] [Instruction] Please fill in the missing logical links in the above text according to the teacher's prompts, and fill in the most critical cause-and-effect link missing in the above text with a complete sentence.

[0138] Get the output (Student_Response): Suppose the output of the low-order inference model is: "Rainwater filled the soil cracks and squeezed out the air".

[0139] Sub-step 1.4.2: Generative marking and entity mapping based on a large model:

[0140] Then, another large model at the GPT-5 level is used as the "examiner" to accurately determine whether the student's natural language response hits the target concept in the knowledge graph.

[0141] Instructions for marking papers (design of prompts):

[0142] [Background Knowledge]: Standard concept in knowledge graph (Target Node): "{Target_Node}" (e.g., lack of oxygen); [Student Response]: "{Student_Response}" (e.g., rainwater filled the cracks in the soil, squeezing out the air);

[0143] The task for the "exam grader" model: Please grade papers like a strict physics / biology teacher.

[0144] 1. Analyze the deeper meaning behind the students' answers.

[0145] 2. Determine whether the student's description is scientifically equivalent to or necessarily leads to the concept in "[Background Knowledge]".

[0146] 3. If yes, output the extracted standard concept ID; otherwise, output None.

[0147] The output format of the "Examiner" model is as follows: "reasoning": "The student's answer describes the process of water replacing air, which leads to a decrease in oxygen content in the soil environment, which is biologically equivalent to 'hypoxia'.", "mapped_entity": "{Target_Node} or None".

[0148] Operation logic:

[0149] Fill the above Prompt with Student_Response and the masked Target_Node.

[0150] Call the large model API to get JSON output.

[0151] Decision logic:

[0152] If the mapped_entity field is not empty and equal to Target_Node, then the review is considered passed.

[0153] If mapped_entity is None, it means the student's answer was off-topic or incorrect, and the grading will fail.

[0154] Sub-step 1.4.3: Final judgment and data entry:

[0155] Passing Data: A C-CoT (Critical Chain of Thought) data entry will only be written to the final training set finetune_dataset.json if all three of the following strict conditions are met:

[0156] Response validity: The simulated student in substep 1.4.1 output non-empty text.

[0157] The marking was passed: In sub-step 1.4.2, the large model determined that the student's answer hit the target node in terms of scientific essence (i.e., mapped_entity == Target_Node), proving that the teacher model's "heuristic questioning" does indeed have the ability to guide students to discover the truth on their own.

[0158] Leakage prevention security check: Perform string checks on Heuristic_Questioning (teacher's heuristic probing questions) to ensure that they do not directly contain the literal name of Target_Node. For example, if the target is "oxygen deficiency", the question must not contain words such as "oxygen deficiency" or "oxygen", and must be guided indirectly (such as "Where did the air go?").

[0159] Data Drop: Any data that does not meet the following conditions (such as a student answering incorrectly, the grading judgment being off-target, or the teacher's question directly revealing the answer) is considered low-quality invalid data and will be directly discarded and not included in the training set.

[0160] Based on the above steps, an automated CoT (Chain of Thought) data construction process (including data validation) can be designed.

[0161] In step S102, constructing the dual-channel language processing learning guidance model includes:

[0162] Logical skeleton extraction based on dependency parsing;

[0163] Physical grouping and parallel computation of dual attention channels; and

[0164] Feature splicing and end-to-end task-driven fine-tuning.

[0165] The logical skeleton extraction in dependency parsing involves constructing a topological mask reflecting the deep structure of a sentence using dependency parsing techniques. Dependency parsing describes the "social relationships" between words. Logic often follows the main "subject-verb-object" structure, while adjectives typically attach to branches representing "modification relationships." Therefore, by calculating the distance and relationship type of words in the syntax tree, it's possible to precisely retain only the "trunk" and remove the "branches."

[0166] Logical skeleton extraction based on dependency parsing includes:

[0167] Sub-step 2.1.1: Dependency parsing and relation definition; and

[0168] Sub-step 2.1.2: Construction of topology mask based on dependency path analysis.

[0169] In this embodiment, the dependency tree structure of the input text is parsed by a dependency parser; a set of core logical relations and a set of modification relations are defined; the dependency paths of any word pair in the text are analyzed, and only word pair connections whose paths are all core logical relations are retained, while word pair connections containing modification relations are masked, generating a topological logic mask matrix.

[0170] In substep 2.1.1, the input sequence is parsed into a dependency tree structure using a syntactic parser (SpaCy). Each word t i With its parent node t j There exists a labeled edge R(t) between them. i ,t j ).

[0171] Define a "core set of core logical relations Rcore", which includes: subject, object, core predicate, compound words, adverbial clauses, attribute definitions, etc.

[0172] At the same time, we define "the set of modification relations Rnoise" (i.e., noise), which includes: adjective modification, adverb modification, determiner, punctuation and auxiliary verb, etc.

[0173] In sub-step 2.1.2, the fully connected mask matrix M is initialized. logic For any word pair (t) in the input sequence i ,t j Perform the following topology analysis:

[0174] Path search: Searching for join t in the dependency syntax tree. i With t j The only dependency path p i-j .

[0175] Relationship pruning: Traversing path p i-j All dependency tags on.

[0176] Masking condition: If the path contains any tag belonging to Rnoise, it indicates that t i With t j The connection depends on non-core modifying components, which will M i,j Set as (Forced blocking).

[0177] Retention condition: Only when all tags on the path belong to Rcore, it indicates that the two are connected by a strong logical link such as subject-verb, verb-object, or compound words, thus connecting M. i,j Set to 0 (allow attention).

[0178] An illustrative, complete demonstration of topology mask construction based on dependency path analysis includes:

[0179] Taking the input sentence "A red apple falls due to gravity" as an example, the logical mask constructed by the system will look like this:

[0180] (1) The final retained logical skeleton (M=0):

[0181] (Apple, dropped): The path is nsubj (subject), reserved.

[0182] (Gravity, falling): The path is pobj -> prep (due to the preposition), keep it.

[0183] (Because, dropped): The path is prep (prepositional modifier), and is retained.

[0184] (Gravity, because): The path is pobj (prepositional object), preserved.

[0185] Logical view: Apple -- falls -- because -- gravity (forming a complete "entity-cause-entity" chain).

[0186] (2) The final masking modifier (M=-∞):

[0187] (Red, Apple): Path is amod (adjective modifier), masked.

[0188] (Red, dropped): The path contains amod, so it's blocked.

[0189] (Down, fall): The path is advmod (adverb modifier / directional complement, depending on the parser's classification, usually classified as noise), masked.

[0190] Effect: Logic header H c By making the color (red) and tendency (down) completely "invisible," one can avoid being disturbed by non-causal information.

[0191] This logic skeleton extraction based on dependency parsing achieves logical purification based on semantic structure and can accurately cut off the interference of modifying components on the core reasoning, forcing attention to focus on the causal backbone composed of "entity (apple / gravity) - relation (drop)".

[0192] The physical grouping and parallel computation of the conceptual restatement content in the aforementioned semantic listening channel and logical auditing channel include:

[0193] The multi-head attention heads of the learning guidance model are physically divided into semantic listening channels and logical auditing channels according to a certain ratio, wherein the number of attention heads allocated to the semantic listening channels is greater than that of the logical auditing channels.

[0194] The semantic listening channel receives a standard causal mask and outputs semantic features according to the self-attention calculation logic; wherein the standard causal mask is an upper triangular mask matrix, which is used to constrain the semantic listening channel to capture contextual semantics according to the text temporal order, so as to ensure the temporal consistency of semantic understanding.

[0195] The logic audit channel injects a logic mask matrix and outputs logic features through self-attention calculation with logical constraints. These logic features are used to capture the logical dependencies between entities in the text and encode the logical association state of the text. When the learner's concept restatement content has a broken logical chain or loopholes, the logic features present an abnormal signal, providing a core judgment basis for subsequent logical integrity verification and knowledge blind spot location.

[0196] In this embodiment, when constructing a learning guidance model with multiple attention heads, the model divides attention heads according to their functional proportions, binds standard causal masks to paraphrased content, and uses end-to-end fine-tuning driven by loss functions. This allows the learning guidance model to retain the model's natural language understanding capabilities while forcing it to focus on the logical skeleton of the input text through physical division of labor and masking mechanisms. This enables the model to achieve low-level perception of logical breakpoints and solves the technical problem of existing models that focus on semantic understanding but have poor logical and completeness examination capabilities.

[0197] The specific working process is as follows: Taking a Transformer model with 32 attention heads as an example, 24 are assigned to the semantic listening channel (Hs) and 8 to the logical auditing channel (Hc). Among them, the semantic listening channel (Hs) is still responsible for "understanding" the user's language to ensure that the dialogue can continue. The core function of the logical auditing channel (Hc) is to "distinguish between false and true".

[0198] Semantic listening channel (Hs):

[0199] Semantic listening channel (Hs) configuration: allocate 75% of attention heads (i.e., Ns=24).

[0200] The input to the semantic listening channel (Hs) is a standard causal mask, without imposing any additional logical restrictions.

[0201] Calculation logic:

[0202] ;

[0203] Logical audit channel (Hc):

[0204] Configuration of the logical audit channel (Hc): Allocate 25% of the attention head (i.e. indivual).

[0205] Input to the logic audit channel (Hc): Inject the above logic mask matrix M logic ;

[0206] Calculation logic:

[0207] ;

[0208] Due to M logic Due to constraints, the logic audit channel (Hc) cannot focus on "modifying words". It is forced to establish long-distance dependencies between nouns (entities) and verbs (relationships). If there is a break in the logical chain of user input (such as the absence of a corresponding noun entity after a causal word), the attention distribution of the logic audit channel (Hc) will be abnormal, thus encoding logical fallacy signals in the underlying features.

[0209] For example, when a student says something illogical but eloquent (such as "rain gently summons earthworms out"), the traditional Transformer sees "gently summons" as semantically coherent and tends to agree with it.

[0210] In this embodiment, the logic audit channel (Hc) is forcibly shielded from "gentle" interference, only seeing the two isolated entities, "rainwater" and "earthworm," with no physical connection between them. At this point, the logic audit channel (Hc) will generate a "feature mismatch" signal. This signal is the physiological basis for the model's "questioning" behavior.

[0211] In this embodiment, the learning guidance model also concatenates the output features of the semantic listening channel (Hs) and the logical auditing channel (Hc) and performs end-to-end task-driven fine-tuning, thereby adaptively coordinating the operation of the two channels, the semantic listening channel (Hs) and the logical auditing channel (Hc).

[0212] First, the output features of the semantic listening channel (Hs) and the output features of the logical auditing channel (Hc) are concatenated and then mapped back to the original feature dimensions through a linear projection layer.

[0213] Then, the concatenated model is fine-tuned end-to-end using the training data of Critical Thinking Chain (C-CoT), enabling the model to adaptively coordinate the dual-channel feature weights during backpropagation, thus achieving dynamic collaboration between logical judgment and semantic interaction.

[0214] In this embodiment, multi-view feature fusion is performed first. The semantic listening channel output Output_s and the logical audit channel output Output_c are concatenated along the feature dimension and then mapped back to the original dimension through a linear projection layer.

[0215] ;

[0216] in This indicates a concatenation operation, where Wo is a learnable weight matrix.

[0217] Next, task-driven implicit collaboration is executed, including:

[0218] Training Data: The learning guidance model is trained using the previously constructed logic verification and heuristic training data. Specifically, the logic verification and heuristic training data can serve as data for Critical Thinking Chains (C-CoT). This data contains explicit... <thought>The tags require learning the guided model's output of blind spot localization and logical deduction processes.

[0219] Cooperative Mechanism: During backpropagation, to minimize the loss of the C-CoT generation task, the gradient automatically updates the weights related to the output. The model will discover that to accurately generate... <thought>Logical deduction in the process relies more efficiently on the "skeleton features" provided by the logical audit channel (Hc) than on the semantic listening channel (Hs).

[0220] Result: Through this data-driven approach, the model naturally learns to "focus on listening to the left ear (logic audit channel)" when making logical judgments and "focus on listening to the right ear (semantic listening channel)" when engaging in casual conversation.

[0221] The principle of the aforementioned collaborative mechanism is as follows:

[0222] Training on C-CoT data (i.e., logic verification and heuristic-guided training data) establishes a conditioned reflex in the learning-guided model:

[0223] When the logic audit channel (Hc) detects "skeleton connectivity", it calls the semantic listening channel (Hs) to provide a compliant response (such as "You're right").

[0224] When the logic audit channel (Hc) issues a "feature mismatch" alarm (i.e., a skeleton break is detected), the model is trained to suppress the compliant tendency of the semantic listening channel (Hs) and instead activate the "questioning module" to output heuristic follow-up questions. Therefore, the learning-guided model focuses on logical loopholes, and once a loophole is found, the fine-tuning mechanism forces it to ask follow-up questions instead of continuing to elaborate on the user's words (or even making things up).

[0225] Therefore, in this embodiment, the above-mentioned improvements to the learning guidance model solve the technical pain point of "language fluency masking logical flaws" in the "Feynman Learning Technique" scenario:

[0226] Combating statistical co-occurrence bias: Traditional models, based on probabilistic predictions, tend to assume the co-occurrence of "rain" and "earthworms" is natural, easily overlooking intermediate causal jumps. This invention addresses this bias through M... logic The masking mechanism forces the logic audit channel to block descriptive words, preventing it from utilizing the shallow statistical patterns of language and forcing it to directly observe the sparse connections between entities, thereby keenly perceiving logical breaks.

[0227] Achieving "Dual Listening": The physical splitting mechanism enables the model to possess both "emotional understanding (semantic head)" and "rational examination (logical head)." The former ensures the natural and smooth interaction, while the latter ensures the rigor of blind spot detection, perfectly matching the Feynman Learning Technique's role requirement of "being both a patient listener and a rigorous error corrector."

[0228] The practical application of this learning guidance model also includes:

[0229] Receive the learner's concept restatement text (a restatement or explanation of a concept), for example: "Plants can magically grow organic matter as long as they are exposed to warm sunshine and watered."

[0230] Extract the logical skeleton of the text in real time and generate the corresponding logical mask, including:

[0231] This learning-guided model utilizes a built-in lightweight dependency parser to process the input text in real time.

[0232] Operation: Identify "warm" (modifier) ​​and "as long as / magically" (adverb / noise).

[0233] Generate a mask: Construct a logical mask matrix (M) logic The above-mentioned modifying components are masked (set to -∞), and only the connections of the core framework such as "plant, sun, sun, watering, growth, organic matter" are retained.

[0234] The text is encoded in parallel through a dual-channel path. If a logical feature mismatch signal is detected, compliant responses are suppressed and a heuristic follow-up question generation module is activated to output follow-up questions that guide learners to complete the logic.

[0235] The specific working process is as follows: After receiving the concept restatement text input by the learner, the learning guidance model processes it in real time through a built-in lightweight dependency parser, identifies modifying components, and generates logical masks; the dual-channel path encodes the text simultaneously, the semantic listening channel captures the fluency of expression, and the logical auditing channel detects the logical connection status. When a logical feature mismatch is detected (such as a lack of necessary logical connections between entities), the learning guidance model suppresses compliant responses such as "You're right" and activates the follow-up question generation module; for example, for the restatement of "Plants can magically grow organic matter as long as they get warm sunshine and water," the model outputs the follow-up question "Your explanation covers energy and solvents, but plants are carbon-based life forms, so where does the 'carbon' element that makes up their bodies come from?"

[0236] This learning guidance model enables real-time responses to learners' retelling of content, accurately triggers logical follow-up questions, effectively guides learners to independently complete the logical chain, and helps deepen their understanding of knowledge.

[0237] Please see Figure 4 In another embodiment, a learning device 400 based on a dual-channel neural network attention mechanism is provided, the learning device 400 comprising:

[0238] Storage module 401 is used to store logic verification and heuristic training data;

[0239] A dual-channel language processing learning guidance model 402 is configured as a listener for paraphrasing comprehension learning and adopts a semantic and logical dual-channel parallel structure. The semantic listening channel is used to capture the semantic fluency and referential relationships of the text context, and the logical auditing channel is used to focus on the core logical skeleton of the text. The learning guidance model includes, but is not limited to, Transformer-like models and other approximate language processing models based on multi-head attention mechanisms.

[0240] Application module 403 is used by the learning guidance model to receive the learner's concept restatement content, and to analyze and judge the logical coherence and knowledge gaps of the content by combining semantic listening and logical auditing, and to generate corresponding heuristic follow-up questions to guide the learner to complete the logical chain.

[0241] In some embodiments, the learning device 400 based on the dual-channel neural network attention mechanism further includes a training module, which is used to perform targeted training on the learning guidance model using the training data, so that the learning guidance model has the function of verifying the logical integrity of the conceptual restatement content output by the learner and locating knowledge blind spots.

[0242] like Figure 5 As shown, in another embodiment, a computer-readable storage medium 500 is provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the learning method based on the dual-channel neural network attention mechanism described in any of the above embodiments.

[0243] The computer-readable storage medium can be volatile memory or non-volatile memory, or both. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), ferromagnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); the magnetic surface memory can be disk storage or magnetic tape storage.

[0244] The volatile memory may be random access memory (RAM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), synchronous static random access memory (SSRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synclink dynamic random access memory (SLDRAM), and direct memory bus random access memory (DRRAM). The computer-readable storage media described in the embodiments of the present invention are intended to include these and any other suitable types of memory.

[0245] In some embodiments, the processor may be implemented by software, hardware, firmware, or a combination thereof, and may be a circuit, one or more of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, or a microprocessor, thereby enabling the processor to execute some or all of the steps or any combination thereof in the learning method based on the dual-channel neural network attention mechanism described in the various embodiments of this application.

[0246] Finally, it should be noted that although the above embodiments have been described in the text and drawings of this application, this should not limit the scope of patent protection of this application. Any technical solutions that are based on the essential concept of this application and utilize the content described in the text and drawings of this application, resulting in equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this application.< / thought> < / thought> < / thought>

Claims

1. A learning method based on a dual-channel neural network attention mechanism, characterized in that, include: Pre-construct logical verification and heuristic-guided training data; A dual-channel language processing learning guidance model is constructed. The learning guidance model is configured as a listener for paraphrasing comprehension learning and adopts a dual-channel parallel structure of semantic listening and logical auditing. The semantic listening channel is used to capture the semantic fluency and referential relationships of the text context, and the logical auditing channel is used to focus on the core logical skeleton of the text. The learning guidance model includes, but is not limited to, Transformer-like models and other approximate language processing models based on multi-head attention mechanisms. The learning guidance model receives the learner's concept restatement, and analyzes and judges the logical coherence and knowledge gaps of the content through semantic listening and logical auditing, generating corresponding heuristic follow-up questions to guide the learner to complete the logical chain.

2. The learning method based on a dual-channel neural network attention mechanism according to claim 1, characterized in that, The construction of the dual-channel language processing learning guidance model includes: physically grouping the semantic listening channel and the logical auditing channel, wherein the physical grouping includes: physically dividing the multi-head attention heads of the learning guidance model into semantic listening channels and logical auditing channels in proportion; The semantic listening channel and the logical auditing channel are capable of parallel computation, which includes: The semantic listening channel receives a standard causal mask and outputs semantic features according to the self-attention calculation logic; wherein the standard causal mask is an upper triangular mask matrix, which is used to constrain the semantic listening channel to capture contextual semantics according to the text temporal order, so as to ensure the temporal consistency of semantic understanding. The logic audit channel injects a logic mask matrix and outputs logic features through self-attention calculation with logical constraints. These logic features are used to capture the logical dependencies between entities in the text and encode the logical association state of the text. When the learner's concept restatement content has a broken logical chain or loopholes, the logic features present an abnormal signal, providing a core judgment basis for subsequent logical integrity verification and knowledge blind spot location.

3. The learning method based on a dual-channel neural network attention mechanism according to claim 2, characterized in that, After physically grouping the semantic listening channel and the logical auditing channel, the process also includes feature splicing and end-to-end task-driven fine-tuning of the semantic listening channel and the logical auditing channel. The output features of the semantic listening channel and the output features of the logical auditing channel are concatenated and mapped back to the original feature dimension through a linear projection layer; By using training data from the critical thinking chain, the concatenated learning guidance model is fine-tuned end-to-end, enabling the learning guidance model to adaptively coordinate the dual-channel feature weights during backpropagation, thus achieving dynamic collaboration between logical judgment and semantic interaction.

4. The learning method based on a dual-channel neural network attention mechanism according to claim 1, characterized in that, Before combining semantic listening and logical auditing to analyze and judge the logical coherence and knowledge gaps of the content, the method further includes the step of: training the learning guidance model in a targeted manner using the training data, so that the learning guidance model has the function of verifying the logical integrity of the concept restatement content output by the learner and locating knowledge blind spots.

5. The learning method based on a dual-channel neural network attention mechanism according to claim 1, characterized in that, The construction of the logical verification and heuristic-guided training data adopts a hybrid graph-based strategy, specifically including: Obtain basic logical connections from the aforementioned open-source common sense graph; Semantic segmentation, structured extraction, and natural language reasoning verification are performed on private domain documents to construct a private domain graph from unstructured private domain documents; The logical paths in the private domain graph are perturbed programmatically using graph perturbation operators to generate erroneous input samples containing logical loopholes; the erroneous input samples are then reverse-processed using a higher-order teacher model to generate training data containing a critical thinking chain that includes logical verification processes and heuristic probing questions.

6. The learning method based on a dual-channel neural network attention mechanism according to claim 5, characterized in that, The construction of the logical verification and heuristic-guided training data also includes a closed-loop quality verification step based on graph path reconstruction: The learner's response to the heuristic probing questions was simulated using a low-order reasoning model; Generative grading of simulated responses is performed using a high-order model to determine whether the responses hit the target logical nodes in the graph. Only when the simulated responses are valid, the grading determines that the target node has been hit, and the heuristic follow-up questions do not reveal the name of the target node, is the training data of the corresponding critical thinking chain included in the final training set.

7. The learning method based on a dual-channel neural network attention mechanism according to claim 2, characterized in that, Before injecting the logical audit channel into the logical mask matrix, the logical mask matrix is ​​constructed. The steps include: parsing the dependency tree structure of the input text through a dependency parser; defining the core logical relation set and the modification relation set; analyzing the dependency path of any word pair in the text, retaining only word pair connections whose paths are all core logical relations, and masking word pair connections containing modification relations, thereby generating a topological logical mask matrix.

8. The learning method based on a dual-channel neural network attention mechanism according to claim 1, characterized in that, It also includes a follow-up questioning step to complete the logic: after receiving the learner's concept restatement text, the learning guidance model extracts the logical skeleton of the text in real time and generates a corresponding logical mask; the text is encoded in parallel through the dual-channel path; if a logical feature mismatch signal is detected, compliant responses are suppressed and the heuristic follow-up questioning generation module is activated to output follow-up questions to guide the learner to complete the logic.

9. A learning device based on a dual-channel neural network attention mechanism, characterized in that, include: The storage module is used to store logical verification and heuristic training data; A dual-channel language processing learning guidance model is configured as a listener for paraphrasing comprehension learning and adopts a dual-channel parallel structure of semantic listening and logical auditing. The semantic listening channel is used to capture the semantic fluency and referential relationships of the text context, while the logical auditing channel is used to focus on the core logical skeleton of the text. The learning guidance model includes, but is not limited to, Transformer-like models and other approximate language processing models based on multi-head attention mechanisms. The application module is used by the learning guidance model to receive the learner's concept restatement content, capture the semantic fluency and referential relationship of the context using the semantic listening channel, evaluate the logical coherence and knowledge gaps of the content using the logical auditing channel, and generate heuristic follow-up questions based on these to guide the learner to improve the logical chain.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the learning method based on a dual-channel neural network attention mechanism as described in any one of claims 1-8.

Images