A post-reading subsequent writing intelligent correction and classroom simulation interaction platform

Abstract

This application relates to the fields of natural language processing and smart education infrastructure technology, and discloses an intelligent grading and classroom simulation interaction platform for reading and writing continuation. The platform extracts reference text to construct a basic state map, extracts characters from handwritten text to construct a continuation state map, and generates a multi-dimensional defect feature matrix by comparing and extracting topological breakage features. Based on the feature matrix, it clusters execution objects and calculates defect complementarity to generate a grouping strategy. Based on the grouping strategy and historical writing fluctuation characteristics, it dynamically generates typesetting instructions with isolation buffers. It scans the physical medium to extract multi-region handwritten text data, reconstructs it into an interactive text tree, compares it with the basic state map, calculates penalty values, and outputs evaluation data. This invention transforms subjective semantic differences into topological structural differences, realizes complementary grouping of execution objects and dynamic typesetting of the medium, avoids cross-boundary optical scanning, and takes into account both continuation deviation and plot omission verification.

Description

Technical Field

[0001] This invention relates to the field of natural language processing and smart education infrastructure technology, specifically to a reading-to-writing intelligent grading and classroom simulation interactive platform. Background Technology

[0002] With the development of educational informatization, AI-powered text correction technology has been widely applied in auxiliary teaching scenarios. Reading continuation writing, a complex writing format requiring students to logically extend from given materials, not only tests learners' basic language expression abilities but also emphasizes plot coherence and the rigor of logical thinking. Most existing intelligent correction systems rely on basic natural language processing models, primarily targeting and correcting surface features such as spelling, vocabulary, and conventional grammar.

[0003] However, current systems have significant limitations when handling complex read-to-write tasks. Existing technologies focus excessively on correcting local language norms, failing to establish a structured mapping of deep narrative logic between the reference text and the continued text. They cannot quantitatively compare the topological features of plot logic coherence, resulting in a lack of objective and operational indicators for assessing content deviations. Furthermore, existing teaching interaction mechanisms often employ homogeneous or random grouping patterns, lacking multi-dimensional extraction and clustering analysis of the logical and linguistic flaws of the target audience, and failing to achieve targeted collaborative resource allocation based on the complementarity of flaws. In addition, in teaching sessions integrating offline paper-and-pen writing with online digital assessment, existing equipment lacks dynamic physical typesetting and optical isolation control mechanisms. When dealing with collaborative physical media, conventional optical scanning is prone to cross-boundary extraction and recognition errors, making it difficult to reconstruct discrete, multi-region handwritten text in memory according to accurate logical sequence. This creates a data gap between the physical writing space and the digital assessment loop, ultimately leading to one-sided assessment feedback and failing to meet the needs of in-depth interactive teaching.

[0004] Therefore, this invention proposes a reading-and-writing intelligent grading and classroom simulation interactive platform to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent grading and classroom simulation interaction platform for reading and writing continuation. It solves the problems of existing technologies, such as the lack of deep logical topology comparison indicators when processing reading and writing continuation, the inability to achieve targeted collaborative grouping based on defect complementarity, and the lack of dynamic typesetting isolation control when multiple people collaborate on physical media writing, resulting in data discontinuity in optical extraction and reconstruction.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a reading-and-writing intelligent grading and classroom simulation interactive platform, comprising: The graph extraction module extracts the core entities and relational edges of the standard reference text to construct a basic state graph and marks the dangling nodes in the basic state graph. The fault-tolerant analysis module extracts character nodes from handwritten text images to construct a continuation state map, compares the continuation state map with the basic state map to extract topological break features, and combines language standard features to generate a multi-dimensional defect feature matrix. The feature clustering module clusters the execution objects based on the multidimensional defect feature matrix and calculates the defect complementarity to generate grouping strategy data. The dynamic typesetting module divides the writing area according to the grouping strategy data, generates an isolation buffer zone between adjacent writing areas based on historical writing characteristics, and outputs typesetting instructions. The optical reconstruction module scans the collaborative physical medium generated based on the typesetting instructions and extracts multi-region handwritten text data along the isolation buffer zone; The verification closed-loop module reconstructs the multi-region handwritten text data into an interactive text tree, compares it with the basic state map to calculate the total penalty value, and outputs the evaluation data.

[0007] Preferably, the graph extraction module is specifically used for: calling a pre-trained entity extraction model to perform sequence labeling on the standard reference text to obtain the core entities; obtaining the association probability by calculating the concatenation matrix of the hidden feature vectors between the core entities; establishing the core entities with an association probability greater than or equal to a preset association probability threshold as valid relation edges; traversing each node in the basic state graph and calculating the degree value; calling a semantic analysis algorithm to calculate the semantic completeness of the terminal nodes with a degree value of zero; and marking nodes with an out-degree value of zero and a semantic completeness value lower than the semantic completeness threshold as dangling nodes with unresolved attributes.

[0008] Preferably, the fault-tolerant analysis module is specifically used for: performing morphological filtering on the handwritten text image, extracting the character nodes using an optical character recognition model, and constructing the continuation state graph after removing character nodes whose recognition confidence weight is lower than a preset threshold; using the suspended nodes as logical reference points for comparison, calculating the node mapping and edge mapping relationship between the reference substructure of the basic state graph and the continuation state graph using a subgraph isomorphic matching algorithm; calculating the topological breakage degree of the reference substructure based on the nodes and relation edges of the missing matching items, combined with the node weight coefficient and the relation edge weight coefficient, and generating the topological breakage feature by dimensional combination according to the node sequence.

[0009] Preferably, the feature clustering module is specifically used for: calling the principal component analysis algorithm to perform orthogonal transformation on the multidimensional defect feature matrix to extract principal component features and obtain a dimensionality-reduced feature matrix; extracting the feature vectors corresponding to the execution objects in the dimensionality-reduced feature matrix, and using the K-means clustering algorithm to divide the execution objects into defect type clusters; obtaining the defect complementarity by calculating the vector inner product and magnitude between the feature vectors of the execution objects in different defect type clusters, and pairing the execution objects whose defect complementarity is greater than or equal to the complementarity threshold to generate the grouping strategy data.

[0010] Preferably, the dynamic typesetting module is specifically used for: parsing the grouping strategy data to obtain the number of execution objects and the object number sequence; dividing the preset available writing area into multiple independently allocated writing areas according to the proportion; extracting the logical starting point coordinates of the suspended nodes; generating a logical timing encoding graphic based on the collaboration order of the execution objects and inserting it into a visual typesetting file; calculating the width of the isolation buffer band through the preset minimum optical segmentation spacing, boundary fault tolerance coefficient, and standard deviation of the writing fluctuation of the execution objects; and superimposing the two-dimensional boundary of the writing area, the logical timing encoding graphic, and the isolation buffer band layer to generate the typesetting instruction.

[0011] Preferably, the optical reconstruction module is specifically used for: acquiring a two-dimensional image matrix of the cooperative physical medium; locating the visual anchor pixel coordinates of the two-dimensional image matrix using an edge detection algorithm; comparing the visual anchor pixel coordinates with the standard coordinates inside the system and constructing a system of linear equations to solve for the rotation, translation, and scaling parameters of the affine transformation matrix; resampling the two-dimensional image matrix using a bilinear interpolation algorithm based on the affine transformation matrix; and outputting a corrected image matrix.

[0012] Preferably, the optical reconstruction module is further configured to: cut the correction image matrix into multiple independently distributed sub-region images according to the coordinate position of the isolation buffer; parse the logical temporal coding graphics in the correction image matrix to obtain the logical splicing order; arrange the sub-region images in memory in one dimension according to the logical splicing order; and input the arranged sub-region images into the optical character recognition model in sequence to extract the multi-region handwritten text data.

[0013] Preferably, the verification closed-loop module is specifically used for: allocating data storage nodes for the multi-region handwritten text data in memory space, parsing the logical splicing order to establish a directed data pointer; using the directed data pointer, performing memory addressing and head-to-tail connection of the multi-region handwritten text data according to the logical topology direction; and instantiating the multi-region handwritten text data into the interactive text tree through the head-to-tail connection.

[0014] Preferably, the verification closed-loop module is specifically used to calculate the total penalty value by: calling the text feature extraction model to calculate the cosine similarity between the text nodes contained in the interactive text tree and the constraint rules contained in the basic state graph, and calculating a positive penalty term for content deviation by combining the positive penalty weight coefficient; calculating the cosine similarity between the text nodes contained in the interactive text tree and the necessary nodes contained in the basic state graph, and calculating a negative penalty term for plot omission by combining the negative penalty weight coefficient; and accumulating the positive penalty term and the negative penalty term to generate the total penalty value.

[0015] Preferably, the verification closed-loop module is specifically used to: obtain a preset initial benchmark score, subtract the product of the total penalty value and the preset score conversion coefficient from the initial benchmark score, calculate the final evaluation score of the corresponding execution object; encapsulate the interactive text tree, the total penalty value and the final evaluation score into structured data and store them in the system database.

[0016] This invention provides an intelligent grading and classroom simulation interaction platform for reading and writing continuation. It has the following beneficial effects: 1. This invention constructs a basic state graph and a continuation state graph, employs a subgraph isomorphic matching algorithm to calculate the node and edge mapping relationships between the reference substructure and the continuation state graph, and combines the extracted topological break features and language normative features into a multidimensional defect feature matrix. This feature processing mechanism transforms the subjective semantic differences of natural language sequences into quantitative data differences in network topology, enabling the system to accurately locate the logical deviations of the continuation text through feature numerical calculations, thus improving the objectivity and accuracy of text defect feature extraction.

[0017] 2. This invention uses a feature matrix to cluster execution objects, generates a grouping strategy by calculating the complementarity of feature vectors of different defect types, and dynamically calculates the width of the isolation buffer zone between writing areas of the collaborative medium based on the historical writing fluctuation standard deviation of the execution objects. This typesetting control process achieves targeted collaborative grouping of execution objects based on the defect complementarity dimension, and dynamically adjusts the physical typesetting spacing by introducing individual writing fluctuation characteristics, avoiding optical scanning boundary crossing and image segmentation extraction errors that are prone to occur in fixed typesetting modes.

[0018] 3. This invention obtains the splicing order by parsing the logical temporal encoding graphics on the physical medium, instantiates the extracted multi-region discrete handwritten text into an interactive text tree in memory, and introduces a two-way penalty mechanism to calculate positive penalty terms for content deviation and negative penalty terms for missing necessary plot points. This mechanism establishes a data mapping chain from the physical medium spatial text to the system logical tree, achieving automatic alignment and reconstruction of cross-region handwritten text while taking into account both the constraints of continuation direction and the completeness of core plots, ensuring the comprehensiveness of the closed-loop evaluation score. Attached Figure Description

[0019] Figure 1 This is a structural block diagram of the intelligent grading and classroom simulation interactive platform for reading and writing continuation according to the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a schematic diagram of the defect feature clustering and complementary pairing distribution state of the present invention; Figure 4 This is a schematic diagram of the three-dimensional response curve for optimizing the dynamic isolation buffer zone of the present invention; Figure 5 This is a schematic diagram illustrating the evolution trend of the total penalty value within the system's operating cycle according to the present invention.

[0020] Among them, 100 is the knowledge graph processing layer; 110 is the graph extraction module; 120 is the fault tolerance analysis module; 200 is the physical medium scheduling layer; 210 is the feature clustering module; 220 is the dynamic typesetting module; 300 is the multimodal reconstruction layer; 310 is the optical reconstruction module; and 320 is the verification closed-loop module. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See attached document Figure 1 , Figure 1 This is a structural block diagram of a read-and-write intelligent grading and classroom simulation interaction platform according to an embodiment of the present invention. The present invention provides a read-and-write intelligent grading and classroom simulation interaction platform, including a knowledge graph processing layer 100, a physical media scheduling layer 200, and a multimodal reconstruction layer 300.

[0023] The knowledge graph processing layer 100 is used to extract information from standard reference text and to compare the logical structure of the collected test text data. The knowledge graph processing layer 100 includes a graph extraction module 110 and a fault-tolerant analysis module 120. The graph extraction module 110 is used to construct the basic network topology. The fault-tolerant analysis module 120 is used to process noisy input text and generate the corresponding structured state graph.

[0024] The physical media scheduling layer 200 is used to control the terminal hardware to generate physical media with spatial sequence metadata based on the topological defects of the map, and output visual labels. The physical media scheduling layer 200 includes a feature clustering module 210 and a dynamic layout module 220. The feature clustering module 210 is used to perform clustering calculations on the feature matrix and generate a grouping strategy. The dynamic layout module 220 is used to generate layout instructions based on the grouping strategy and drive external output devices.

[0025] The multimodal reconstruction layer 300 is used to extract physical medium information through the image acquisition components, eliminate physical image distortion, and complete data topology connection in a virtual coordinate system. The multimodal reconstruction layer 300 includes an optical reconstruction module 310 and a verification closed-loop module 320. The optical reconstruction module 310 processes two-dimensional scanned images and extracts serialized text. The verification closed-loop module 320 performs logical verification on the extracted text data and updates the system database.

[0026] See attached document Figure 2 , Figure 2 This is a flowchart of a read-and-write intelligent grading and classroom simulation interaction method according to an embodiment of the present invention. The present invention provides a read-and-write intelligent grading and classroom simulation interaction method, comprising the following steps: S100: Generate a basic state graph for the standard reference text, traverse the topological degree of the nodes in the basic state graph, and mark the nodes with zero out-degree and unresolved attributes as dangling nodes. S200: Acquire handwritten text sequence images collected by the terminal and extract the handwritten text sequence, calculate the confidence weight of the handwritten text sequence, and construct a continuation state map based on the confidence weight; S300: Compare the basic state map with the continuation state map, extract topological fracture features, and combine the topological fracture features with language standard features to generate a multi-dimensional defect feature matrix. S400 clusters and complements the execution objects according to the multidimensional defect feature matrix, and controls the printing device to output cooperative physical media according to the grouping results. The cooperative physical media has logical timing coding and isolation buffer bands for region segmentation. S500 performs optical scanning on the cooperative physical medium to obtain a two-dimensional image matrix, analyzes the visual anchor points in the two-dimensional image matrix, performs affine transformation calculations based on the visual anchor points, and dynamically aligns the transformed spatial sequence. S600 concatenates the extracted multi-region text in the system memory according to the logical topology direction based on logical timing encoding. It performs closed-loop verification between the concatenation result and the basic state map by calculating a bidirectional penalty function, and generates evaluation data based on the verification result to synchronously update the system database.

[0027] To further clarify the implementation of each technical aspect of the present invention, the following will provide a detailed description of the implementation of each functional module involved above and its internal processing flow.

[0028] In this embodiment, the graph extraction module 110 is used to construct the network topology structure of the standard reference text, which can be achieved through the following steps: S111, Core Entity and Relationship Edge Extraction Calculation: Receive standard reference text data and use natural language processing algorithms to extract core entities and relationship edges from the standard reference text. In this embodiment, the classification system of core entities specifically includes person entities, location entities, time entities, and event entities; the classification system of relationship edges specifically includes causal relationships, temporal logical relationships, action subordinate relationships, and referential relationships. Specifically, the graph extraction module 110 calls a pre-trained entity extraction model to perform sequence labeling on the standard reference text to obtain the core entities in the text. In this embodiment, the entity extraction model adopts a bidirectional long short-term memory network structure. This entity extraction model is trained in a supervised manner based on an educational corpus containing large-scale labeled text. The training samples in the corpus contain manually labeled entity classification tags. During training, backpropagation is performed by calculating the cross-entropy loss function to update the model network parameters. Core entities and relationship edges constitute the main plot of the standard reference text. By extracting the above information and transforming the core entities into vector space features, the logical connections between entities can be quantified, thereby filtering out weakly correlated interference data. Furthermore, after extracting the core entities, the graph extraction module 110 inputs the core entities into the output hidden layer feature vectors of the bidirectional long short-term memory network structure, and calculates the association probability between the core entities through feature concatenation. The formula for calculating the association probability is as follows: ; In the formula: Represents the core entity With core entities The probability of association between them; This represents the activation function; in this embodiment, the Sigmoid function is used. This represents the system's preset weight matrix, which is determined through backpropagation iterations during the model training phase. Represents the core entity The hidden layer feature vectors; Represents the core entity The hidden layer feature vectors; Represents the core entity Hidden feature vectors and core entities The concatenation matrix of hidden layer feature vectors; This indicates the calculation of the bias term, the value of which is updated synchronously during model training.

[0029] After obtaining the association probability, the graph extraction module 110 sets an association probability threshold. Preferably, the association probability threshold ranges from 0.75 to 0.85, with a system default value of 0.80. When the association probability is greater than or equal to the association probability threshold, the graph extraction module 110 establishes the corresponding core entities as valid relation edges. For the network layer construction and backpropagation training process of the entity extraction model, those skilled in the art can use conventional deep learning frameworks, and its structure and compilation are well-known technologies in the field, and will not be elaborated here.

[0030] S112, Generate the basic state graph: Based on the core entities and relational edges, initialize a directed graph data structure in system memory to generate the basic state graph. In the specific implementation, the graph extraction module 110 traverses all nodes in the basic state graph, counting and calculating the in-degree and out-degree values ​​of each node. The out-degree value of a node represents the total number of relational edges originating from that node and pointing to other nodes in the basic state graph. The graph extraction module 110 stores the out-degree and in-degree values ​​as the topological degree attributes of the nodes in the system database.

[0031] S113, Identify Suspended Nodes: Identify and mark suspended nodes with unresolved attributes in the basic state graph. In this embodiment, the graph extraction module 110 extracts terminal nodes with a degree value of zero through traversal operations and calls a semantic analysis algorithm to calculate the semantic completeness of the terminal nodes. Specifically, the graph extraction module 110 inputs the text fragments corresponding to the terminal nodes into a pre-trained text classifier, analyzes the syntactic dependency tree structure of the text fragments to determine whether they are missing core grammatical components, and outputs a value between 0 and 1 as the semantic completeness based on the terminal punctuation features. The text classifier adopts a network model architecture that includes a self-attention mechanism. The data construction method of its training samples is as follows: extract sentences and their preceding context fragments from a large-scale standard education corpus, and manually label the corresponding completion state according to whether the overall semantic logic forms a closed loop. To avoid the failure of single-node judgment when the text fragment is a complex sentence or has cross-sentence logical dependencies, the graph extraction module 110 introduces a context window containing a preset number of predecessor nodes (such as the first 2 to 3 nodes) when extracting the text fragment corresponding to the terminal node. The feature concatenation sequence of the terminal node and its predecessor nodes is input into the text classifier to accurately define the boundary of semantic completion. Furthermore, the graph extraction module 110 sets a semantic completion threshold of 0.60 with a precision requirement of 0.01. When the semantic completion degree of the terminal node is lower than the semantic completion threshold, the graph extraction module 110 determines that the corresponding node is not complete in terms of logical event and assigns an unresolved attribute label to the terminal node.

[0032] Subsequently, the graph extraction module 110 performs a conditional search on the basic state graph, identifying nodes that simultaneously meet the criteria of having an out-degree value of zero and an unresolved attribute label as dangling nodes. The graph extraction module 110 then marks the dangling nodes with coordinate data in the graph database. Each dangling node carries its topological coordinate data in the basic state graph, which is stored and retrieved as the logical starting point coordinates for the subsequent text continuation task executed by the physical media scheduling layer 200.

[0033] In this embodiment, the fault-tolerant analysis module 120 is used to process noisy input text, construct a structured state graph of the corresponding handwritten text, and extract relevant defect features. Specifically, this can be achieved through the following steps: S121: Acquire handwritten text sequence images captured by the terminal, and perform morphological filtering on the handwritten text sequence images to eliminate background noise. In specific implementations, the original images captured by the terminal device usually contain noise such as paper wrinkles or ink stains. Specifically, the fault-tolerant analysis module 120 uses an opening operation of erosion followed by dilation to remove isolated noise in the image, and a closing operation of dilation followed by erosion to fill in the broken areas of character strokes, thereby restoring the continuity of handwritten characters. For the setting of the convolution kernel size and the specific matrix operation of the image morphological filtering, those skilled in the art can configure it according to the actual acquisition resolution. The processing process is a well-known technology in the field and will not be described in detail here. Further, after processing, the fault-tolerant analysis module 120 calls a pre-trained optical character recognition model to extract characters from the processed handwritten text sequence images to obtain each discrete character node; in this embodiment, the optical character recognition model adopts a convolutional recurrent neural network architecture containing convolutional layers, recurrent layers, and fully connected classification layers, and is trained on an image dataset with character position and category labels.

[0034] S122, calculate the recognition confidence weight of each character node in the handwritten text sequence, remove invalid characters, and construct a continuation state graph. In this embodiment, the recognition confidence weight of each character node is output by the final classification layer of the optical character recognition model. This value reflects the probability that the model correctly judges the current character form. The fault tolerance analysis module 120 sets the recognition confidence threshold internally. Preferably, to ensure the accuracy of text extraction, the recognition confidence threshold is set to a range of 0.85 to 0.95, with a system default value of 0.90. In practical operation, the fault-tolerant analysis module 120 traverses all character nodes. To avoid excessive sparsity in the constructed continuation state graph due to continuous character removal, and to address the issue of high-confidence but semantically incorrect characters (such as misidentification of similar-looking characters) not being captured, the system introduces a fusion mechanism of character-level and word-level confidence and a context correction mechanism: For low-confidence characters or characters suspected of semantic conflict, the fault-tolerant analysis module 120 extracts their local context, inputs it into a pre-trained masked language model to calculate the word-level context probability of the candidate character, and performs a weighted fusion of the character-level confidence of optical recognition with the word-level probability. If the fused confidence is still lower than the recognition confidence threshold set by the system, the removal operation is performed; otherwise, the candidate character with the highest probability output by the language model is used for dynamic replacement and correction. For the valid text composed of the remaining nodes, the fault-tolerant analysis module 120 performs the same entity extraction and relation calculation logic as the graph extraction module 110, constructing a continuation state graph representing the logical structure of the handwritten text in the system memory.

[0035] S123, compare the continuation state graph with the basic state graph to extract topological break features. In this embodiment, the core principle of graph comparison is to transform abstract textual semantic deviations into specific network topological differences, enabling the system to accurately locate the fault positions in the text that deviate from the reference logic using numerical calculations. In specific implementation, the basic state graph contains the core plot of the standard reference text, and the fault-tolerant analysis module 120 uses the suspended nodes in the basic state graph as the logical benchmark for comparison. The fault-tolerant analysis module 120 uses a subgraph isomorphic matching algorithm to calculate the node mapping and edge mapping relationships between the reference substructures of the basic state graph (i.e., local topological branches extending backward to a preset depth with the corresponding suspended nodes as the logical starting point) and the continuation state graph. If a node or relation edge existing in the basic state graph cannot find a corresponding mapping item in the continuation state graph, the fault-tolerant analysis module 120 determines that a logical deviation has occurred at the corresponding position. In order to transform structural differences into computable data, the fault-tolerant analysis module 120 calculates the topological break degree for each reference substructure. The formula for calculating topological breakage is as follows: ; In the formula: This represents the calculated topological breakage degree; This represents the node weight coefficient, used to adjust the influence of missing nodes on the overall fracture characteristics, with a value ranging from 0.4 to 0.6. This represents the total number of nodes to be matched in the basic state graph. Indicates the first One node to be matched; This represents a node state indicator function, when the node... The function value is 1 when there is a missing item in the continuation state graph, and 0 when there is a matching item. This represents the relation edge weight coefficient, used to adjust the impact of missing relation edges. Its value range is also from 0.4 to 0.6, and the sum of the system's mandatory constraint node weight coefficient and the relation edge weight coefficient is 1. This represents the total number of relation edges to be matched in the basic state graph; Indicates the first One relation edge to be matched; This represents the state indicator function for relation edges, when the relation edge... The function value is 1 when a node is missing from the continuation state graph, and 0 when a matching term exists. It should be noted that, to avoid the penalty of repeated calculations for missing related edges due to missing nodes, the relationship edge state indicator function follows a conditional constraint: the function value is 1 only when both ends of the relationship edge have matching terms in the continuation state graph, but the link between them is missing; otherwise, it is 0. Furthermore, to prevent the calculated topological breakage from being dominated by a few extreme missing terms, the system performs a smooth logarithmic transformation on the calculation results before the final output. Through the above calculations, the fault-tolerant analysis module 120 combines the topological breakage values ​​corresponding to each reference substructure according to the dimension of the graph node sequence to generate a topological breakage feature vector.

[0036] During the comparison process, to improve the error tolerance, in addition to judging strict existence, node mapping also introduces word vector semantic similarity calculation to accommodate entity aliases and near-synonyms. When the basic state map is empty due to the reference text being too short, or when the continuation text fails to construct a map due to OCR recognition failure, the system sets up an abnormal rollback mechanism, directly assigning the highest breakage penalty. Furthermore, the fixed thresholds involved in this embodiment (such as 0.80, 0.60, etc.) are baseline empirical values ​​calculated based on the system's pre-built corpus. The system supports dynamic adaptive adjustment of the above thresholds according to the genre attributes of the input text (such as the difference in completion requirements between narrative and argumentative texts).

[0037] S124, the topological break features and language standardization features are vector-combined to generate a multi-dimensional defect feature matrix. In a preferred embodiment, in addition to comparing the logical topological dimension, the fault-tolerant analysis module 120 also calls the language standardization detection model to perform syntactic analysis on the text composed of the remaining nodes. This language standardization detection model adopts a sequence-to-sequence network architecture based on the Transformer structure, which is trained on a large-scale text corpus containing grammatical error labels and is used to identify lexical and syntactic error types in the text sequence. The fault-tolerant analysis module 120 converts the identified specific grammatical errors into independent feature values ​​of the corresponding dimension and merges them to form a language standardization feature vector. Further, the fault-tolerant analysis module 120 performs max-min normalization processing on the extracted topological break feature vector and language standardization feature vector to eliminate dimensional differences, and after aligning in the feature space, performs feature concatenation operation to generate a comprehensive feature vector for a single handwritten text. The fault-tolerant analysis module 120 arranges and stacks the comprehensive feature vectors corresponding to multiple handwritten texts collected concurrently by the system in a row-wise order to generate a multi-dimensional defect feature matrix. Ultimately, this multidimensional defect feature matrix comprehensively includes logical break defects in the global topological structure of the text under test and language standardization defects in the local sentence construction. It is then output to the physical media scheduling layer 200. This matrix is ​​not only used to quantify the individual continuation deviation of the execution object, but also serves as the data basis for the subsequent K-means clustering algorithm based on defect type to perform differentiated complementary grouping, thereby forming a complete teaching closed loop from defect detection to customized physical collaborative intervention.

[0038] In this embodiment, the feature clustering module 210 is used to perform grouping calculations of execution objects based on the feature matrix, which can be implemented through the following steps: S211, the multidimensional defect feature matrix output by the fault tolerance analysis module 120 is received, and dimensionality reduction processing is performed on the multidimensional defect feature matrix. Specifically, the multidimensional defect feature matrix contains the comprehensive feature vectors of multiple execution objects. In the classroom simulation interaction scenario of this embodiment, the above-mentioned execution objects are specifically mapped to each student terminal device participating in the continuation task and its corresponding operating user, and the system pre-assigns a unique identity identifier for data traceability to each execution object. Further, in order to reduce the complexity of subsequent calculations and eliminate multicollinearity between feature dimensions, the feature clustering module 210 calls the principal component analysis algorithm to perform orthogonal transformation on the multidimensional defect feature matrix to extract the core principal component features whose cumulative variance contribution rate reaches the set requirements. Preferably, the cumulative variance contribution rate is set to a range of 85% to 95%, and the system default value is 90%. For the covariance matrix calculation and eigenvalue decomposition process in the principal component analysis algorithm, those skilled in the art can directly implement it based on the principles of matrix algebra, and its dimensionality reduction processing is a well-known technology in the field, which will not be elaborated here. The core technical significance of this dimensionality reduction process lies in the fact that the topological fracture features and language specification features contained in the multidimensional defect feature matrix are correlated in the original dimensional space. Through the orthogonal transformation of principal component analysis, the two can be decoupled into linearly independent principal component dimensions. This allows subsequent clustering operations to be performed based on independent feature axes that best reflect the essence of the defects of the execution object, thereby improving the accuracy of defect type cluster division and the effectiveness of cross-cluster complementary pairing. After dimensionality reduction, the feature clustering module 210 obtains the dimensionality-reduced feature matrix.

[0039] S212, clustering and complementary grouping calculations are performed on the execution objects based on the dimensionality-reduced feature matrix. In this embodiment, the core logic of complementary grouping is to assign execution objects with different types of text defects to the same collaborative group. Through differentiated feature combinations, the execution objects can promote mutual error correction and inspiration in subsequent collaborative physical media. In specific operation, the feature clustering module 210 extracts the feature vector corresponding to each execution object in the dimensionality-reduced feature matrix on a per-execution-object basis. Subsequently, the feature clustering module 210 uses the K-means clustering algorithm to divide execution objects with similar defect features into the same defect type cluster based on the Euclidean distance between the feature vectors. To achieve substantial complementarity, the feature clustering module 210 selects execution objects for cross-cluster pairing between different defect type clusters and calculates the defect complementarity between execution objects. The formula for calculating the defect complementarity is as follows: ; In the formula: Indicates the execution object With the execution object The degree of complementarity between defects; Indicates the execution object eigenvectors; Indicates the execution object eigenvectors; This represents the vector dot product operation; This represents the magnitude of the vector. It should be noted that this formula reflects the differences in defect types by calculating the cosine distance between feature vectors. Specifically, the larger the angle between the vectors, the lower the overlap between the two execution objects in terms of error type and logical deviation direction, and the higher the calculated defect complementarity value. In specific cross-cluster pairing operations, the feature clustering module 210 preferentially adopts a greedy algorithm: it traverses the execution objects within different defect type clusters, pairing the two objects with the highest defect complementarity greater than the complementarity threshold one-to-one; when the number of remaining execution objects within a cluster is unequal, the system allows the use of a one-to-many pairing mode to ensure that all execution objects can be assigned to the corresponding cooperative group according to the principle of optimal complementarity.

[0040] S213, Set a complementarity threshold and convert the grouping results into a grouping strategy data structure. To ensure that paired execution objects have the required complementary effect, the feature clustering module 210 sets a complementarity threshold internally. Preferably, the complementarity threshold is set to a range of 0.60 to 0.80, with a system default value of 0.70. For execution objects with a defect complementarity greater than or equal to the complementarity threshold in cross-cluster calculations, the feature clustering module 210 combines and pairs them to form the final complementary groups. After completing the pairing calculations for all execution objects, the feature clustering module 210 maps the member information of each complementary group to the corresponding object number sequence. Finally, the feature clustering module 210 generates structured grouping strategy data based on the object number sequence and the system's preset data format specifications, and transmits this grouping strategy data to the dynamic layout module 220 as the control basis for subsequently generating physical layout instructions.

[0041] In this embodiment, the dynamic layout module 220 is used to execute the layout control process of converting the grouping strategy into a physical medium with spatial attributes, which can be achieved through the following steps: S221, Receive grouping strategy data and calculate layout space coordinates. In this embodiment, the dynamic layout module 220 parses the grouping strategy data passed in by the front-end module to obtain the number of execution objects in the same group and the corresponding object number sequence. Further, the dynamic layout module 220 obtains the standard size parameters of the collaborative physical medium preset by the system. Based on the number of execution objects in the same group, after deducting the physical medium margins preset by the system, the dynamic layout module 220 divides the usable write area of ​​the physical medium into multiple independently allocated writing areas unequally according to the average writing size in the historical learning records of each execution object, and calculates the boundary coordinates of each writing area in two-dimensional space.

[0042] S222, Generate a logical timing encoding graphic and insert it into the visual layout file. In specific implementation, to clarify the sequential order and logical connection relationship of the continuation tasks of different execution objects on the collaborative physical medium, the dynamic layout module 220 extracts the logical starting coordinates of the corresponding suspended nodes from the graph database, and performs data packetization processing in combination with the collaboration order of the execution objects in the grouping strategy. The packetized data specifically includes a group identifier code, a unique identifier of each execution object in the group, a collaboration order number, coordinate parameters of the corresponding suspended nodes, and a check bit used to verify the integrity of the information. Subsequently, the dynamic layout module 220 converts the packetized data into a machine-readable matrix optical barcode graphic, such as a data matrix code or a QR code, to form a logical timing encoding graphic. Further, the dynamic layout module 220 inserts this logical timing encoding graphic into the preset coordinate position of the generated visual layout file as the positioning and sorting benchmark for subsequent device scanning and recognition.

[0043] S223, calculate and draw isolation buffer bands between the writing areas of different execution objects. In this embodiment, to prevent the handwriting of the execution object from crossing boundaries during actual writing, thereby avoiding boundary overflow and image segmentation errors in the subsequent optical scanning process, the dynamic typesetting module 220 needs to draw rectangular geometric blank bands as isolation buffer bands between adjacent writing areas. The core calculation principle of this step is that if a buffer band of fixed width is used, it is easy for execution objects with standardized writing to waste media space, while it cannot play an isolation role for execution objects with arbitrary writing; therefore, the system introduces historical writing features for dynamic calculation to balance media space utilization and the isolation security of optical scanning. The formula for calculating the width of the isolation buffer band is as follows: ; In the formula: This represents the calculated width of the isolation buffer zone; This represents the minimum optical segmentation spacing preset by the system to meet the basic resolution requirements of external scanning equipment. Preferably, its value ranges from 5.0 mm to 10.0 mm, and the system default value is 8.0 mm. This represents the boundary tolerance coefficient, used to adjust the degree of writing fluctuation of the target audience at different ages. The value ranges from 1.2 to 1.5, and the specific value is determined by the system based on the grade attribute of the target audience. Lower grade target audiences correspond to larger values, and higher grade target audiences correspond to smaller values. This represents the standard deviation of the overflow length of the handwriting of the corresponding execution object exceeding the standard boundary in the system's historical records. This standard deviation is calculated statistically based on the absolute distance of the measured handwriting pixel coordinates exceeding the standard writing frame boundary in the historical scan images of the corresponding execution object from the most recent preset number of scans (e.g., the last 10 scans). For users with historical records, the system extracts the latest overflow data and continuously updates this standard deviation value using a sliding window mechanism after each newly submitted physical media scan image is acquired and processed. Furthermore, when the corresponding execution object is an initial user lacking historical records, the system defaults to... The value is set to 2.0 mm. Through the above calculations, the dynamic typesetting module 220 can dynamically adjust the width of the isolation buffer band according to the historical writing characteristics of the execution object, thereby generating typesetting instructions with the isolation buffer band.

[0044] S224, driving the external output device to perform media generation. In specific operation, the dynamic typesetting module 220 overlays typesetting instructions containing the two-dimensional boundary of the writing area, logical timing encoding graphics, and isolation buffer parameters to generate complete cooperative physical media typesetting instructions. Further, the dynamic typesetting module 220 calls the device driver to convert the cooperative physical media typesetting instructions into a standard printer control language. Finally, the dynamic typesetting module 220 sends the converted printer control language to the connected external output device through the communication interface to control the external output device to print the cooperative physical media. The compilation of the printer control language and the underlying communication transmission mechanism can be implemented using existing page description language protocols, which are well-known technologies in the field and will not be elaborated upon here.

[0045] In this embodiment, the optical reconstruction module 310 is used to perform an image processing procedure that extracts spatial text information from the cooperative physical medium and performs coordinate alignment. Specifically, this can be achieved through the following steps: S311, a two-dimensional image matrix of the collaborative physical medium is acquired through optical scanning. In a specific implementation, the optical reconstruction module 310 controls the image acquisition component to perform an optical scanning operation on the collaborative physical medium after writing, so as to generate a two-dimensional image matrix and store it in the system memory or video memory. Specifically, the image acquisition component can be a scanner or a document camera, etc. Further, in order to ensure the accuracy of subsequent image processing and character recognition, the system internally presets the optical scanning resolution of the image acquisition component. Preferably, the optical scanning resolution is set to a range of 300 DPI to 600 DPI, and the system default value is 300 DPI. After scanning is completed, the optical reconstruction module 310 converts the acquired analog image signal into a digitized two-dimensional image matrix and temporarily stores it in the system's storage space as the basis data for subsequent calculations.

[0046] S312, analyze the visual anchor points and perform affine transformation calculations to eliminate physical deformation. In this embodiment, the collaborative physical medium is prone to physical deformation during actual circulation and writing due to human folding or tilting, leading to misalignment in subsequent text segmentation. It should be noted that the core principle of the spatial coordinate mapping here is to reverse the pixel offset caused by paper tilting or deformation by establishing a mathematical mapping relationship between the physical scanning space and the standard typesetting space, thereby ensuring the boundary accuracy when the image is segmented by region. Based on this principle, the optical reconstruction module 310 uses an edge detection algorithm to traverse the two-dimensional image matrix and locate the pixel coordinates of the visual anchor points at the four corners of the two-dimensional image matrix. The above-mentioned visual anchor points are positioning marks pre-generated in the previous dynamic typesetting stage, and their specific form is preferably a solid cross mark or an L-shaped corner mark with high contrast. The edge detection algorithm, combining morphological template matching and pixel grayscale thresholding, can distinguish visual anchor points from ordinary handwriting. If a visual anchor point cannot be fully extracted due to handwriting occlusion or contamination, the system will use the remaining unoccluded visual anchor points and the intersection points of the physical edge lines of the paper to perform coordinate estimation and redundancy repair. Subsequently, the optical reconstruction module 310 extracts the actual coordinates of each visual anchor point and compares them with the preset standard coordinates within the system. In specific operation, the optical reconstruction module 310 uses the correspondence between the actual coordinates and standard coordinates of these four visual anchor points to construct a system of linear equations and uses the least squares method to solve for the parameters in the affine transformation matrix. The formula for calculating the pixel coordinate transformation is as follows: ; In the formula: Represents the x-coordinate of the target pixel after transformation; Represents the ordinate of the target pixel after transformation; This represents the affine transformation matrix calculated based on the visual anchor point. The matrix contains spatial transformation parameters for rotation, translation, and scaling. Its matrix dimension is 3 rows and 3 columns, and to satisfy the mathematical constraints of affine transformation, the elements in its third row are fixed as [0,0,1]. Represents the actual pixel x-coordinate in the original two-dimensional image matrix; This represents the actual pixel ordinate in the original two-dimensional image matrix. Further, the optical reconstruction module 310, based on the calculated pixel coordinate correspondence, uses a bilinear interpolation algorithm to interpolate and resample the original two-dimensional image matrix, thereby outputting a corrected image matrix after eliminating physical deformation. The specific implementation of the edge detection algorithm and the matrix operation logic of the bilinear interpolation can be directly called using existing computer vision algorithm libraries by those skilled in the art; the processing is well-known in the field and will not be elaborated upon here.

[0047] S313, dynamically align the transformed spatial sequence and extract multi-region handwritten text data. In specific implementation, the optical reconstruction module 310 reads the coordinate position parameters of the isolation buffer zone generated by the pre-module. Subsequently, the optical reconstruction module 310 cuts the correction image matrix into multiple independently distributed sub-region images according to the coordinate position of the isolation buffer zone. At the same time, the optical reconstruction module 310 calls the decoding algorithm to parse the logical temporal encoding pattern in the correction image matrix. After verifying the data integrity through check bits, it extracts the group identification code, identity identifier, and collaboration sequence number to obtain the logical splicing order corresponding to different sub-region images. Further, the optical reconstruction module 310 uses the parsed logical splicing order as an array index to arrange each independently distributed sub-region image in the system memory in a one-dimensional order, thereby completing the dynamic alignment of the spatial sequence. Finally, the optical reconstruction module 310 inputs the aligned sub-region image sequence into the pre-trained optical character recognition model to extract the corresponding multi-region handwritten text data, and outputs the multi-region handwritten text data to the subsequent module as the original text input for full text evaluation.

[0048] In this embodiment, the closed-loop verification module 320 is used to perform text topology splicing, closed-loop verification, and system data update processes, which can be implemented through the following steps: S321, multi-region text is concatenated in memory according to logical topological direction based on logical timing encoding. In this embodiment, the verification closed-loop module 320 receives multi-region handwritten text data extracted by the optical reconstruction module 310 and parses the logical concatenation order obtained in the previous step. Further, the verification closed-loop module 320 allocates corresponding data storage nodes for each discrete text sequence in the system's memory space and establishes directed data pointers according to the parsed logical concatenation order. Subsequently, the verification closed-loop module 320 uses the above-mentioned directed pointers to perform continuous memory addressing and head-to-tail connection of the discrete text sequences according to the logical topological direction, thereby instantiating the text sequence into a complete interactive text tree data structure in memory.

[0049] It should be noted that in this interactive text tree, nodes represent independently extracted sentences or core event semantic components, and the edges formed by directed pointers represent the temporal and logical relationships mapped by the spatial arrangement of the text sequence on the physical medium. This tree structure is a digital representation of the reconstructed handwritten text, and its data format is fully compatible with the aforementioned basic state graph and continuation state graph, and can be directly used for subsequent bidirectional penalty and node comparison calculations.

[0050] S322, the bidirectional penalty function is calculated to perform a closed-loop verification between the splicing result and the basic state graph. In specific implementation, the verification closed-loop module 320 compares the instantiated interactive text tree with the system's preset basic state graph at both the structural and content levels. The basic state graph internally defines a set of constraint rules used to define the logical direction of the continuation, as well as a set of necessary nodes representing the core plot. It should be noted that the core calculation principle of introducing the bidirectional penalty function here is that the system needs to both constrain and record errors that deviate from the preset plot direction and verify errors that omit core plots. When performing node comparison, the system needs to convert natural language text into computer-processable quantitative features. The formula for calculating the total penalty value is as follows: ; In the formula: This represents the calculated total penalty value; This represents the positive penalty weighting coefficient, used to adjust the proportion of logical deviation errors in the total penalty value. Preferably, its value range is set to 0.4 to 0.6, and the system default value is 0.5. This represents the total number of text nodes in the interactive text tree. Represents the first in the interactive text tree One text node; This indicates that the node deviates from the judgment function. Specifically, the verification closed-loop module 320 calls the pre-trained text feature extraction model to classify the text nodes. The constraint rules of the basic graph are mapped to text feature vectors, and the cosine similarity between them is calculated. This text feature extraction model is a pre-defined neural network model within the system used to convert text sequences into fixed-dimensional vectors. Specifically, this text feature extraction model adopts a dual-tower sentence representation network architecture; during the model training phase, positive samples are logically coherent context text pairs, and negative samples are randomly concatenated text pairs without logical connection. During model training, the network parameters are optimized by calculating a contrastive learning loss function, so that text vectors with similar semantics and logic are closer to each other in the multi-dimensional feature space. When the cosine similarity is lower than the system's pre-defined semantic matching threshold, it is determined to be a content deviation. The value is 1; the function takes a value of 0 when the constraint rules are met. Preferably, the semantic matching threshold is set to a range of 0.70 to 0.85, with a system default value of 0.75. This represents the reverse penalty weighting coefficient, used to adjust the proportion of missing plot errors in the total penalty value. Its value range is set from 0.4 to 0.6, with a system default value of 0.5. Internally, the system forces the sum of the positive penalty weighting coefficient and the reverse penalty weighting coefficient to be 1. This represents the total number of necessary nodes defined in the basic state graph. Represents the first in the basic state diagram One necessary node; This represents the function for determining missing nodes. Similarly, the closed-loop verification module 320 calculates the relationship between each node and the necessary nodes in the interactive text tree. The vector similarity is calculated as follows: when no necessary node with a similarity greater than or equal to the semantic matching threshold is matched in the interactive text tree, the function takes a value of 1; when a necessary node is matched, the function takes a value of 0. Using the above formula, the closed-loop verification module 320 completes the calculation of the total penalty value.

[0051] S323, Generate evaluation data and synchronously update the system database. In specific operation, the verification closed-loop module 320 obtains the system's preset initial benchmark score. Preferably, the system default value of this initial benchmark score is set to 100. The verification closed-loop module 320 subtracts the product of the total penalty value and the preset score conversion coefficient from the initial benchmark score to calculate the final evaluation score of the corresponding execution object. The preset score conversion coefficient is used to map the penalty quantity to the actual deduction value, and its value range is set to 2.0 to 5.0, with a system default value of 2.5. Subsequently, the verification closed-loop module 320 structurally encapsulates the interactive text tree, the total penalty value, and the final evaluation score to generate evaluation data in a standard format. Finally, the verification closed-loop module 320 uploads the evaluation data to the system database through the data communication interface and performs data matching using the unique identifier of the corresponding execution object. The closed-loop verification module 320 appends the evaluation data to the corresponding database table and triggers the system to recalculate the historical evaluation mean and standard deviation of writing fluctuations for the corresponding target, thereby achieving a closed-loop update of the learning data in the entire interactive system. This closed-loop update process includes recursive calculation logic. The recalculated "historical evaluation mean" is the latest result obtained by weighted averaging the final evaluation score calculated in this iteration within the historical sequence. Similarly, the "standard deviation of writing fluctuations" is refitted based on the boundary overflow values ​​measured in this scan, ensuring that the system's intervention strategy can adaptively evolve as the target's capabilities change.

[0052] This embodiment uses an English reading and writing task in an environment without a mobile communication terminal as an application scenario, and elaborates on the specific operation process of each module of the system and the verification principle shown in the attached diagram: Scene Preset and Standard Topology Graph Generation: The system receives reference text input as the starting point of interaction (e.g., the text material where the protagonist Anna receives encouragement from the old man in the coffee shop). The graph extraction module 110 calls a deep learning model to perform semantic dependency analysis, extracting core entities and preceding feature nodes from the text (e.g., the old man's dialogue node "You have a real gift"). The system identifies that the out-degree of this preceding feature node points to a target event node that has not yet been instantiated within the text range, marking it as a "foreshadowing feature node." In the subsequent comparison operation of the loop verification module 320, this node will correspond to a dual semantic mapping that has both "constraint rules" (i.e., the content direction of the continued text should respond to this clue) and "necessary nodes" (i.e., as an indispensable structural element constituting the core plot). Subsequently, the system traverses the end of the text, marking the endpoint nodes that have not formed a plot loop as "suspended nodes," thereby constructing a standard topology graph containing foreshadowing features and suspended nodes as an interaction benchmark.

[0053] Defect Feature Matrix Construction and Complementary Grouping Matching: The system retrieves historical interaction data of the execution object (student). The fault tolerance analysis module 120 extracts a multi-dimensional defect feature matrix through comparison. This matrix specifically includes: plot relevance defects (the degree of topological breakage caused by failure to respond to foreshadowing feature nodes), rhetorical usage defects, and basic language standardization defects.

[0054] The feature clustering module 210 performs dimensionality reduction and clustering calculations on the multidimensional defect feature matrix, calculating the defect complementarity between objects in different clusters. The system configuration strategy determination rule is as follows: when the complementarity between objects exceeds a set threshold, a collaborative grouping strategy is generated; for example, objects with high plot relevance defects but good language standardization are bound to objects with low rhetorical defects but high logical discontinuity as the same collaborative group.

[0055] Combined with appendix Figure 3 Note: Attached Figure 3 This diagram illustrates the data distribution of the aforementioned multidimensional defect features after dimensionality reduction and clustering. The figure maps the data components representing the degree of logical topological breakage to the data components representing the language specification error rate. The clustering algorithm automatically divides the sample data into four feature clusters, and the cross-cluster lines marked in the figure represent the complementary pairing relationships calculated by the system. This data distribution verifies that the multidimensional defect features possess computable mathematical separability, thus providing accurate data support for the system to generate grouping strategies.

[0056] Dynamic typesetting and offline physical media generation: The system obtains the historical writing fluctuation standard deviation parameters of each object within the current group. The dynamic typesetting module 220 calculates the dynamic width of the isolation buffer band according to a preset compensation formula. Considering the physical limitations of the execution object being in an environment without mobile communication terminals, the system generates printing instructions for offline collaborative physical media. These instructions include rendering elements such as: optical visual anchor points at the four corners, a logical timing-encoded graphic (e.g., a QR code) containing object identity information and splicing order, the dynamically calculated isolation buffer band mentioned above, and guiding placeholders embedded based on the defect characteristics of each object. The system sends the instructions to the output device to generate a customized physical paper carrier.

[0057] Combined with appendix Figure 4 Note: Attached Figure 4 The three-dimensional response surface of the dynamic isolation buffer band optimization calculation is shown. This surface reflects the mapping relationship between the handwriting out-of-bounds standard deviation parameter, the boundary tolerance coefficient, and the optical segmentation accuracy. It should be noted that the aforementioned guiding placeholders refer to the dashed boxes or prompts specifically reserved by the system during typesetting based on the defects in the object language specification; the optical segmentation accuracy refers to the probability that the optical reconstruction module 310, when performing image segmentation along the buffer band, does not truncate valid handwriting and successfully separates each region. The height variation trend of the surface indicates that as the handwriting out-of-bounds fluctuation parameter increases, the system dynamically increases the value of the isolation buffer band, ensuring that the mapping area representing the segmentation accuracy always remains above the set target threshold, effectively avoiding text segmentation breaks caused by fixed typesetting templates.

[0058] Scan Reconstruction and Topology Loop Closure Verification: After the execution object completes interactive text writing on the physical medium, the system acquires the image recorded by the optical acquisition device. The optical reconstruction module 310 extracts four visual anchor points from the image to establish an affine transformation matrix to correct image distortion, and automatically binds the corresponding execution object and its splicing order through parsing the logical timing encoded graphics. The system segments the image along the isolation buffer band and extracts the text characters.

[0059] The closed-loop verification module 320 extracts characters using optical character recognition technology and constructs an interactive text tree. The system traverses and compares the interactive text tree with the standard topology graph: if a closed-loop backtracking edge pointing to a "foreshadowing feature node" is detected in the interactive text tree, such as a recall or echo of previous node content in the continued text, the reverse penalty value of that object in the bidirectional penalty function calculation is reduced. Finally, the system outputs quantitative evaluation data and triggers an offline batch printing command to generate a multi-dimensional diagnostic report in paper format, completing the data closed loop in the offline environment.

[0060] Combined with appendix Figure 5 Note: To verify the long-term effectiveness of the system's intervention, see attached... Figure 5The figure illustrates the evolution trend of the total penalty value over the system's operating cycle. The two trend lines represent the experimental group employing the closed-loop verification and complementary heuristic mechanism of this invention, and the control group employing a fixed layout and random grouping mechanism, respectively. The trend evolution shows that after introducing the dynamic layout, foreshadowing closed-loop verification, and defect complementary intervention nodes of this invention, the curve representing the total penalty value of the system output exhibits a significant convergent decreasing trend. This demonstrates that the system can effectively reduce the logical reconstruction error rate in multimodal data interaction and continuously improve the plot construction capability of objects.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A reading-and-writing intelligent grading and classroom simulation interactive platform, characterized in that, include: The graph extraction module is used to extract the core entities and relation edges of the standard reference text to construct a basic state graph, and to mark the dangling nodes in the basic state graph. The fault-tolerant analysis module is used to extract character nodes from handwritten text images to construct a continuation state map, extract topological break features by comparing the continuation state map with the basic state map, and generate a multi-dimensional defect feature matrix by combining language standard features. The feature clustering module is used to cluster the execution objects based on the multidimensional defect feature matrix and calculate the defect complementarity to generate grouping strategy data. The dynamic typesetting module divides the writing area according to the grouping strategy data, generates an isolation buffer zone between adjacent writing areas based on historical writing characteristics, and outputs typesetting instructions. The optical reconstruction module scans the collaborative physical medium generated based on the typesetting instructions and extracts multi-region handwritten text data along the isolation buffer zone; The verification closed-loop module reconstructs the multi-region handwritten text data into an interactive text tree, compares it with the basic state map to calculate the total penalty value, and outputs the evaluation data.

2. The intelligent grading and classroom simulation interactive platform for reading and writing continuation as described in claim 1, characterized in that, The map extraction module is specifically used for: The pre-trained entity extraction model is invoked to perform sequence labeling on the standard reference text to obtain the core entities. The association probability is obtained by calculating the concatenation matrix of the hidden feature vectors between the core entities. Establish the core entities with an association probability greater than or equal to a preset association probability threshold as valid relationship edges; Traverse each node in the basic state graph and calculate the degree value. Call the semantic analysis algorithm to calculate the semantic completion degree of the terminal node with a degree value of zero. Mark the node with an out-degree value of zero and a semantic completion degree lower than the semantic completion threshold as the dangling node with the unresolved attribute.

3. The intelligent grading and classroom simulation interactive platform for reading and writing continuation as described in claim 1, characterized in that, The fault-tolerant analysis module is specifically used for: A morphological filtering operation is performed on the handwritten text image, and the character nodes are extracted using an optical character recognition model. After removing character nodes whose recognition confidence weight is lower than a preset threshold, the continuation state map is constructed. Using the suspended node as the logical reference point for comparison, the subgraph isomorphic matching algorithm is used to calculate the node mapping and edge mapping relationship between the reference substructure of the basic state graph and the continuation state graph. Based on the nodes and relation edges with missing matching items, the topological breakage degree of the reference substructure is calculated by combining the node weight coefficient and the relation edge weight coefficient, and the topological breakage feature is generated by dimensional combination according to the node sequence.

4. The intelligent grading and classroom simulation interactive platform for reading and writing continuation as described in claim 1, characterized in that, The feature clustering module is specifically used for: The principal component analysis algorithm is invoked to perform an orthogonal transformation on the multidimensional defect feature matrix to extract principal component features and obtain a dimension-reduced feature matrix. Extract the feature vector corresponding to the execution object in the reduced-dimensional feature matrix, and use the K-means clustering algorithm to divide the execution object into defect type clusters; The defect complementarity is obtained by calculating the vector inner product and modulus between the feature vectors of the execution objects in different defect type clusters. The execution objects with defect complementarity greater than or equal to the complementarity threshold are paired to generate the grouping strategy data.

5. The intelligent grading and classroom simulation interactive platform for reading and writing continuation as described in claim 1, characterized in that, The dynamic layout module is specifically used for: The grouping strategy data is parsed to obtain the number of execution objects and the object number sequence, and the preset available writing area is divided into multiple independently allocated writing areas according to the proportion; Extract the logical starting point coordinates of the suspended node, combine them with the collaboration order of the execution objects to generate a logical timing code diagram, and insert it into a visual layout file; The width of the isolation buffer is calculated by using the preset minimum optical segmentation spacing, boundary fault tolerance coefficient, and standard deviation of the writing fluctuation of the execution object. The two-dimensional boundary of the writing area, the logical timing encoding graphic, and the isolation buffer layer are superimposed to generate the typesetting instruction.

6. The intelligent grading and classroom simulation interactive platform for reading and writing continuation as described in claim 1, characterized in that, The optical reconstruction module is specifically used for: A two-dimensional image matrix of the cooperative physical medium is obtained, and the visual anchor point pixel coordinates of the two-dimensional image matrix are located using an edge detection algorithm; The pixel coordinates of the visual anchor point are compared with the standard coordinates inside the system, and a system of linear equations is constructed to solve for the rotation, translation and scaling parameters of the affine transformation matrix. The two-dimensional image matrix is ​​resampled using a bilinear interpolation algorithm based on the affine transformation matrix, and a corrected image matrix is ​​output.

7. The intelligent grading and classroom simulation interactive platform for reading and writing continuation as described in claim 6, characterized in that, The optical reconstruction module is also used for: The corrected image matrix is ​​divided into multiple independently distributed sub-region images based on the coordinate positions of the isolation buffer band; The logical temporal coding pattern in the corrected image matrix is ​​parsed to obtain the logical splicing order, and the sub-region images are arranged in one dimension in memory according to the logical splicing order; The arranged sub-region images are sequentially input into the optical character recognition model to extract the multi-region handwritten text data.

8. The intelligent grading and classroom simulation interactive platform for reading and writing continuation as described in claim 1, characterized in that, The verification closed-loop module is specifically used for: In the memory space, data storage nodes are allocated for the handwritten text data in the multi-region area, and the logical splicing order is parsed to establish directed pointers for the data; Using the directed pointers, the multi-region handwritten text data is memory-addressed and connected end-to-end according to the logical topology direction; The multi-region handwritten text data is instantiated into the interactive text tree by connecting the beginning and end.

9. The intelligent grading and classroom simulation interactive platform for reading and writing continuation as described in claim 8, characterized in that, The verification closed-loop module is specifically used to calculate the total penalty value as follows: The text feature extraction model is invoked to calculate the cosine similarity between the text nodes contained in the interactive text tree and the constraint rules contained in the basic state graph, and a positive penalty term for content deviation is calculated by combining the positive penalty weight coefficient. Calculate the cosine similarity between the text nodes contained in the interactive text tree and the necessary nodes contained in the basic state graph, and calculate the reverse penalty term for missing plot by combining the reverse penalty weight coefficient. The positive penalty term and the negative penalty term are summed to generate the total penalty value.

10. The intelligent grading and classroom simulation interactive platform for reading and writing continuation as described in claim 1, characterized in that, The verification closed-loop module is specifically used for outputting evaluation data as follows: Obtain a preset initial baseline score, subtract the product of the total penalty value and the preset score conversion coefficient from the initial baseline score, and calculate the final evaluation score of the corresponding execution object; The interactive text tree, the total penalty value, and the final evaluation score are encapsulated as structured data and stored in the system database.

Images