A method and system for intelligent writing and correction

Abstract

The application belongs to the field of intelligent human-computer interaction, and relates to a kind of intelligent writing and correcting method and system, comprising: utilizing flexible pressure sensor matrix and coordinate positioning system, the multidimensional data of sensing input pen writing, and constructing four-dimensional writing characteristic space;Adopt three-level correcting system, the logical reasoning verification of answer is carried out through depth semantic understanding model;The three-level correcting system successively includes OCR character shape matching, semantic rectification based on large language model and intelligent correction based on process reward model;Cognitive diagnosis model is introduced, and the degree of hesitation of answering is judged by analyzing answering index;Fusion question solving information, the knowledge point of question is identified using large model, and according to the question answering condition, the weak point of knowledge mastery is inferred;With the intelligent writing capture technology based on multi-modal sensor fusion, the writing information is accurately collected, and the intelligent correcting integrated method is constructed combined with cognitive diagnosis model, the efficient identification and accurate correction of writing content are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent human-computer interaction, and specifically discloses a method and system for intelligent writing and correction. Background Technology

[0002] Traditional paper-based digitization processes suffer from significant technical bottlenecks: optical scanning-based document conversion requires manual intervention, with single-page processing latency typically ranging from 15 to 30 seconds, and OCR recognition accuracy is easily affected by handwriting clarity. The field of educational assessment faces three major technical challenges: First, reliance on electromagnetic induction electronic pens or specialized pressure-sensitive paper results in high hardware costs, increasing the barrier to entry; second, traditional subjective question grading primarily relies on keyword matching, lacking the ability to deeply analyze logical chains, leading to poor results in in-depth analysis of essay questions; third, existing technologies have fundamental deficiencies in the cognitive diagnostic dimension, failing to establish a mapping relationship between writing dynamics and cognitive processes, making personalized learning intervention difficult.

[0003] In view of this, this application proposes a method and system for intelligent writing and correction. This system and method utilize intelligent writing capture technology based on multimodal sensor fusion to accurately collect writing information and combine it with a cognitive diagnostic model to construct an integrated intelligent correction method, achieving efficient recognition and accurate correction of written content. It has broad application prospects and is suitable for various scenarios such as educational assessment, remote collaboration, and intelligent office work, effectively improving the intelligence level and efficiency of writing instruction and collaborative office work. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent writing and correction method, which solves the problem of accurately collecting writing information based on multimodal sensor fusion intelligent writing capture technology, and combining it with a cognitive diagnostic model to construct an integrated intelligent correction method, thereby achieving efficient recognition and accurate correction of written content; the specific solution is as follows:

[0005] The system includes: S1: Utilizing a flexible pressure sensor matrix and coordinate positioning system, it senses multi-dimensional data during pen writing and constructs a four-dimensional writing feature space; the multi-dimensional data includes pressure changes, pen tip trajectory, and writing speed; the four-dimensional writing feature space includes pressure P, trajectory F, speed V, and time T; S2: Employing a three-level grading system, it uses a deep semantic understanding model to logically verify the answers; the three-level grading system includes OCR character matching, semantic correction based on a large language model, and intelligent grading based on a process reward model; S3: Introducing a cognitive diagnostic model to determine the degree of hesitation in answering questions by analyzing answer indicators; the answer indicators include pressure change rate and pen movement fluctuation; S4: Integrating question answer information, it uses a large model to identify the knowledge points to which the questions belong, and infers weak points in knowledge mastery based on the question answering situation.

[0006] Furthermore, S2 includes: S21: performing image segmentation and character recognition on the handwritten image to obtain the handwritten answer content; S22: constructing a structured semantic input template and inputting the structured semantic input template into the large language model for semantic correction to obtain the corrected answer content; the structured semantic input template includes question stem information units, answer content units, and model instruction units; S23: modeling the reasoning steps in the candidate's answering process through a logical reasoning model, and evaluating the logical rationality of each reasoning step to obtain a step score.

[0007] Furthermore, S23 includes: S231: Based on the scoring rules and standard answers, decompose and output scoring points using a general reasoning model; S232: Train the general reasoning model based on the first training dataset and the second training dataset to obtain an intelligent grading model; S233: Input key steps into the intelligent grading model, independently judge each standard step s_n, and determine the step judgment result; the step judgment result is used to indicate whether the user's answer A_u contains or satisfies the core points of the step; S234: Calculate the score for each step and the final score based on the step judgment result and the normalized weight value.

[0008] Furthermore, S3 includes: S31: When the user answers questions, real-time collection of writing pressure signals and pen speed signals, and determination of the examinee's mastery of knowledge points based on the writing pressure signals and pen speed signals; S32: Acquiring the question stem text content of the questions to be evaluated, and inputting the question stem text content into the intelligent grading model, inferring and outputting key knowledge point tags based on the semantic features and language structure of the question stem; and encoding the key knowledge point tags to construct the mapping relationship between the question stem and knowledge points; S33: Dividing the third training dataset into multiple classes, constructing multi-class training sets, and training the cognitive diagnostic model through multi-class training sets.

[0009] Furthermore, S31 includes: determining the real-time pressure change rate based on the writing pressure signal; obtaining the average pressure change rate based on the real-time pressure change rate; determining the pen speed fluctuation amplitude based on the pen speed signal; calculating the standard deviation of the pen speed fluctuation amplitude; determining the mastery and proficiency of knowledge points during the answering process based on the standard deviation; and obtaining the answer feature vector.

[0010] Furthermore, S33 includes: S331: Training the cognitive diagnostic model using a third training dataset; the third training dataset includes historical answer results and knowledge point labels; the training loss function is to minimize the prediction error of knowledge point mastery; S332: The expert scoring system collects scoring features for each question in real time, inputs them into the trained cognitive diagnostic model, and outputs the mastery proficiency of each knowledge point.

[0011] Furthermore, step S231 includes: constructing prompt words and extracting key steps and corresponding scores based on the prompt words; inputting the key steps and corresponding scores into a general reasoning model for processing, outputting a set of key steps and the scores corresponding to the key steps; normalizing the scores to obtain normalized weight values; and obtaining a standardized scoring template based on the question, standard answer, set of key steps, and normalized weight values.

[0012] This invention also provides an intelligent writing and grading system, including a writing module, a transmission module, and a processing module; the writing module is used to capture the writing content; the writing content includes multi-dimensional data during writing; the transmission module is used to send the writing content to a server; the processing module is used to automatically score objective questions based on the writing content, intelligently analyze and grade subjective questions, and provide problem-solving ideas and feedback suggestions.

[0013] Furthermore, the writing module includes a pad, an input pen, and writing paper; the pad is equipped with a pressure sensor and a coordinate positioning system for capturing the writing content; the input pen is a stylus with a magnetoelectric sensor.

[0014] Furthermore, the processing module includes a writing feature construction unit, a reasoning verification unit, an analysis unit, and an inference unit. The writing feature construction unit uses a flexible pressure sensor matrix and a coordinate positioning system to sense multi-dimensional data during pen writing and construct a four-dimensional writing feature space. The multi-dimensional data includes pressure changes, pen tip trajectory, and writing speed. The four-dimensional writing feature space includes pressure P, trajectory F, speed V, and time T. The reasoning verification unit uses a three-level grading system to perform logical reasoning verification on the answers through a deep semantic understanding model. The three-level grading system includes OCR character matching, semantic correction based on a large language model, and intelligent grading based on a process reward model. The analysis unit introduces a cognitive diagnostic model to determine the degree of hesitation in answering questions by analyzing answer indicators. The answer indicators include pressure change rate and pen movement fluctuation. The inference unit integrates the question answer information, uses a large model to identify the knowledge points to which the questions belong, and infers the weak points in knowledge mastery based on the question answer situation.

[0015] The present invention has the following advantages and beneficial effects:

[0016] This invention has broad application prospects and is applicable to various scenarios such as educational assessment, remote collaboration, and smart office. It can effectively improve the intelligence level and efficiency of writing instruction, collaborative office work, and other tasks. Attached Figure Description

[0017] Figure 1 An exemplary flowchart of an intelligent writing and correction method provided by the present invention;

[0018] Figure 2This is an exemplary schematic diagram of an intelligent writing and correction system according to the present invention;

[0019] Figure 3 This is an exemplary flowchart of multimodal data processing for an intelligent writing and correction system according to the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] like Figure 1 As shown, the multimodal data processing flow in this invention includes the following:

[0022] Step S1: By utilizing the flexible pressure sensor matrix and coordinate positioning system in the pad under the paper, the input pen is used to sense the content being written on the paper. This writing content can include multi-dimensional data such as pressure changes, pen tip trajectory, and writing speed, constructing a four-dimensional writing feature space (P, F, V, T) of "pressure P-trajectory F-speed V-time T". The pad sends the sensed user input to a remote cloud server. The remote cloud server then acquires and stores this input content.

[0023] Step S2: A three-tiered grading system is adopted: "OCR character matching → semantic correction based on a large language model → intelligent grading based on a process reward model." This system uses a deep semantic understanding model to perform logical reasoning verification on the answers, ensuring the accuracy and in-depth analytical capabilities of subjective question grading. Specifically, this includes:

[0024] Step S21: Segment the handwritten image to obtain the font text, and perform OCR character matching on the font text to obtain the handwritten answer content. Image segmentation is implemented using a CNN-BiLSTM-based handwritten character recognition model.

[0025] A first dataset for training is obtained, consisting of a large number of labeled handwritten images and their corresponding text (the font text corresponding to the handwritten images). A Convolutional Neural Network (CNN) is used to extract features from the handwritten images, and a Bidirectional Long Short-Term Memory (BiLSTM) network is used to encode the extracted temporal features, achieving accurate recognition of complex handwriting. This model combines spatial and temporal features, maintaining a high recognition rate under diverse handwriting conditions. Handwritten images refer to images of handwritten characters, such as student answers, homework, and essays on exam papers. Corresponding text refers to the text corresponding to the student's handwritten images. The corresponding text is obtained by extracting and recognizing characters from the handwritten images. The input to the CNN is the handwritten image, and the output is a multi-dimensional vector, such as a 3x768 matrix. The input to the BiLSTM is temporal features, used to adjust the output from the CNN network and output a character sequence. For example, the handwritten text image can be segmented into individual character images, and then the individual character images can be recognized to identify which script the character belongs to. The CNN-BiLSTM model was trained using the first dataset to obtain a handwritten character recognition model.

[0026] Step S22: Construct a structured semantic input template. Based on the semantic correction mechanism of the large language model, input the structured semantic input template into the large language model for semantic correction to obtain the corrected answer content.

[0027] After recognizing the handwritten answer content, a general open-source LLM (Large Language Model; e.g., Deepseek, Qwen3) is introduced to perform semantic correction, combining the question stem and contextual information. By performing semantic understanding and reasoning on the recognized text (i.e., the handwritten answer content), semantic deviations caused by non-standard handwriting or recognition errors are corrected, improving the ability to reconstruct the examinee's true intentions. The inputs to the LLM include the handwritten answer content, the question stem, and contextual information. Contextual information can include the subject and question type of the exam paper. Contextual information can be obtained through layout analysis and OCR recognition. The question stem can be obtained through OCR text recognition and can include all question stem content. For example, for a specific answer, such as options ABC, the aforementioned handwritten answer content, question stem, and contextual information are all input into the large language model, and prompt words are written, allowing the large model to output corresponding structured information. For example, if the handwritten answer content is recognized as "O" in ABCD, it will be corrected to "D".

[0028] A structured semantic input template is adopted, which consists of three parts: a question unit (Q), an answer unit (A), and an instruction unit (I), denoted as {QAI}. The specific structure is as follows: The question unit (Q) provides the original question text, including the question stem, which forms the basic semantic background for the model's understanding of the context. The answer unit (A) contains the original handwritten answer text obtained from OCR, including the handwritten answer content, serving as the semantic correction content for model inference. This text may contain character recognition errors, grammatical anomalies, or incoherent contexts, requiring high-level semantic reconstruction by the model; for example, the handwritten answer content obtained from OCR is input into the large model, which returns a corrected document. The instruction unit (I) provides task-oriented information, including contextual information, in the form of natural language or templated instructions, guiding the model to perform specified operations based on its understanding of the question stem and answer. Using a structured semantic input template can improve the accuracy of large language models in correcting biases, thus accurately describing the task and improving accuracy. Finally, the prompt will be in the form of: "Please correct the semantic error in the answer {A} below based on the question stem: {Q}. Please maintain the original meaning and correct the recognition error."

[0029] Step S23: Intelligent grading module based on process reward model.

[0030] To ensure the logical correctness of the answers, a process-reward mechanism-based logical reasoning model is introduced. This model models the reasoning steps in the test-taker's answer process, evaluates the logical rationality of each step, and uses it as one of the important criteria for scoring. The reward mechanism is based on reinforcement learning, providing positive incentives for reasonable reasoning processes to promote the model's accurate judgment of the completeness and correctness of the reasoning chain. Specifically, the question content Q, user answer text Au, scoring details S, and standard answer A are structurally decomposed and standardized to form a unified scoring input format, which serves as the input to the large language model. The logical reasoning model can be a post-trained instructive LLM, such as the qwen2.5-7b instruction; the input is text, and the output is text.

[0031] Step S231: Based on the scoring rules and standard answers, use a general reasoning model (such as DeepSeek R1, QW3, etc.) to break down and output the scoring points. This includes the following:

[0032] 1) Construct prompts (including task description instructions, such as "Based on question Q, scoring rules S, standard answer A, extract key steps and corresponding scores").

[0033] 2) Input the question Q, scoring details S, and standard answer A into the general reasoning model for processing.

[0034] 3) The set of key steps output by the large inference model and the corresponding scores for key steps. In C, the score for each item is extracted from the scoring criteria and normalized to obtain the normalized weight value. The normalization formula is:

[0035] ;

[0036] in, Represents the normalized weight of the i-th inference step; This represents the score of the i-th reasoning step (i.e., the step involving key knowledge points); Let represent the score of the j-th reasoning step; i and j both represent the reasoning step variables. n represents the total number of reasoning steps (i.e., the steps for key knowledge points).

[0037] The final standardized scoring template is as follows:

[0038] ;

[0039] For example, the original scoring rules might be: 1 point for writing out Newton's First Law, 2 points for calculating the ball's speed, and full marks for a correct answer. If the question is worth 5 points, then simply writing the answer would actually earn 2 points. A standardized process score representation is output using a large model. Finally, a standard scoring template (i.e., a formatted version of the scoring rules) is obtained.

[0040] Step S232: Construct a large-scale intelligent grading model. Based on the general inference model, use the first training dataset and the second training dataset to obtain the large-scale intelligent grading model through supervised fine-tuning (SFT). The model was fine-tuned under supervision based on two downstream subtasks: knowledge point extraction and multi-step scoring. The specific process includes:

[0041] Task 1: Fine-tuning of the knowledge point extraction task:

[0042] Objective: To automatically extract the core knowledge points from the test question text to provide semantic reference for subsequent scoring.

[0043] Input: Question Q;

[0044] Output: Collection of knowledge points ;

[0045] Data preparation: Constructing the first training dataset ,in These are actual test questions. This is a set of standard knowledge points marked by teaching and research experts; The variable representing the reasoning step, i.e., the first step. Each rating item; The variables represent the knowledge points in the actual test questions. These knowledge points can come from the course knowledge graph, the teaching syllabus, or the classification label system, and have a unified naming convention and semantic definition.

[0046] Model fine-tuning: Use a pre-trained large language model as the base model (such as qwen2.5-7b); perform supervised fine-tuning on the model (using the existing fine-tuning framework) to enable it to learn to abstract semantically related knowledge points from natural language prompts.

[0047] Task 2: Fine-tuning of the multi-step scoring task:

[0048] Objective: To build an intelligent judgment model for candidates' subjective answers based on the scoring criteria;

[0049] Input: Standardized scoring template and user response text ;

[0050] Output: Scoring results ;in, Indicates the relationship with the first The scoring results of the key knowledge points corresponding to each reasoning step. For example, a subjective question with a total of 15 points is broken down into 5 scoring steps. Y represents whether each of these 5 steps is scored. The score of each step is [2,4,1,4,4], Y=[0,1,1,0,1], so the final score is 4+1+4=9.

[0051] ;

[0052] Data preparation: Constructing a second training dataset ;

[0053] in, This represents the user's answer text for question j; This indicates the j-th question. Each reasoning step corresponds to a standard step; This indicates the j-th question. Normalized weights for each inference step; labels For manual scoring and labeling, mark whether the answer satisfies the condition of question j. Scoring criteria; This indicates the j-th question. The rating item (the first) (Score of each reasoning step); represents the reasoning steps for each question; j represents the question variable.

[0054] Model fine-tuning: A pre-trained large language model is used as the base model (such as qwen2.5-7b); supervised fine-tuning of the model is performed (using the existing fine-tuning framework); the input prompt words are organized in a structured way (QAS format) to improve the model's understanding of the scoring task;

[0055] Step S233: Utilize the intelligent grading model obtained in step S232 For each standard step Independent judgment, that is, the model judges the user's answer. Does it contain or satisfy the core requirements described in this step? The input is the standard step S for each step and the user's answer Au. This is a loop that checks only one key step at a time. For example, writing out Newton's first law earns 1 point, calculating the ball's speed earns 2 points, and a correct answer earns full marks. Three checks are required.

[0056] The decision function is:

[0057] ;

[0058] in, Indicates the first A decision function for each reasoning step; Indicates the first The standard steps corresponding to each reasoning step (i.e., the key steps).

[0059] Step S234: Combine the results of the previous steps with the weight values ​​in the scoring criteria to calculate the final score:

[0060] ;

[0061] The scoring system calls the large model in a loop. The system ultimately outputs the total score and detailed scores for each step, enabling learners to analyze the reasons for their mistakes and teachers to track their answer logic and knowledge mastery.

[0062] Step S3: Introduce a cognitive diagnostic model to determine the degree of hesitation in answering questions by analyzing the rate of change in stress and pen movement fluctuation indicators.

[0063] Step S4: Integrate the question and answer information, use a large model to identify the knowledge points to which the questions belong, and infer the weak points in knowledge based on the answers. This achieves a technological leap from "result feedback" to "process intervention".

[0064] Specifically, it includes:

[0065] Step S31: While the user is answering questions, the writing pressure signal and pen speed signal are collected in real time. The writing pressure signal P(t) represents the continuous function of the user's pressure changing with time during the question-answering process; the pen speed signal v(t) represents the trajectory of the pen tip's movement speed changing with time.

[0066] Using the real-time rate of pressure change as a measure of hesitation in answering questions:

[0067] ;

[0068] in, Indicates the degree of hesitation in answering the question; This indicates the difference in writing pressure signals; d represents the time difference; d represents the derivative sign. The rate of change of pressure is used to measure the degree of hesitation during the answering process. Significant fluctuations before and after key steps indicate that the user has cognitive uncertainty about that knowledge point.

[0069] Mean pressure change rate:

[0070] ;

[0071] in, dt represents the average pressure change rate; T represents a constant that controls the range of values; dt represents the derivative with respect to time; d* represents the derivative sign.

[0072] Fluctuation range of pen stroke speed:

[0073] ;

[0074] in, Indicates the fluctuation range of pen stroke speed; std indicates standard deviation; std indicates standard deviation calculation. This indicates the speed signal of pen strokes.

[0075] By calculating the standard deviation of pen stroke speed, we can determine the user's mastery of knowledge points during the answering process. Greater fluctuations indicate unstable mastery or repetitive thought processes. The final answer feature vector is obtained as follows:

[0076] ;

[0077] Where x represents the final response feature vector; m represents the total number of features; Represents the space of real numbers.

[0078] Step S32: Obtain the text content of the question stem of the user's question to be evaluated, and input the text into the intelligent grading model. Based on the semantic features and language structure of the question, the system automatically infers and outputs the core knowledge points tested in the question. The output is a set of key knowledge point tags. The tag can come directly from a pre-set knowledge point base, or be generated by standardizing knowledge point names through a model. The system encodes the key knowledge point tag and constructs a mapping relationship between questions and knowledge points.

[0079] Step S33: Divide the training samples of the third training dataset into multiple classes according to manual annotation or expert scoring system (a system used to annotate data through expert knowledge), construct a training set, train the cognitive diagnostic model CDM, and use the cognitive diagnostic model to identify knowledge gaps and generate personalized learning paths:

[0080] ;

[0081] in, Represents a cognitive diagnostic model; This represents user's answer data (including answer content, results, and time information). This represents the feature set extracted from the pressure change rate and velocity fluctuations. This represents the k-th key knowledge point in the set of knowledge points; Indicates the user's understanding of knowledge points The probability of mastering ; o represents the total number of key knowledge points. The answer content can be the score for each question (converted to 0-1). For example, a physics problem is worth 5 points, and if the score is 4 points, then 0.8 is the weighted average score for that question under the Newton's First Law knowledge point tag. The result represents the level of mastery of that knowledge point, ranging from 0-1. Time represents the time information for each question. If the answer is interrupted, the cognitive diagnostic model can determine that the test taker is not yet proficient in that knowledge point.

[0082] Step S331 Training phase: Use the labeled third training dataset (containing the user's historical answer results and knowledge point labels, where historical answer results refer to the candidate's past answer results) to train the cognitive diagnostic model. The optimization objective is to minimize the knowledge point mastery prediction error (such as cross-entropy or MSE).

[0083] S332 Reasoning Stage: In real-world use, the expert scoring system collects data in real time for each question. The input is fed into the trained CDM, and the output is a probability estimate of the mastery level of each knowledge point.

[0084] This invention also provides an intelligent writing and correction system, such as... Figure 2 As shown, it includes a writing module, a transmission module, and a processing module.

[0085] The writing module consists of an input pen 12 (such as a stylus with a magnetoelectric sensor), a pad 11 with a pressure sensor and coordinate positioning system, and plain paper 13, used to capture the written content. Figure 3 As shown, the pad under the paper senses the content written on the paper by the input pen through a pressure sensor and a coordinate positioning system. For example, a composite sensing scheme using a 1024×1024 flexible pressure sensor matrix (resolution 0.01N) fabricated with MEMS technology and a 60GHz millimeter-wave radar array (three-dimensional positioning accuracy 0.1mm).

[0086] Transmission module 15: such as Figure 3 As shown, the pad sends the user's handwriting data to a remote cloud server and provides real-time feedback, using various wireless communication protocols such as Bluetooth, WIFI, LTE, and 5G.

[0087] Processing Module: The processing module primarily consists of cloud server 14, possessing AI large-scale model processing capabilities. This includes AI large-scale models deployed on remote servers, automatically scoring objective questions, intelligently analyzing and grading subjective questions, and providing problem-solving strategies and feedback suggestions. Figure 3 As shown, the remote cloud server acquires and stores the user's handwritten content data. For exam paper information, it can be recognized and transmitted in one go. The question stem content is matched based on the exam paper coordinates. Objective questions can include multiple-choice and true / false questions, which can be directly recognized and matched with the correct answers. The problem-solving approach and feedback suggestions are standardized scoring items specified in the scoring rubric.

[0088] The processing module includes a writing feature construction unit, a reasoning verification unit, an analysis unit, and an inference unit;

[0089] The writing feature construction unit is used to sense multi-dimensional data when the input pen is writing using a flexible pressure sensor matrix and a coordinate positioning system, and to construct a four-dimensional writing feature space; the multi-dimensional data includes pressure changes, pen tip trajectory and writing speed; the four-dimensional writing feature space includes pressure P, trajectory F, speed V and time T.

[0090] The reasoning and verification unit is used to perform logical reasoning verification of the answer through a three-level grading system and a deep semantic understanding model. The three-level grading system includes OCR character matching, semantic correction based on a large language model, and intelligent grading based on a process reward model.

[0091] The analysis unit is used to introduce a cognitive diagnostic model to determine the degree of hesitation in answering questions by analyzing answer indicators, including the rate of change of pressure and pen movement fluctuations.

[0092] The inference unit is used to integrate the information from the answers to the questions, use a cognitive diagnostic model to identify the knowledge points to which the questions belong, and infer the weak points in knowledge mastery based on the answers to the questions, thereby generating a personalized learning path.

[0093] Example 1

[0094] In a smart education scenario, students use ordinary ballpoint pens to answer questions on standard answer sheets laid with sensor pads. The pads simultaneously collect pressure (0.01N resolution) and three-dimensional coordinates (0.1mm accuracy) data at a 200Hz sampling rate. Edge computing nodes perform real-time data denoising and feature extraction, transmitting the data to a cloud server via a 5G network (transmission latency <100ms). A cloud-based multimodal large model performs three levels of processing: recognizing character shapes, analyzing writing dynamics (pressure fluctuations, pen speed, etc.), understanding semantic logic, and performing intelligent grading. Especially for subjective questions, it generates multi-dimensional feedback reports that include logical chain analysis, error type labeling (such as "conceptual confusion," "reasoning jumps," etc.), and recommendations for similar questions. The teacher-side visualization system presents a heatmap of class learning and automatically generates differentiated teaching suggestions (such as specialized training programs for frequently missed knowledge points).

[0095] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent writing and correction, characterized in that, include: S1: Utilizing a flexible pressure sensor matrix and coordinate positioning system, multi-dimensional data during pen writing is sensed, and a four-dimensional writing feature space is constructed; The multidimensional data includes pressure changes, pen tip trajectory, and writing speed; the four-dimensional writing feature space includes pressure P, trajectory F, speed V, and time T. S2: A three-level grading system is adopted, which uses a deep semantic understanding model to perform logical reasoning verification on the answers; the three-level grading system includes OCR character matching, semantic correction based on a large language model, and intelligent grading based on a process reward model. S3: Introducing a cognitive diagnostic model to determine the degree of hesitation in answering questions by analyzing answer indicators; these indicators include the rate of change in stress and penmanship fluctuation; S3 includes: S31: While the user is answering questions, the writing pressure signal and pen speed signal are collected in real time, and based on these signals, the user's mastery of the knowledge points is determined; the writing pressure signal P(t) represents a continuous function of the user's pressure changing with time during the question-answering process; the pen speed signal v(t) represents the trajectory of the pen tip's movement speed changing with time; S31 includes: Based on writing pressure signals, the real-time pressure change rate is determined; this rate is used as the degree of hesitation in answering questions. Significant fluctuations before and after key steps indicate cognitive uncertainty for the user regarding that knowledge point. ； in, Indicates the degree of hesitation in answering the question; This indicates the difference in writing pressure signals; Indicates the time difference; d represents the derivative sign; Based on the real-time pressure change rate, the average pressure change rate is obtained: ； in, dt represents the rate of change of average pressure; T represents a constant; dt represents the derivative with respect to time; d* represents the derivative sign; This represents the writing pressure signal at time t; Determine the amplitude of pen stroke speed fluctuation based on the pen stroke speed signal: ； in, Indicates the fluctuation range of pen stroke speed; std indicates standard deviation; std indicates standard deviation calculation. This indicates the speed signal of pen strokes; Calculate the standard deviation of the fluctuation range of pen stroke speed, and determine the mastery and proficiency of knowledge points during the answering process based on the standard deviation, and obtain the answer feature vector; the greater the fluctuation, the more unstable the mastery or the repetitive thinking path. S32: Obtain the text content of the question stem of the question to be evaluated, and input the text content of the question stem into the intelligent grading model. Based on the semantic features and language structure of the question stem, reason and output key knowledge point tags; and encode the key knowledge point tags to construct the mapping relationship between the question stem and knowledge points. S33: Divide the third training dataset into multiple classes to construct a multi-class training set. Train the cognitive diagnostic model using the multi-class training set, and use the cognitive diagnostic model to identify knowledge gaps and generate personalized learning paths. ； in, Represents a cognitive diagnostic model; This represents user answer data; This represents the feature set extracted from the pressure change rate and velocity fluctuations. This represents the k-th key knowledge point in the set of knowledge points; Indicates the user's understanding of knowledge points S4: Integrate the information from the answers to the questions, use a large model to identify the knowledge points to which the questions belong, and infer the weak points in knowledge mastery based on the answers to the questions.

2. The intelligent writing and correction method according to claim 1, characterized in that, S2 include: S21: Perform image segmentation and character recognition on the handwritten image to obtain the handwritten answer content; S22: Construct a structured semantic input template and input the structured semantic input template into the large language model for semantic correction to obtain the corrected answer content; the structured semantic input template includes question stem information unit, answer content unit and model instruction unit; S23: Using a logical reasoning model, model the reasoning steps in the candidate's answer process, evaluate the logical rationality of each reasoning step, and obtain a step score.

3. The intelligent writing and correction method according to claim 2, characterized in that, S23 includes: S231: Based on the scoring rules and standard answers, use a general reasoning model to break down and output the scoring points; S232: Based on the first training dataset and the second training dataset, train the general reasoning model to obtain the intelligent grading model; S233: Input key steps into the intelligent grading model, and process each standard step... Independent judgment is made to determine the result of the step-by-step judgment; the result of the step-by-step judgment is used to represent the user's answer. Does it contain or satisfy the core requirements of this step? The input is the standard step S for each step and the user's answer Au; the process is to iterate through the steps, judging one key step at a time; the decision function is: ； in, Indicates the first A decision function for each reasoning step; Indicates the first The standard steps corresponding to each reasoning step; S234: Based on the step-by-step decision results and normalized weight values, calculate the score for each step and the final score: ; Where Score represents the total score output by the large model; i represents the inference step variable; and n represents the total number of inference steps. This represents the user's response text; This represents the normalized weight of the i-th inference step.

4. The intelligent writing and correction method according to claim 1, characterized in that, S33 includes: S331: The cognitive diagnostic model is trained using a third training dataset; the third training dataset includes historical answer results and knowledge point labels; the training loss function is to minimize the prediction error of knowledge point mastery. S332: The expert scoring system collects scoring features for each question in real time, inputs them into the trained cognitive diagnostic model, and outputs the mastery level of each knowledge point.

5. The intelligent writing and correction method according to claim 3, characterized in that, Step S231 includes: Construct prompts and extract key steps and their corresponding scores based on the prompts; Input the key steps and their corresponding scores into the general reasoning model for processing, and output a set of key steps and the corresponding scores for each key step. The scores are normalized to obtain normalized weight values; A standardized scoring template is obtained based on the question, standard answer, key steps set, and normalized weight values.

6. A system for intelligent writing and correction utilizing the intelligent writing and correction method as described in any one of claims 1-5, characterized in that, It includes a writing module, a transmission module, and a processing module; The writing module is used to capture the written content; the written content includes multi-dimensional data during the writing process. The transmission module is used to send the written content to the server; The processing module is used to automatically score objective questions based on the written content, intelligently analyze and grade subjective questions, and provide problem-solving ideas and feedback suggestions.

7. The intelligent writing and correction system according to claim 6, characterized in that, The writing module includes a pad, a pen, and writing paper; The pad is equipped with a pressure sensor and a coordinate positioning system to capture the written content; The input pen is a stylus with a magnetoelectric sensor.

8. The intelligent writing and correction system according to claim 6, characterized in that, The processing module includes a writing feature construction unit, a reasoning verification unit, an analysis unit, and an inference unit; The writing feature construction unit is used to sense multi-dimensional data during pen writing using a flexible pressure sensor matrix and a coordinate positioning system, and to construct a four-dimensional writing feature space; the multi-dimensional data includes pressure changes, pen tip trajectory, and writing speed; the four-dimensional writing feature space includes pressure P, trajectory F, speed V, and time T; The reasoning and verification unit is used to perform logical reasoning verification of the answer through a three-level grading system and a deep semantic understanding model. The three-level grading system includes OCR character matching, semantic correction based on a large language model, and intelligent grading based on a process reward model. The analysis unit is used to introduce a cognitive diagnostic model to determine the degree of hesitation in answering questions by analyzing answer indicators; these indicators include the rate of change in pressure and pen movement fluctuations. The inference unit is used to integrate the information from the answers to the questions, use a large model to identify the knowledge points to which the questions belong, and infer the weak points in knowledge based on the answers to the questions.

Images