Intelligent marking based on artificial intelligence using text analysis

Abstract

The application discloses an intelligent text analysis method for intelligent paper reading based on artificial intelligence, proposes an integrated solution for the business scene problems of handwriting content recognition, answer correction, error analysis and individualized tutoring in student homework, accurately identifies the stem and answer text in the homework picture through an image understanding module, generates a correction result by comparing a machine learning model with a standard answer, analyzes knowledge points through an error generation module for errors, creates similar exercise questions, and simultaneously generates detailed problem solving guidance text, and finally integrates into a feedback document containing error review and tutoring suggestions. Further, the model parameters are optimized through a dynamic adjustment mechanism to improve the correction accuracy. The application seamlessly connects homework correction and individualized tutoring, realizes full-process automation from recognition to feedback, significantly improves the education efficiency and learning pertinence, and provides accurate learning support for students.

Description

Technical Field

[0001] This invention relates to the field of educational technology, and in particular discloses an intelligent text analysis method for grading papers based on artificial intelligence. Background Technology

[0002] In recent years, the field of educational technology has demonstrated significant value in improving teaching efficiency and personalized learning, particularly in the area of ​​homework and test grading, where technological innovation is considered a key path to improving educational quality and reducing teachers' workload. Automation and intelligentization through technological means can not only optimize the allocation of teaching resources but also provide students with more precise learning support. However, despite the considerable attention this area has received, existing methods and tools still face numerous challenges in practical application, requiring breakthroughs.

[0003] Currently, homework and exam grading relies primarily on manual operation. While some technological solutions have attempted to introduce automation, these solutions often fall short when dealing with complex content. The problem isn't just the limited scope of the technology; existing tools struggle to adapt to diverse writing styles and expressions when handling varied questions and student answers, especially when identifying and understanding non-standardized content, often resulting in errors. These errors not only affect the accuracy of grading but also limit the precise analysis of students' learning progress.

[0004] A deeper technical challenge lies in accurately understanding and parsing the complex content of assignments or exams, especially those containing special symbols, graphics, or handwritten text. The diversity and unstructured nature of this content pose significant challenges to the system's extraction and judgment. The core issue focuses on the comprehensive ability to analyze image content, because without correctly identifying every detail in the questions and answers, subsequent grading and analysis are impossible. For example, in math exams, handwritten formulas or symbols might be misread by the system due to illegible handwriting or non-standard writing, leading to discrepancies between the graded results and the actual answers. This misjudgment directly affects the reliability of the entire grading process.

[0005] Therefore, accurately interpreting text, symbols, and expressions in images from diverse assignments and test papers, and ensuring the correctness of the grading results, has become a key issue that urgently needs to be addressed in the field of educational technology. Summary of the Invention

[0006] This invention provides an intelligent text analysis method for marking papers based on artificial intelligence, aiming to solve at least one of the defects existing in the above-mentioned prior art.

[0007] This invention relates to an intelligent text analysis method for marking examination papers based on artificial intelligence, comprising the following steps: S100. The image understanding module processes the homework images uploaded by the user. The image understanding module integrates optical character recognition function and uses an artificial intelligence model to perform semantic analysis and recognition on the handwritten content and mathematical symbols in the homework images to obtain the extracted question text and student answer text. S200. Based on the extracted question text and student answer text, a machine learning model is used to compare the student answer text with the preset standard answer to determine the correctness of the student answer text and obtain a list of grading results. S300. If the grading results list shows errors, the errors are classified and analyzed through the error generation module to identify the knowledge points involved in the errors. An artificial intelligence model is then used to create similar questions based on the knowledge points in error, resulting in a personalized set of practice questions for each student. S400. Based on the personalized practice question set, the tutoring module generates a step-by-step sequence of problem-solving steps, resulting in a detailed guidance text containing logical reasoning. S500 integrates the grading results list and detailed guidance text through an automated workflow, determines the final output report, and obtains a feedback document containing a review of incorrect questions and tutoring suggestions; S600. Based on the feedback document, a dynamic adjustment mechanism is used to iteratively update the parameters of the machine learning model and the artificial intelligence model to obtain the optimized grading accuracy.

[0008] Further, step S100 includes: S110. Obtain the original job image uploaded by the user and generate an enhanced job feature map containing clear foreground information; S120. Divide the printed question stem region and the handwritten handwriting region according to the enhanced task feature map to obtain the basic question stem character stream and the original handwritten symbol sequence; S130. Map the basic question stem character stream and the original handwritten symbol sequence into character semantic vectors to generate question stem semantic representation and answer semantic representation; S140. Analyze the semantic representation of the question stem and the semantic representation of the answer, and reorganize them to generate a standardized sequence of question stem text and student answer text, so as to achieve accurate extraction of question stem and answer from homework images.

[0009] Further, step S200 includes: S210. Obtain the preset standard answer sequence from the question text, and use a deep neural network in the machine learning model to encode the student's answer text and the preset standard answer sequence to obtain the hidden layer semantic tensor. S220. Calculate the attention weight distribution based on the hidden layer semantic tensor to identify key problem-solving steps in the student's answer text; S230. The key problem-solving steps are aligned with the standard deduction nodes in the preset standard answer sequence. If the matching confidence is lower than the preset threshold, it is judged as a logical deviation and a comprehensive evaluation score is calculated. S240. Based on the comprehensive evaluation score, generate structured evaluation data and output a list of grading results.

[0010] Further, step S300 includes: S310. Obtain the incorrect question data from the grading results list, extract the original question text and the student's incorrect answer text, and perform entity extraction on the original question text to obtain the question entity set; S320. Map the set of question entities to the subject knowledge graph to locate knowledge nodes. If the semantic deviation between the student's wrong answer text and the standard explanation of the knowledge node exceeds the preset deviation threshold, then identify the wrong knowledge points that need to be strengthened. S330. Construct a weak knowledge subgraph based on the knowledge points to be strengthened, and input it into the question generation model to obtain a sequence of candidate variant questions. S340. Filter the candidate variation question sequence and arrange it according to the difficulty gradient to obtain a personalized practice question set for students.

[0011] Further, step S400 includes: S410. Obtain the structural features of each question in the personalized practice question set, and identify the starting and ending points of the implicit logical reasoning chain in the questions; S420. Construct a reasoning dependency graph based on the starting and ending points of the logical reasoning chain, and insert detailed problem-solving step nodes into the reasoning dependency graph; S430. Traverse the problem-solving steps and nodes to generate a thought-guiding path with causal relationships. Input the thought-guiding path into the natural language generation model for semantic and logical mapping to obtain a detailed thought-guiding text containing logical reasoning. The detailed thought-guiding text is generated by the natural language generation model after transforming the abstract reasoning process.

[0012] Further, step S500 includes: S510. Obtain the error type codes and thought breakpoint location data from the detailed thought process guidance text in the correction results list. S520. Based on the error type code and the data of the thought breakpoint location, execute the integration mapping rules to associate and match the error correction logic in the detailed thought guidance text with the correction result list. S530. Generate tutoring suggestions based on error correction logic and construct error review content blocks; S540. Fill the tutoring suggestions module and the wrong question review content block into the document layout template and render it to obtain a feedback document containing wrong question review and tutoring suggestions.

[0013] Further, step S600 includes: S610. Parse the feedback document to obtain the error correction logic text, and convert the error correction logic text into a standardized semantic feature vector; S620. Calculate the Euclidean distance between the semantic feature vector and the original output vector, and determine the correction deviation value based on the Euclidean distance. S630. Calculate the gradient descent direction based on the correction deviation value, and perform weighted correction on the weight matrix in the model parameter set of the machine learning model and the artificial intelligence model in combination with the gradient descent direction. S640. Use the modified weight matrix to perform secondary inference on the validation set data. When the inference result meets the convergence baseline, fix the parameters to obtain the optimized correction accuracy.

[0014] The beneficial effects achieved by this invention are as follows: This invention provides an AI-based intelligent text analysis method for marking assignments. Addressing the challenges of handwritten content recognition, answer correction, error analysis, and personalized tutoring in student assignments, it proposes an integrated solution. Through an image understanding module, it accurately identifies the question stem and answer text in assignment images. A machine learning model compares the results with standard answers to generate corrections. For errors, an error generation module analyzes the relevant knowledge points, creates similar practice questions, and generates detailed solution guidance text. Finally, these are integrated into a feedback document containing a review of errors and tutoring suggestions. Furthermore, a dynamic adjustment mechanism optimizes model parameters to improve correction accuracy. Its core lies in seamlessly integrating assignment correction with personalized tutoring, achieving full automation from recognition to feedback, significantly improving educational efficiency and learning relevance, and providing students with precise learning support. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating an embodiment of the intelligent text analysis method for marking papers based on artificial intelligence according to the present invention. Detailed Implementation

[0016] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0017] like Figure 1 As shown, the first embodiment of the present invention proposes a text analysis method for intelligent marking based on artificial intelligence, including the following steps: Step S100: The image understanding module processes the homework images uploaded by the user. The image understanding module integrates optical character recognition function and uses an artificial intelligence model to perform semantic analysis and recognition on the handwritten content and mathematical symbols in the homework images to obtain the extracted question text and student answer text.

[0018] Using student-uploaded homework images (including scanned copies and photos of handwritten assignments, supporting single-page / multi-page and mixed question types) as the processing object, the image understanding module, which integrates Optical Character Recognition (OCR) functionality, completes accurate image-to-text conversion and semantic parsing. 1. Image preprocessing: First, perform noise reduction, tilt correction, edge cropping, and sharpness enhancement on the work images to eliminate the effects of uneven lighting, shooting angle deviation, background interference, etc., and ensure that the text area is clear and recognizable. 2. Character and Symbol Recognition: Employing deep learning-driven OCR models (such as CRNN (Convolutional Recurrent Neural Network) and Transformer-based OCR (Transformer-based Optical Character Recognition) models), it accurately recognizes handwritten Chinese / English text, numbers, mathematical symbols (such as formulas, square roots, integral symbols, geometric figure annotations), punctuation marks, etc. in images, and supports cursive writing and adaptation to different writing styles. 3. Semantic Analysis and Text Extraction: The identified character sequences are semantically analyzed using a Natural Language Processing (NLP) model to distinguish between the question stem text and the student answer text. Based on format features (such as question stem number and question identifier) ​​and semantic logic (such as keywords such as "ask", "solve", and "answer"), independent question stem units (including question number, question content, known conditions, and question requirements) and corresponding student answer units (including solution steps and final answer) are segmented and extracted. Finally, a structured "question stem text set" and "student answer text set" are output to ensure that the text and the question correspond one-to-one without confusion or omission.

[0019] Step S200: Based on the extracted question text and student answer text, a machine learning model is used to compare the student answer text with the preset standard answer to determine the correctness of the student answer text and obtain a list of grading results.

[0020] Using the question stem text and student answer text extracted in step S100 as input, and combined with a pre-set standard answer database (including standard answers for objective questions, scoring criteria for subjective questions, and key points of problem-solving steps), a machine learning model adapted to the characteristics of the question type is used to perform intelligent comparison and correctness judgment: 1. Model Selection and Adaptation: For objective questions (multiple choice, fill-in-the-blank, true / false), precise matching algorithms (such as string similarity comparison and hash matching) are used to directly determine the correctness of the answer; for subjective questions (short answer, essay, calculation, proof), NLP (Natural Language Processing) semantic understanding models (such as BERT (Bidirectional Encoder Representations from Transformers) and RoBERTa (Robustly Optimized BERT Pretraining Approach)) are used to fine-tune the model, comparing the answer with the scoring criteria from dimensions such as semantic consistency, logical completeness, key point coverage, and step standardization, and quantifying the score and the conclusion of correctness.

[0021] 2. Grading Logic Execution: For each question, the student's answer text is compared point by point with the corresponding standard answer / grading criteria. Correct questions, incorrect questions, and partially correct questions (subjective questions) are marked. The location of the error (such as a certain solution step or core argument) and the type of error (such as conceptual error, calculation error, or logical flaw) are recorded. Finally, the results are integrated into a "grading result list"—containing core information such as question number, question summary, original student answer, correct or incorrect conclusion, score, error marking (applicable to incorrect questions), and scoring basis, ensuring that the grading results are traceable and explainable.

[0022] Step S300: If the grading result list shows errors, the errors are classified and analyzed through the error generation module to identify the knowledge points involved in the errors. An artificial intelligence model is then used to create similar questions based on the knowledge points in error, resulting in a personalized set of practice questions for each student.

[0023] If there are incorrect questions in the grading results list, the incorrect question generation module will be activated to perform in-depth classification analysis of the errors and generate targeted exercises: 1. Error Classification Analysis: Machine learning classification models (such as decision trees and random forests) are used to classify error questions in multiple dimensions. First, the core knowledge points corresponding to the errors are identified (such as "solving quadratic equations in one variable" in mathematics, "correcting grammatical errors" in Chinese, and "using the past perfect tense" in English). Then, the errors are classified a second time according to the error type (conceptual confusion, calculation error, logical confusion, unclear understanding of the question, non-standard expression, etc.) and difficulty level, forming "error knowledge point tags" and "error type tags". 2. Personalized Exercise Creation: Based on error knowledge point tags, AI generation models (such as GPT (Generative Pre-trained Transformer) series models and domain-specific exercise generation models) are invoked. Combining the student's error type and learning level (derived through the difficulty of the wrong questions and answering behavior), similar questions are created—ensuring a high degree of matching between the exercises and the error knowledge points, while also adapting to the student's learning ability (e.g., reducing the difficulty of exercises for weak foundations, and increasing the extension for advanced errors). Finally, these are integrated into a "personalized exercise set," which includes the exercise stem, question type, difficulty level, and corresponding knowledge point tags. The number of questions is dynamically adjusted according to the number of errors (e.g., 3-5 questions / error knowledge point).

[0024] Step S400: Based on the personalized practice question set, the tutoring module generates a step-by-step sequence of problem-solving steps, resulting in a detailed guidance text containing logical reasoning.

[0025] Based on personalized practice sets, the tutoring module generates a "step-by-step solution sequence" and a "detailed thought process guide" for each exercise, achieving a "teaching a man to fish" tutoring effect: 1. Problem-solving step sequence generation: Using logical reasoning models (such as knowledge graph-based reasoning models and symbolic computing engines), detailed step-by-step problem-solving steps are generated according to subject-specific problem-solving standards. These steps include analysis of known conditions, problem-solving approaches, application of key formulas / theorems, intermediate derivation processes, and the final answer, ensuring that the steps are clear, logically rigorous, and in line with students' cognitive patterns. 2. Writing Guidelines for Problem Solving: Combining the problem-solving steps, write guidelines for problem solving in plain and easy-to-understand natural language—including core problem-solving ideas (such as mathematical "combination of number and shape" and "classification and discussion"), reminders of common mistakes (related to students' previous error types), extensions of knowledge points (such as formula derivation and the application scenarios of theorems), and summaries of problem-solving techniques (such as quick problem-solving methods and the application of the elimination method), forming a "detailed guideline for problem solving" that includes the logical reasoning process, avoiding the problem of only providing answers without process guidance.

[0026] Step S500: Integrate the grading results list and detailed guidance text through an automated workflow, determine the final output report, and obtain a feedback document containing a review of incorrect questions and tutoring suggestions.

[0027] By automating the workflow, the grading result list from step S200 is systematically integrated with the detailed guidance text from step S400 to generate a structured and visual "final output report" (i.e., feedback document). 1. Integrated Logic: The content is organized according to the structure of "Overall Evaluation - Review of Wrong Answers - Personalized Tutoring - Learning Suggestions"—Overall evaluation includes homework scores, accuracy rate, strengths, and main weaknesses; Review of Wrong Answers summarizes all wrong questions, linking error annotations, knowledge point tags, and error types; Personalized Tutoring corresponds to personalized practice question sets and provides solution steps and thought process guidance for each question; Learning Suggestions provide targeted learning plans based on the wrong answers (such as key review chapters and adjustments to learning methods).

[0028] 2. Output format: Supports multiple formats (PDF, Word, web version), with clear layout and highlighted key content (such as wrong questions and key steps) for easy viewing by students and teachers. It also supports exporting and sharing (such as sending to student accounts or teacher management backend).

[0029] Step S600: Based on the feedback document, the parameters of the machine learning model and the artificial intelligence model are iteratively updated using a dynamic adjustment mechanism to obtain the optimized grading accuracy.

[0030] Using the data in the feedback documents (including grading results, student feedback on incorrect answers, and manual corrections (if any)) as incremental training data, the parameters of the machine learning models (grading model, classification model) and artificial intelligence models (OCR model, exercise generation model, and thought process guidance model) in the system are iteratively updated through a dynamic adjustment mechanism: 1. Iterative Logic: If a grading deviation is found (such as a manually corrected erroneous grading case), the case is included in the model training set, and the feature weights and decision boundaries of the model are adjusted; if the generated exercises do not match the needs of students (such as feedback that the exercises are too difficult / too easy), the difficulty adaptation parameters of the generated model are optimized; if there are errors in OCR recognition, samples with corresponding handwriting styles and symbol types are added to improve the recognition accuracy. 2. Optimization Goals: Continuously improve the model's core performance indicators—grading accuracy (target ≥ 95%), OCR recognition accuracy (target ≥ 98%), exercise generation adaptability, and clarity of thought guidance—to ensure that the system is continuously optimized as usage data accumulates, adapting to more writing styles, question types, and teaching scenarios.

[0031] Furthermore, the text analysis method for intelligent marking based on artificial intelligence provided in this embodiment includes step S100 as follows: Step S110: Obtain the original job image uploaded by the user and generate an enhanced job feature map containing clear foreground information.

[0032] The pixel values ​​of the enhanced job feature map are obtained using the following formula: (1) In formula (1), The enhanced job feature map is represented in coordinates. Pixel value at that location, This represents the pixel value at this location in the original job image uploaded by the user. The weight coefficients represent the weights of the original image. This represents the weighting coefficients of the enhancement filter. Indicates the number of Gaussian filters. Indicates the first The weights of each filter, The standard deviation is expressed as Gaussian kernel function in coordinates The value at that location.

[0033] When processing user-uploaded original homework images, the first step is to generate an enhanced homework feature map containing clear foreground information using image enhancement techniques. Specifically, this enhancement process employs contrast adjustment and noise removal methods. For example, when processing a scanned image of a math homework assignment, the original image may have foreground blurring due to uneven lighting or paper wrinkles. In this case, histogram equalization is used to improve the overall contrast of the image, making printed text and handwritten handwriting clearer. Then, medium-level filtering is applied to remove random noise, ultimately resulting in a feature map with prominent foreground information. The edges of the printed text in the question stem are sharp, and the outlines of the handwritten answers are also enhanced, which facilitates subsequent region segmentation operations.

[0034] Step S120: Divide the printed question stem area and the handwritten handwriting area according to the enhanced task feature map to obtain the basic question stem character stream and the original handwritten symbol sequence.

[0035] The recognition result of the printed title area is obtained using the following formula: (2) In formula (2), Indicates position The results of identifying the printed title area are as follows: and Indicates the width and height of the feature window. The enhanced job feature map is represented in coordinates. eigenvalues ​​at that location Indicates the position of the character feature template The weighting coefficients.

[0036] (3) In formula (3), This represents the segmentation result of the handwritten handwriting area. These represent different region labels. Indicates the total number of pixels. Indicates the first Gradient strength at each pixel Indicates the first Feature identifiers of class regions Indicates the first Feature type of each pixel This represents the Dirac function.

[0037] The process of dividing the printed question stem region and the handwritten handwriting region based on the enhanced homework feature map is achieved using optical character recognition combined with region segmentation algorithms. For example, for a homework image containing multiple-choice questions, the system first identifies text blocks in the image using edge detection algorithms such as the Canny operator. Then, based on texture analysis, it distinguishes between printed and handwritten text. Printed text usually has a regular font structure and uniform spacing, while handwritten text shows irregular curves and pen pressure variations. After this segmentation, the basic question stem character stream is obtained, namely the character sequence of the printed question stem, such as "2+3=?", and the original handwritten symbol sequence, such as the student's handwritten "5". The key to this step is accurate segmentation to avoid handwriting interference with the question stem region, thereby ensuring the purity of the character stream.

[0038] Step S130: Map the basic question stem character stream and the original handwritten symbol sequence into character semantic vectors to generate question stem semantic representation and answer semantic representation.

[0039] The following formula describes the mapping process of converting a basic question stem character stream into a semantic vector: (4) In formula (4), Represents a character semantic vector. This represents the basic question stem character stream. Represents the embedding weight matrix. The function represents a character embedding mapping operation. This represents a mapping function.

[0040] The final semantic representation of the question stem is generated using the following formula: (5) In formula (5), This represents the semantic representation of the question stem. Indicates the sequence length. Indicates the first Attention weights for each position, Indicates the first There are hidden state vectors. Represents the global context weight. This represents the global context vector.

[0041] The following formula maps a sequence of handwritten symbols to a semantic representation of the answer: (6) In formula (6), This represents the semantic representation of the answer. Represents the original handwritten symbol sequence. Represents a vectorized dictionary. Functions represent mapping operations. This represents the tensor product operation. This represents the bias coefficient. This represents the bias vector.

[0042] The process of mapping the basic question stem character stream and the original handwritten symbol sequence into character semantic vectors involves the application of natural language processing models. Specifically, pre-trained word embedding models such as BERT (Bidirectional Encoder Representations from Transformers) are used to convert the character sequence into high-dimensional vectors. For example, for the question stem character stream "calculate the area of ​​a circle", the system decomposes it into words and maps them into a semantic vector space, capturing the mathematical semantic relationship between words such as "calculate" and "area". Similarly, for handwritten symbol sequences such as handwritten formulas, the system can map the characters into semantic vectors. The handwriting recognition is mapped to vectors to generate semantic representations of the question stem and the answer. These representations not only retain the surface form of the characters but also incorporate contextual semantics, thus providing richer semantic information for subsequent parsing. This mapping can effectively handle ambiguous characters and improve the overall extraction accuracy.

[0043] Step S140: Analyze the semantic representation of the question stem and the semantic representation of the answer, and reorganize them to generate a standardized sequence of question stem text and student answer text, so as to achieve accurate extraction of question stem and answer from the homework image.

[0044] The standardized sequence of question stem texts is derived using the following formula: (7) In formula (7), This represents a standardized sequence of question text extracted from the assignment image. Represents a candidate text sequence. This represents the input image of the assignment. Indicates the first Image feature extraction functions Indicates the first A text semantic encoding function, Indicates the first The weight coefficients of each feature, Indicates the total number of feature dimensions. This represents the semantic similarity calculation function. This represents the set of candidate text sequences.

[0045] The student answer text sequence is derived using the following formula: (8) In formula (8), This represents the recombined sequence of student answer text. Represents a sequence of candidate answers. Indicates the first A text tag, Indicates the first Image feature representation of each region This represents the total length of the answer sequence. Represents the conditional probability distribution function. This represents the set of candidate answer sequences.

[0046] The process of parsing the semantic representations of the question stem and the answer, and then recombining them to generate standardized question stem text sequences and student answer text sequences, is implemented using a sequence-to-sequence model such as Transformer. For example, when processing images of Chinese language homework, the system parses the semantic representation of the question stem, such as the vector representation of a reading comprehension question, identifies key sentence structures, and then recombines it into a standardized text sequence: "According to the article, what is the author's viewpoint?". Simultaneously, for the semantic representation of the answer, such as a handwritten paragraph, the system parses and generates a standardized sequence: "The author believes friendship is important." This process ensures the accuracy of extraction and brings the convenience of automatic grading in educational assessment, improving the efficiency and consistency of homework processing.

[0047] Furthermore, the text analysis method for intelligent marking based on artificial intelligence provided in this embodiment includes step S200: Step S210: Obtain the preset standard answer sequence of the question stem text, and use the deep neural network in the machine learning model to encode the student answer text and the preset standard answer sequence to obtain the hidden layer semantic tensor.

[0048] The generation process of semantic tensors in hidden layers of deep neural networks is described by the following formula: (9) In formula (9), Indicates the first The implicit semantic tensor of the layer, This represents the activation function. Indicates the first The weight matrix of the layer, This represents the output tensor of the previous layer. Indicates the first The bias vector of the layer.

[0049] The process of encoding student answer text using a deep neural network can be represented by the following formula: (10) In formula (10), The encoded vector representing the student's answer text. This represents a sequence of student answer texts. This represents the set of parameters of the encoder. This represents the encoding function of a deep neural network.

[0050] The process of encoding a preset standard answer sequence using a deep neural network can be represented by the following formula: (11) In formula (11), The encoding vector representing the pre-defined standard answer. This indicates a pre-defined sequence of standard answers. This represents the set of parameters of the encoder. This represents the encoding function of a deep neural network.

[0051] When processing math homework, the first step is to extract a pre-defined sequence of standard answers from the problem text. This involves retrieving the correct solution path from a database that matches the problem text. Specifically, for a problem involving calculating the area of ​​a triangle, the system extracts a sequence of standard answers such as "base times height divided by 2". This sequence includes step-by-step calculations to ensure a clear benchmark for subsequent comparisons. In one implementation, a deep neural network is used to encode the student's answer text and the pre-defined sequence of standard answers, resulting in a semantic tensor in the hidden layer. Here, the deep neural network can be a recurrent neural network such as LSTM, which transforms the text sequence into a high-dimensional tensor through layer-by-layer processing. For example, for the student's answer text "base is 5, height is 4, area is 10" and the standard sequence, the network first converts each word into an embedding vector, and then captures semantic relationships through hidden layers to form a tensor representing the overall meaning. This process starts from the input layer, encoding word by word and gradually building a context-dependent representation.

[0052] Step S220: Calculate the attention weight distribution based on the hidden layer semantic tensor to identify key problem-solving steps in the student's answer text.

[0053] The attention weight distribution is derived using the following formula: (12) In formula (12), Indicates the first The position of the student's answer to the first Attention weights for each hidden layer feature. Indicates the first The position and the first Correlation score between features This represents the total number of hidden layer features. This partially ensures that the sum of all attention weights is 1.

[0054] The key problem-solving steps in the student's answer text are derived using the following formula: (13) In formula (13), This represents the feature vector of the key problem-solving steps identified. Indicates the first The importance weight of each time step Indicates the first Semantic representation of steps Represents the characteristic transformation matrix. Indicates the first The implicit state of the step, This represents the bias vector. This represents the total number of steps required to solve the problem. This represents the hyperbolic tangent activation function.

[0055] Continuing based on the aforementioned tensor, the attention weight distribution is calculated to identify key problem-solving steps in the student's answer text. The attention mechanism assigns weights to highlight important parts. Specifically, in the answer to a physics problem, if a student writes "speed equals distance divided by time," the system calculates weights using the softmax function, emphasizing keywords such as "distance" and "time," thereby extracting steps such as "identifying variables" and "applying formulas" as features.

[0056] Step S230: Perform time-series alignment between the key problem-solving steps and the standard deduction nodes in the preset standard answer sequence. If the matching confidence is lower than the preset threshold, it is determined to be a logical deviation link and a comprehensive evaluation score is calculated.

[0057] The match confidence score is calculated using the following formula: (14) In formula (14), Indicates the match confidence level. This represents the total number of steps to solve the problem. Indicates the first Key characteristics of problem-solving steps Indicates the first A standard simulation node, The function represents the similarity calculation. The maximum value of the step length is used for normalization.

[0058] The following formula is used to define the judgment condition for logical deviation: (15) In formula (15), This indicates a logical deviation from the judgment result. This represents a preset threshold. When the confidence level is lower than the threshold, it is determined to be 1, indicating a logical deviation; otherwise, it is 0, indicating no deviation.

[0059] The overall evaluation score is derived using the following formula: (16) In formula (16), This indicates the overall evaluation score. Indicates the correctness score. Indicates the degree of deviation. Indicates the timing alignment score. , , These represent the weight coefficients of each evaluation indicator, and the sum of the weights is 1.

[0060] The key problem-solving steps are aligned temporally with the standard deduction nodes in the preset standard answer sequence. This can be achieved through a dynamic time warping algorithm, which aligns the student's step sequence with standard nodes such as "Step 1: Formulate, Step 2: Calculate" and calculates the similarity. For example, in a chemistry equilibrium problem, if a student omits "considering the effect of temperature," and the matching confidence level is below the threshold of 0.8, it is judged as a logical deviation, and a weighted summation is used to calculate the comprehensive evaluation score, such as deducting the corresponding points.

[0061] Step S240: Based on the comprehensive evaluation score, generate structured evaluation data and output a list of grading results.

[0062] The output format of structured grading results is defined by the following formula: (17) In formula (17), A data structure representing a list of grading results. This indicates the total number of items that need to be revised. Indicates the first A unique identifier for each project Indicates the first The evaluation level classification of each project Indicates the first The final evaluation results for each project.

[0063] The system generates structured evaluation data based on the comprehensive assessment score and outputs a list of grading results. The structured data includes the score, deviation points, and suggestions, and is output in tabular form. Specifically, for Chinese composition, the system generates data such as "Content score: 8, Structure score: 7, Total score: 15, Suggestion: Strengthen the support of the argument," which facilitates teacher review.

[0064] Preferably, the text analysis method for intelligent marking based on artificial intelligence provided in this embodiment includes step S300: Step S310: Obtain the erroneous question data from the grading results list, extract the original question text and the student's incorrect answer text, and perform entity extraction on the original question text to obtain the question entity set.

[0065] The following formula describes the process of entity extraction from the original question text: (18) In formula (18), This represents the set of extracted question entities. This represents a named entity recognition function. This represents the original question text. This indicates the text of a student's incorrect answer.

[0066] When processing the grading results of math homework, the first step is to retrieve the data on the questions that students answered incorrectly from the grading results list. For example, for a question about a quadratic equation, the original text might be "Solve..." The question asks for the root of "+3x-4=0", while the student's incorrect answer was "x=1 and x=-4". Named entity recognition (NER) is used to extract entities from the original question text. This technique, typically based on a pre-trained language model like BERT, scans the text and identifies key entities such as "quadratic equation," "root," and "coefficient," forming a set of question entities. This process involves segmenting the text and inputting it into the model. The model uses an attention mechanism to capture the context and outputs labeled entity tags, resulting in a set like {quadratic equation, root, coefficient 3, constant -4}.

[0067] Step S320: Map the question entity set to the subject knowledge graph to locate knowledge nodes. If the semantic deviation distance between the student's wrong answer text and the standard explanation of the knowledge node exceeds the preset deviation threshold, then the wrong knowledge points to be reinforced are identified.

[0068] The following formula is used to achieve the optimal mapping from question entities to knowledge nodes by minimizing the distance: (19) In formula (19), This represents the mapping function from the question entity to the knowledge graph. Represents the set of entities in the question. This represents the set of knowledge nodes in a subject knowledge graph. Indicates the first Each title entity, Indicates the first Each knowledge node A distance metric function representing the distance between an entity and a knowledge node.

[0069] The semantic bias is quantified by calculating the mean absolute difference between the two text embedding vectors of the student's incorrect answer text and the standard explanation of the knowledge node using the following formula: (20) In formula (20), This represents the semantic deviation distance between a student's incorrect answer and the standard explanation. This represents the text of a student's incorrect answer. The standard explanatory text representing knowledge nodes, This represents a text embedding function. This represents the dimension of the embedding vector. This represents the dimension index of the embedded vector.

[0070] The following formula was used to implement a threshold-based error knowledge point determination mechanism: (twenty one) In formula (21), This indicates the results of identifying incorrect knowledge points. This represents the calculated semantic deviation distance value. This represents the preset deviation threshold. When the semantic deviation distance exceeds the preset threshold, output 1 to indicate that it is identified as an incorrect knowledge point that needs to be strengthened; otherwise, output 0 to indicate that it does not need to be strengthened.

[0071] This set of question entities is mapped to a subject knowledge graph, a structured graph database where nodes represent knowledge concepts and edges represent relationships. For example, in a mathematical knowledge graph, the node "quadratic equation" connects to the node "quadratic formula". The mapping process uses a similarity matching algorithm, such as cosine similarity, to align entities with graph nodes, locating specific knowledge nodes such as "quadratic equation root finding". Then, the semantic deviation distance between the student's incorrect answer text and the standard explanation of that node is calculated. Here, the standard explanation might be "using the quadratic formula". The deviation distance is calculated by embedding vectors, such as generating vectors using Word2Vec and then calculating the Euclidean distance. If it exceeds the threshold of 0.7, then the knowledge points that need to be reinforced, such as "the application of the quadratic formula", are identified.

[0072] Step S330: Construct a weak knowledge subgraph based on the knowledge points to be strengthened, and input it into the question generation model to obtain a sequence of candidate variant questions.

[0073] Construct a weak knowledge subgraph using the following formula: (twenty two) In formula (22), Represents a subgraph with weak knowledge. This represents the set of nodes representing weak knowledge points. This represents the set of edges between weak knowledge points. This represents the set of incorrect knowledge points that need to be reinforced. Indicates the first Weak knowledge points Representing knowledge points Adjacent knowledge points The associated edges between them Representing knowledge points The set of adjacent nodes.

[0074] Based on these identified knowledge gaps, a weak knowledge subgraph is constructed. This involves extracting a subgraph from the knowledge graph, for example, using "quadratic formula" as the central node and expanding related nodes such as "discriminant" and "real root condition" to form a subgraph structure containing nodes and edges. This subgraph is input into a question generation model, such as a Transformer-based generative model. It generates a sequence of variant questions by encoding the subgraph information. The model first converts the subgraph into a sequence input, then decodes and outputs candidate questions, such as "solve...". "+5x+6=0" or more complex variations.

[0075] Step S340: Filter the candidate variation question sequence and arrange it according to the difficulty gradient to obtain a personalized practice question set for students.

[0076] The following formula describes the rules for constructing personalized practice question sets, ensuring that questions are arranged from easy to difficult: (twenty three) In formula (23), Represented as the first Personalized practice question sets generated by each student This indicates the number after sorting by difficulty. A question, This represents the total number of questions in the question set. This indicates the difficulty level of the corresponding question.

[0077] When selecting these candidate variation problem sequences, a difficulty assessment algorithm is used to arrange them in a gradient. For example, the complexity of each problem is assessed based on the number of entities and the depth of logic, starting with lower difficulties such as simple root problems, and then moving to higher difficulties such as parametric equations. This results in a personalized set of practice problems, such as a list format where the first problem is at difficulty level 1, the second at difficulty level 3, etc., helping students gradually strengthen their weaknesses. This method effectively improves the focus of learning and reduces repetitive errors through targeted practice.

[0078] Furthermore, the text analysis method for intelligent marking based on artificial intelligence provided in this embodiment includes step S400 as follows: Step S410: Obtain the question structure features of each question in the personalized practice question set, and identify the starting and ending points of the implicit logical reasoning chain in the question.

[0079] The starting point of the implicit logical reasoning chain in the question can be derived using the following formula: (twenty four) In formula (24), The concept node representing the starting point of a logical reasoning chain. Indicates the first... One candidate concept node, Represents concept nodes The number of in-degrees, Represents concept nodes The complexity evaluation value, This represents the complexity weight adjustment parameter.

[0080] The endpoint of the implicit logical reasoning chain in the question can be obtained using the following formula: (25) In formula (25), This represents the endpoint of a logical reasoning chain. Indicates the first One possible endpoint of reasoning, This represents the total number of reasoning paths. Indicates the first The credibility weight of each reasoning path Indicates the distance from the starting point to the end point. After the first Connectivity score for each path.

[0081] When processing personalized practice sets, the first step is to identify the structural features of each question. For example, for a math question like "Calculate the area of ​​a triangle, given a base of 5 and a height of 4," structured parsing techniques are used to identify the start and end points of its implicit logical reasoning chain. This parsing technique is typically based on syntax trees, decomposing the question text into the body, conditions, and target parts, with the starting point being "given the base and height" and the ending point being "find the area value." Specifically, this process involves inputting the text into a pre-trained parsing model. This pre-trained model captures the sentence structure through a hierarchical attention mechanism and outputs a tree-like representation, where leaf nodes correspond to basic elements, such as the numbers 5 and 4, while the root node represents the overall solution target, thus clearly identifying the start and end points of the reasoning chain. For example, in practical applications, for a geometry question like "Prove two lines are parallel, given that the corresponding angles are equal," the starting point is identified as "the fact that the corresponding angles are equal," and the ending point is "the conclusion that the two lines are parallel," which helps to build more accurate guidance later.

[0082] Step S420: Construct a reasoning dependency graph based on the starting and ending points of the logical reasoning chain, and insert detailed problem-solving step nodes into the reasoning dependency graph.

[0083] Construct the inference dependency graph using the following formula: (26) In formula (26), Represents the dependency graph for inference. This represents the set of nodes that form the starting point of the reasoning chain. Represents the set of endpoint nodes in the reasoning chain. Represents the set of all vertices in the graph. Represents a set of directed edges. Represents the set of intermediate inference nodes. and Represents the nodes in the graph. Indicates the direction of inference dependencies.

[0084] Based on these logical reasoning chains, a reasoning dependency graph is constructed from the starting and ending points. For example, the starting node is connected to the intermediate assumption node, and then extended to the ending point, forming a directed graph structure where edges represent logical dependencies, such as from "corresponding angles are equal" to "application of the parallel axiom". Detailed problem-solving step nodes are inserted into this reasoning dependency graph. Specifically, nodes are added using graph expansion algorithms, such as inserting "recall the definition of parallel lines" and "apply the isotropic theorem". These nodes are connected by weighted edges to ensure the coherence of the reasoning dependency graph. This method effectively captures the levels of problem-solving and helps generate comprehensive guidance paths. For example, for area calculation problems, the dependency graph might include the starting point "base 5 and height 4", inserting the step node "recall the area formula S=(1 / 2)bh", then connecting to "substitute numerical calculation", and finally reaching the ending point "obtain the area 10", thus forming a complete reasoning framework.

[0085] Step S430: Traverse the problem-solving step nodes to generate a thought-guiding path with causal relationships. Input the thought-guiding path into the natural language generation model for semantic logic mapping to obtain a detailed thought-guiding text containing logical reasoning. The detailed thought-guiding text is generated by the natural language generation model after transforming the abstract reasoning process.

[0086] The quality assessment value of the detailed thought guidance text generated after the abstract reasoning process is transformed is obtained using the following formula: (27) In formula (27), This represents the quality assessment value of the detailed thought process guidance text generated after the abstract reasoning process is transformed. This represents the total number of nodes in the reasoning steps. Represents a node Importance weights Represents the semantic similarity function. Represents a node The encoding representation, Indicates generated text semantic representation, Represents a node With text Information retention between them This represents the set of nodes for the reasoning steps. The control logic of formula (27) is to first calculate the product term of "node importance weight × semantic similarity (similarity between node encoding and generated text semantics) × information retention (information matching degree between node and text)" for each reasoning step node, then sum the product terms of all nodes, and finally divide the summation result by the total number of reasoning step nodes to obtain the quality assessment value of the detailed reasoning guidance text.

[0087] This process iterates through the problem-solving steps to generate a thought-guided path with causal relationships. For example, starting from the starting node, the steps are linked sequentially to form a path like "First, recall the formula, then substitute the data, and finally verify the result," emphasizing causal relationships such as "Because the formula is true, the calculation is valid." This path is then input into a natural language generation model for semantic and logical mapping. Specifically, this model, such as a variant based on GPT, transforms the abstract path into coherent text, encoding the path sequence and decoding it to generate narrative guidance. The resulting detailed thought-guided text is generated by the model after transforming the abstract reasoning process, for example, outputting "Step 1: Identify the given conditions: base 5 and height 4, this is the starting point for calculation; Step 2: Apply the area formula S=(1 / 2)bh, because this is a basic geometric principle; Step 3: Substitute and calculate to get 10, ensuring logical continuity." For example, in business applications, this guidance text can provide step-by-step solutions to students' weaknesses, improving problem-solving efficiency and enabling personalized feedback in education systems, thereby significantly improving the learning experience.

[0088] Preferably, the text analysis method for intelligent marking based on artificial intelligence provided in this embodiment includes step S500: Step S510: Obtain the error type codes and thought breakpoint location data from the detailed thought process guidance text in the correction results list.

[0089] The following formula is used to calculate the weighted average coding value of all error types in the correction results: (28) In formula (28), This represents the overall evaluation value of the error type code. This indicates the total number of errors in the correction results list. Indicates the first An incorrect weighting coefficient Indicates the first The type-coded value of the error.

[0090] When processing the list of graded results, the first step is to extract error type codes. These codes are typically predefined classification identifiers, such as labeling calculation errors as E01 and logical reasoning errors as E02. This extraction process is based on text parsing technology, breaking down the list content into structured data. For example, regular expressions are used to match coded fields to obtain the E01 code, such as when a student mistakenly performs addition as multiplication in a math problem. Next, the data on thought breakpoints in the detailed guidance text is analyzed. These data represent the locations where the student's reasoning chain is interrupted. For example, in solving an equation, the student might get stuck at the variable isolation step, marked as P3. This can be located by analyzing the text structure using a natural language processing (NLP) model. The NLP model scans the sentence sequence of the guidance text, identifying keywords such as "the distributive law should be applied here" to mark breakpoints, thus forming a list of location data to ensure the accuracy of subsequent integration.

[0091] Step S520: Execute the integration mapping rules based on the error type code and thought breakpoint location data to associate and match the error correction logic in the detailed thought guidance text with the correction result list.

[0092] The following formula is used to define the mapping function between error type codes and thought breakpoint location data: (29) In formula (29), A function representing the mapping between error type codes and thought breakpoint location data. Indicates the first Error type encoding, Indicates the first Data on the location of each thought breakpoint. Indicates the total number of error correction logic. Indicates the first The weighting coefficients of each error correction logic. A function representing the matching degree between error type and error correction logic. Indicates the first A correction logic, The activation function represents the location of the thought breakpoint. The control logic of formula (29) is to first calculate the product of "its weight coefficient × the matching degree between the error type and the error correction logic" for each error correction logic, then sum the product results of all error correction logics, then divide the sum by the sum of the weights of all error correction logics to obtain the weighted matching degree between the error type and the error correction logic, and finally multiply this degree by the activation function result of the thought breakpoint location to obtain the mapping result between the error type code and the thought breakpoint location data.

[0093] Based on the extracted error type codes and thought breakpoint locations, an integrated mapping rule is executed. This involves associating codes with breakpoint data; for example, mapping the E01 code to the P3 breakpoint. The integrated mapping rule uses a conditional matching algorithm. If the code indicates a calculation error and the breakpoint is at the numerical substitution point, then error correction logic such as "check if the substituted value is correct" is generated. In this process, the error correction logic is a correction path derived from the detailed thought process guidance text. During specific matching, each item in the correction result list is traversed and aligned with the corresponding paragraph in the guidance text. For example, if the correction result for a physics problem shows that the student ignored the unit conversion (E03 code), it is matched with the thought breakpoint P5 in the guidance text, generating the logic "remind the student to recalculate after unifying the units," thus achieving association and ensuring the targeted nature of the error correction.

[0094] Step S530: Generate tutoring suggestion module based on error correction logic and construct error review content block.

[0095] The following formula is used to calculate the learner's progress after reviewing incorrect answers, and the learning effect is displayed visually as a percentage: (30) In formula (30), This indicates the rate of mastery of incorrect answers. This indicates a review of the total number of questions. This indicates the number of questions that are still incorrect.

[0096] When generating the tutoring suggestion module, the module content is constructed based on the aforementioned error correction logic. This tutoring suggestion module is a structured sequence of suggestions. It first introduces the cause of the error, and then provides corrective steps. For example, the suggestion for E01 encoding is "Review basic arithmetic rules to avoid confusing symbols." The construction process involves template filling, translating the logic into natural language. For instance, the module begins by describing, "Your error lies in misusing a formula in the calculation steps," followed by step-by-step guidance: "Step one: Verify the input data; Step two: Apply the correct formula." This helps students understand the root cause of the problem. In actual educational practice, this tutoring suggestion module can enhance students' self-learning ability because it not only points out errors but also provides actionable improvement paths.

[0097] The construction of the error review content block is based on the error correction logic to summarize historical errors. For example, the E02 codes of multiple questions are integrated into a review block, which includes "Review of the last logical breakpoint: omission of assumptions in the proof steps". The block is presented in the form of a list, with each item containing the original question, the error point and the correction suggestion to ensure the comprehensiveness of the review.

[0098] Step S540: Fill the tutoring suggestions module and the wrong question review content block into the document layout template and render it to obtain a feedback document containing wrong question review and tutoring suggestions.

[0099] The rendering effect evaluation value of the feedback document is obtained using the following formula: (31) In formula (31), This indicates the rendering effect evaluation value of the feedback document. This indicates the document's layout quality score. This represents the content integration quality coefficient. This represents the document length parameter. This represents the complexity compensation value. The rendering efficiency factor is represented by formula (31). The control logic of formula (31) is to first calculate the product of "document layout quality score × content integration quality coefficient × document length parameter", then calculate the product of "complexity compensation value × rendering efficiency factor", and finally add the two product terms to obtain the rendering effect evaluation value of the feedback document.

[0100] Next, the tutoring suggestions module and the error review content block are filled into a document formatting template. This template is a preset HTML or PDF framework. Data binding technology is used during the filling process, such as inserting module text into specified DIV elements and placing the review block in a table. Then, the document is rendered to generate a feedback document. The rendering process involves applying styles, such as adding color to highlight error points. The final output is a document containing error review and tutoring suggestions. From a business perspective, this provides teachers with an efficient feedback tool, significantly improving teaching efficiency. Specifically, in an example of a chemical equilibrium problem, the grading results list shows error type code E04 as "miscalculation of equilibrium constant", and the thought breakpoint data is P2, which is "incorrect concentration substitution". After integration and mapping, the error correction logic is generated as "verify the initial concentration and recalculate the K value". Then, a tutoring suggestion module is constructed, such as "suggested steps: list the reaction equation, calculate the concentration of each substance, and ensure that the units are consistent". The error review section reviews similar questions, such as "last error: ignoring the effect of temperature, the result after this correction is equilibrium". The document rendered after filling the template is presented in PDF format, with the modules on the top half of the page and the review on the bottom half of the page, which is convenient for students to review.

[0101] Further, step S600 includes: Step S610: Parse the feedback document to obtain the error correction logic text, and convert the error correction logic text into a standardized semantic feature vector.

[0102] The error-correction logic text is converted into a standardized semantic feature vector using the following formula: (32) In formula (32), Represents a standardized semantic feature vector. This represents a semantic transformation function. This represents the input error correction logic text. This represents the total number of semantic units in the logical text. Indicates the first Importance weights of each semantic unit Represents word embedding function, Indicates the first The control logic of formula (32) is to first calculate the result of "semantic unit importance weight × word embedding vector of the unit" for all semantic units of the error correction logic text, then sum these results, and finally divide the sum by the total number of semantic units to obtain the standardized semantic feature vector.

[0103] The process of parsing feedback documents first involves the structured extraction of document content. For example, in an educational grading system, feedback documents typically include sections reviewing incorrect answers and providing tutoring suggestions. By scanning the document format using text parsing tools, text blocks such as those under the "Correction Logic" tag are identified, thus yielding the specific correction logic text. This correction logic text refers to the description of corrections made to student assignment errors. For instance, in a math assignment feedback document, the document might state, "The student neglected to calculate the discriminant when solving the quadratic equation, leading to a radical error." During parsing, natural language processing techniques are used to break down the sentence, extracting key phrases such as "neglected discriminant" and "radical error," forming independent logical text units to ensure the accuracy of subsequent transformations. This parsing helps the system understand the root causes of grading deviations and provides a data foundation for model optimization in business operations.

[0104] The process of converting error-correction logic text into standardized semantic feature vectors requires the introduction of a word embedding model to process the text. For example, the Word2Vec model can be used to vectorize the text. First, the text is segmented into words such as "solve," "quadratic equation," "ignore," and "discriminant." Then, each word is mapped to a high-dimensional vector space, and the overall semantic feature vector is obtained by averaging or weighted summation. In the example of Chinese composition correction, if the error-correction logic text is "students make logical leaps and lack transition sentences when arguing their points," the converted vector may capture the semantic similarity between "logical leaps" and "lack of transitions." Standardization is achieved by comparing with a standard vector library to ensure that the vector dimension is uniform, such as 128 dimensions. This facilitates quantitative comparison of correction quality in business applications.

[0105] Step S620: Calculate the Euclidean distance between the semantic feature vector and the original output vector, and determine the correction deviation value based on the Euclidean distance.

[0106] The Euclidean distance between the semantic feature vector and the original output vector is obtained by the following formula: (33) In formula (33), This represents the Euclidean distance between the semantic feature vector and the original output vector. The semantic feature vector represents the first One portion, Represents the first element of the original output vector. One portion, The dimension of the vector is represented by . The control logic of formula (33) is to first calculate the difference between the corresponding dimension components of the semantic feature vector and the original output vector, then perform a square operation on each difference, then sum the square results of all dimensions, and finally take the square root of this sum to obtain the Euclidean distance between the two vectors.

[0107] The correction deviation value is obtained using the following formula: (34) In formula (34), This indicates the numerical value of the correction deviation. This represents the distance weighting coefficient. Indicates the maximum possible distance value. This represents the basic deviation adjustment parameter. The control logic of formula (34) is to first adjust the current Euclidean distance... and the maximum possible distance Calculate the ratio to obtain the relative proportion of distances, then use the distance weighting coefficient. This percentage is weighted and then adjusted by the basic deviation parameter. This yields the final correction deviation value. .

[0108] When calculating the Euclidean distance between the semantic feature vector and the original output vector, the original output vector originates from the initial model's correction results and represents the preset correct correction features. The distance value is obtained by taking the square root of the sum of the squares of the differences in each dimension of the two vectors. In grading physics lab reports, if the semantic vector reflects "ignoring unit conversion" corrections while the original vector emphasizes "accurate calculations," a larger distance indicates a significant deviation. In one implementation, determining the correction deviation value based on the Euclidean distance involves mapping the distance to a deviation score; for example, a distance less than 0.5 is considered low deviation and is applied to historical assignment data to evaluate model performance.

[0109] Step S630: Calculate the gradient descent direction based on the correction deviation value, and perform weighted correction on the weight matrix of the model parameter set of the machine learning model and the artificial intelligence model in combination with the gradient descent direction.

[0110] The gradient descent direction is calculated using the following formula: (35) In formula (35), Represents the loss function For the weight matrix gradient, This represents the weight matrix of the model. Indicates the number of training samples. Indicates the first The true label of each sample The model represents the first The predicted output for each sample, Indicates the first The input samples. The control logic of formula (35) is to first calculate the difference between the true label and the model prediction output of each training sample and take the square (corresponding to the single sample loss of the loss function), then sum the losses of all samples to get the total loss, and finally take the partial derivative of the total loss with respect to the weight matrix to get the gradient of the loss function as the weight matrix changes (i.e. the direction of gradient descent).

[0111] The weighted update of the weight matrix based on the gradient descent direction is achieved using the following formula: (36) In formula (36), This represents the updated weight matrix. This represents the weight matrix before the update. This represents the learning rate parameter. This represents the partial derivative of the error function with respect to the weight matrix. The control logic of formula (36) is based on the weight matrix before the update. First, it calculates the adjustment amount of the weights by "learning rate × partial derivative of error with respect to the weight matrix" (the direction of the partial derivative represents the direction of the increase of error, so the minus sign is used to indicate the adjustment along the direction of the decrease of error). Then, this adjustment amount is subtracted from the original weight matrix to finally obtain the updated weight matrix.

[0112] Calculating the gradient descent direction based on the grading deviation involves optimization algorithms. The deviation serves as a loss function, and its derivative yields the direction, which is used to adjust model parameters. In a grading system, this gradually reduces errors. For example, when using the gradient descent direction to weight the weight matrices of machine learning and artificial intelligence models, the weight matrix represents the core parameter set of the model, such as the connection weights in a neural network. By multiplying it by the direction vector and updating it, the model is made more suitable for educational scenarios.

[0113] Step S640: Use the corrected weight matrix to perform secondary inference on the validation set data. When the inference result meets the convergence baseline, solidify the parameters to obtain the optimized correction accuracy.

[0114] The following formula is used to determine whether the result of quadratic inference satisfies the convergence condition: (37) In formula (37), This indicates the convergence criterion. Indicates the number of samples in the validation set. Indicates the first The output of this inference, Indicates the first The output of this inference. This represents the convergence baseline threshold. The control logic of formula (37) is to first calculate the "first" threshold for each sample in the validation set. The output of the second reasoning and the first The L2 norm of the "second inference output" (measuring the difference between the two outputs) is then used to calculate the average difference by summing the L2 norms of all samples and dividing by the number of samples in the validation set. Finally, it is determined whether this average difference is less than or equal to the convergence baseline threshold. If the condition is met, the result of the quadratic inference is considered to have reached convergence.

[0115] The optimized grading accuracy rate is calculated using the following formula: (38) In formula (38), This indicates the optimized accuracy rate of corrections. This represents the total number of test samples. Indicates an indicator function, This represents the corrected result of the model prediction. The actual correction label is represented by the indicator function. The control logic of formula (38) is to first determine whether the "model prediction correction result is consistent with the actual correction label" of each test sample through the indicator function (the indicator function outputs 1 when consistent and 0 when inconsistent), then sum the indicator function results of all test samples, and finally divide the summation result by the total number of test samples to obtain the average correct percentage on the test set, which is the optimized correction accuracy.

[0116] The process of using the revised weight matrix to perform secondary inference on the validation set data involves applying the new weights to a simulated homework set, outputting grading results, and checking for improvements. Specifically, parameters are fixed when the inference results meet a convergence baseline, such as an accuracy rate of 95%, resulting in an optimized grading accuracy rate. In practice, this significantly improves the reliability of teaching feedback.

[0117] This embodiment provides an AI-based intelligent text analysis method for marking assignments. Compared with existing technologies, it proposes an integrated solution for business scenarios involving handwritten content recognition, answer correction, error analysis, and personalized tutoring in student assignments. The method uses an image understanding module to accurately identify the question stem and answer text in assignment images, and a machine learning model to compare the results with standard answers to generate corrections. For errors, an error generation module analyzes the knowledge points, creates similar practice questions, and generates detailed problem-solving guidance text. Finally, these are integrated into a feedback document containing a review of errors and tutoring suggestions. Furthermore, a dynamic adjustment mechanism optimizes model parameters to improve correction accuracy. Its core lies in seamlessly integrating assignment correction with personalized tutoring, achieving full automation from recognition to feedback, significantly improving educational efficiency and learning relevance, and providing students with precise learning support.

[0118] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A text analysis method for intelligent marking based on artificial intelligence, characterized in that, Includes the following steps: S100. The image understanding module is used to process the homework images uploaded by the user. The image understanding module integrates optical character recognition function and uses an artificial intelligence model to perform semantic analysis and recognition on the handwritten content and mathematical symbols in the homework images to obtain the extracted question text and student answer text. S200. Based on the extracted question text and student answer text, a machine learning model is used to compare the student answer text with the preset standard answer to determine the correctness of the student answer text and obtain a list of grading results. S300. If the correction result list shows errors, the errors are classified and analyzed by the error generation module to identify the knowledge points involved in the errors, and the artificial intelligence model is used to create similar questions based on the knowledge points in the errors to obtain a personalized set of practice questions for students. S400. Based on the personalized practice question set, a step-by-step sequence of problem-solving steps is generated using the tutoring module, resulting in a detailed guidance text containing logical reasoning, specifically including: S410. Obtain the question structure features of each question in the personalized exercise question set, and identify the starting point and ending point of the implicit logical reasoning chain in the question; S420. Construct a reasoning dependency graph based on the starting and ending points of the logical reasoning chain, and insert detailed problem-solving step nodes into the reasoning dependency graph; S430. Traverse the problem-solving step nodes to generate a thought-guiding path with causal relationships. Input the thought-guiding path into the natural language generation model for semantic logic mapping to obtain a detailed thought-guiding text containing logical reasoning. The detailed thought-guiding text is generated by the natural language generation model after transforming the abstract reasoning process. S500: Integrate the list of correction results and the detailed guidance text through an automated workflow to determine the final output report and obtain a feedback document containing a review of incorrect questions and tutoring suggestions; S600. Based on the feedback document, the parameters of the machine learning model and the artificial intelligence model are iteratively updated using a dynamic adjustment mechanism to obtain the optimized grading accuracy, specifically including: S610. Parse the feedback document to obtain the error correction logic text, and convert the error correction logic text into a standardized semantic feature vector; The error-correction logic text is converted into a standardized semantic feature vector using the following formula: in, Represents a standardized semantic feature vector. This represents a semantic transformation function. This represents the input error correction logic text. This represents the total number of semantic units in the logical text. Indicates the first Importance weights of each semantic unit Represents word embedding function, Indicates the first One semantic unit; S620. Calculate the Euclidean distance between the semantic feature vector and the original output vector, and determine the correction deviation value based on the Euclidean distance; S630. Calculate the gradient descent direction based on the correction deviation value, and perform weighted correction on the weight matrix in the model parameter set of the machine learning model and the artificial intelligence model in combination with the gradient descent direction. S640. Use the modified weight matrix to perform secondary inference on the validation set data. When the inference result meets the convergence baseline, fix the parameters to obtain the optimized correction accuracy.

2. The text analysis method for intelligent marking based on artificial intelligence according to claim 1, characterized in that, Step S100 includes: S110. Obtain the original job image uploaded by the user and generate an enhanced job feature map containing clear foreground information; S120. Divide the printed question stem region and the handwritten handwriting region according to the enhanced task feature map to obtain the basic question stem character stream and the original handwritten symbol sequence; S130. Map the basic question stem character stream and the original handwritten symbol sequence into a character semantic vector to generate a question stem semantic representation and an answer semantic representation; S140. Parse the semantic representation of the question stem and the semantic representation of the answer, and reorganize them to generate a standardized question stem text sequence and student answer text sequence, so as to achieve accurate extraction of the question stem and answer from the homework image.

3. The text analysis method for intelligent marking based on artificial intelligence according to claim 1, characterized in that, Step S200 includes: S210. Obtain the preset standard answer sequence of the question stem text, and use the deep neural network in the machine learning model to encode the student answer text and the preset standard answer sequence to obtain the hidden layer semantic tensor. S220. Calculate the attention weight distribution based on the hidden layer semantic tensor to identify key problem-solving steps in the student's answer text; S230. The key problem-solving steps are aligned with the standard deduction nodes in the preset standard answer sequence. If the matching confidence is lower than the preset threshold, it is determined to be a logical deviation and a comprehensive evaluation score is calculated. S240. Based on the comprehensive evaluation score, generate structured evaluation data and output a list of grading results.

4. The text analysis method for intelligent marking based on artificial intelligence according to claim 1, characterized in that, Step S300 includes: S310. Obtain the erroneous question data from the grading result list, extract the original question text and the student's erroneous answer text, and perform entity extraction on the original question text to obtain the question entity set; S320. Map the set of question entities to the subject knowledge graph to locate knowledge nodes. If the semantic deviation distance between the student's wrong answer text and the standard explanation of the knowledge node exceeds a preset deviation threshold, then determine the wrong knowledge points to be strengthened. S330. Construct a weak knowledge subgraph based on the erroneous knowledge points to be strengthened, and input it into the question generation model to obtain a candidate variant question sequence; S340. Filter the candidate variation question sequence and arrange it according to the difficulty gradient to obtain a personalized practice question set for students.

5. The text analysis method for intelligent marking based on artificial intelligence according to claim 1, characterized in that, Step S500 includes: S510. Obtain the error type codes and thought breakpoint location data from the detailed thought process guidance text in the correction results list. S520. Based on the error type code and the thought breakpoint location data, execute the integration mapping rule to associate and match the error correction logic in the detailed thought guidance text with the correction result list. S530. Generate a tutoring suggestion module based on the error correction logic and construct a wrong question review content block; S540. Fill the tutoring suggestion module and the wrong question review content block into the document layout template and render it to obtain a feedback document containing wrong question review and tutoring suggestions.

6. The text analysis method for intelligent marking based on artificial intelligence according to claim 1, characterized in that, In step S620, the Euclidean distance between the semantic feature vector and the original output vector is obtained by the following formula: in, This represents the Euclidean distance between the semantic feature vector and the original output vector. The semantic feature vector represents the first... One portion, Represents the first element of the original output vector. One portion, Indicates the dimension of the vector; The correction deviation value is obtained using the following formula: in, This indicates the numerical value of the correction deviation. This represents the distance weighting coefficient. Indicates the maximum possible distance value. This indicates the basic deviation adjustment parameter.

7. The text analysis method for intelligent marking based on artificial intelligence according to claim 6, characterized in that, In step S630, the gradient descent direction is calculated using the following formula: in, Represents the loss function For the weight matrix gradient, This represents the weight matrix of the model. Indicates the number of training samples. Indicates the first The true label of each sample The model represents the first The predicted output for each sample, Indicates the first One input sample; The weighted update of the weight matrix based on the gradient descent direction is achieved using the following formula: in, This represents the updated weight matrix. This represents the weight matrix before the update. This represents the learning rate parameter. This represents the partial derivative of the error function with respect to the weight matrix.

8. The text analysis method for intelligent marking based on artificial intelligence according to claim 7, characterized in that, In step S640, the following formula is used to determine whether the result of the quadratic inference satisfies the convergence condition: in, This indicates the convergence criterion. Indicates the number of samples in the validation set. Indicates the first The output of this inference, Indicates the first The output of this inference, Indicates the convergence baseline threshold; The optimized grading accuracy rate is calculated using the following formula: in, This indicates the optimized accuracy rate of corrections. This represents the total number of test samples. Indicates an indicator function, This represents the corrected result of the model prediction. Indicates a genuine correction label.

Images