Student mistake reason induction test paper analysis system based on an AI large model

Abstract

The application belongs to the technical field of artificial intelligence and AI teacher homework correction system, and relates to a student mistake reason induction test paper analysis system based on an AI large model, which comprises: an image acquisition module, which is used for acquiring a question image; a question positioning module, which is used for processing the question image and extracting a question frame of the preprocessed question image; an information extraction module, which is used for extracting MLLM information and OCR information of the question frame, performing question ability tagging and question knowledge point tagging on questions in the MLLM information, generating an ability label and a knowledge point path label of the current question, fusing the OCR information and handwriting answer content of the MLLM information, and correcting the fusion result in combination with the question of the MLLM information and a correction state; and a mistake reason analysis module, which is used for performing mistake reason analysis according to a correction result in combination with the ability label and the knowledge point path label of the current question to obtain a mistake reason result.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and AI teacher homework correction systems, and in particular to a test paper analysis system based on AI large-scale model for summarizing student error causes. Background Technology

[0002] Existing methods mainly suffer from the following five problems: Superficial error analysis: Most systems can only determine whether an answer is right or wrong, but cannot deeply analyze the reasons for errors (such as conceptual misunderstandings, calculation errors, and misinterpretation of the question); Poor subject adaptability: Existing systems are often designed for a single subject, making it difficult to handle multiple subjects such as mathematics, physics, chemistry, and biology uniformly; Reliance on manual annotation: Traditional systems require a large amount of manual pre-annotation of knowledge points and ability tags for questions, resulting in high costs and poor scalability; Insufficient OCR recognition accuracy: The recognition accuracy for handwriting, complex formulas, and mixed content such as charts is low; Lack of inductive and reasoning ability: It is impossible to extract common error causes from incorrect questions, making it difficult to formulate effective teaching suggestions.

[0003] Therefore, there is an urgent need for a test paper analysis system based on AI large-scale models to summarize student error causes and solve the problems existing in current methods. Summary of the Invention

[0004] To address the problems existing in the prior art, the purpose of this invention is to provide a student error analysis test paper system based on an AI large-scale model. This system enables unified cross-disciplinary analysis, supporting question comprehension, ability tag generation, and error analysis across four subjects: mathematics, physics, chemistry, and biology. It features deep error summarization, not only judging right or wrong answers but also identifying error types (such as conceptual errors, calculation errors, and flawed reasoning). Automated tag generation uses a large-scale model to automatically predict question knowledge points and ability tags, reducing reliance on manual intervention. High-precision OCR and structural understanding enhance the recognition and semantic understanding of handwriting, formulas, and charts. Finally, it forms an error knowledge graph, constructs student error profiles, and supports personalized learning path recommendations.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A test paper analysis system for summarizing student error causes based on an AI large-scale model includes:

[0007] The image acquisition module is used to acquire images of the questions.

[0008] The question location module is used to process the question image and extract the question frame of the preprocessed question image;

[0009] The information extraction module is used to extract the MLLM information and OCR information of the question frame, mark the question ability and knowledge points in the MLLM information, generate the ability tag and knowledge point path tag of the current question, merge the OCR information with the handwritten answer content of the MLLM information, and grade the fusion result in combination with the question and grading status of the MLLM information.

[0010] The error analysis module is used to analyze the error causes based on the grading results, combined with the ability tags and knowledge point path tags of the current question, and to obtain the error cause results.

[0011] Optionally, the information extraction module includes:

[0012] The information extraction submodule is used to extract the MLLM information of the title drawing frame;

[0013] The ability tagging submodule is used to decompose the standard answers and explanations of the questions in the MLLM information into cognitive atomic operation sequences, tag the cognitive atomic operation sequences with question abilities, and generate the ability tags of the current question.

[0014] The knowledge point tagging submodule is used to tag the knowledge points of the textbook version for the specified subject and grade level in the MLLM information and generate the knowledge point path tags.

[0015] The OCR recognition submodule is used to perform OCR recognition on the question frame using different OCR recognition strategies for different question types, and generate the OCR information.

[0016] The grading submodule is used to merge the OCR information with the handwritten answer content of the MLLM information, and to grade the merged result in combination with the question and grading status of the MLLM information.

[0017] Optionally, the capability labeling submodule includes:

[0018] An atomic operation generation unit is used to decompose the standard answers and explanations of the questions in the MLLM information into a sequence of cognitive atomic operations using a cognitive atomic operation library, which contains the most microscopic thought steps in the problem-solving process.

[0019] A capability labeling unit is used to map the cognitive atomic operations in the cognitive atomic operation sequence to a mapping matrix of atomic operations-macro capabilities, and generate the contribution weight of each cognitive atomic operation to each macro capability.

[0020] The contribution weights are weighted, summarized, and normalized to obtain an ability vector. The ability vector is compared with a preset ability threshold, and the macro-ability corresponding to the ability vector that is higher than the preset ability threshold is used as the ability label of the current question.

[0021] Optionally, the knowledge point tagging submodule includes:

[0022] A knowledge topology unit is used to construct a graph structure of dynamic knowledge topology and generate a knowledge graph. The knowledge graph includes: edges generated by the nodes corresponding to each knowledge point and the various relationships between the nodes, including: pre-requisite relationships, parallel relationships, and instantiation relationships.

[0023] The encoding and topology-aware unit is used to recall a preset number of similar questions and similar knowledge points from various textbook versions based on the similarity between the question and knowledge points in the Milvus vector database.

[0024] Extract the depth representation vector of the question content from the text and image of the question, and recall candidate knowledge point nodes from the Milvus vector database. Input the candidate knowledge point nodes and their neighbor relationships in the knowledge graph into the graph neural network model to obtain the topology-enhanced representation vector.

[0025] The matching degree between the depth representation vector and each topology enhancement representation vector is calculated using the cross-attention mechanism. The node with the highest matching degree is taken as the core knowledge point, and the core knowledge point is taken as the knowledge point core. A complete path from the root node to the last level knowledge point is generated, and the complete path is taken as the knowledge point path label of the current question.

[0026] Optionally, the knowledge point tagging submodule further includes:

[0027] The result correction unit is used to set up a post-repair logic for recall if the knowledge point does not exist in the textbook version of the specified subject and grade. It performs embedding vector recall on the knowledge tree of the textbook version for the complete path, determines the knowledge point of the specified textbook version that is closest, and corrects the output result of the encoding and topology sensing unit.

[0028] Optionally, the OCR recognition submodule includes:

[0029] The first OCR recognition unit is used to input the fill-in-the-blank questions and problem-solving questions with the question drawing frame into the text detection model and the OCR recognition model respectively, and obtain the handwritten drawing and font recognition results; the text detection model and the OCR recognition model are trained using a training set;

[0030] The second OCR recognition unit is used to call a multimodal large language model to recognize the handwritten solutions of students in the problem-solving questions with the question frame drawn.

[0031] Optionally, the second OCR recognition unit includes:

[0032] The handwriting element deconstruction subunit is used to identify all coherent handwritten text blocks and non-text marks, and to isolate the mathematical expression area. It further records the absolute and relative positions between all handwriting elements, that is, the absolute and relative positions between the handwritten text blocks, the non-text marks, and the mathematical expression area.

[0033] The cognitive reasoning and reconstruction subunit is used to perform problem-solving reasoning based on the text content of the question within the question frame, the solution steps of the standard answer, and all handwriting elements.

[0034] The writing flow reconstruction subunit is used to reconstruct a cognitive writing sequence that conforms to the problem-solving logic based on the problem-solving reasoning results;

[0035] The OCR recognition subunit is used to input the cognitive writing sequence into the multimodal large language model to obtain the handwritten answer result.

[0036] Optionally, the error cause analysis module includes:

[0037] The error analysis submodule is used to pre-generate the error reasons for each incorrect option in multiple-choice and fill-in-the-blank questions based on the question information and insert them into the MongoDB database. When analyzing students' error reasons in batches, the MongoDB database is queried directly based on the students' incorrect options to match the corresponding error reason results.

[0038] When analyzing fill-in-the-blank questions, a fallback error reason is also set up. That is, when students actually answer the fill-in-the-blank questions, if not all the wrong answers can be matched, a fallback error reason will be given by default.

[0039] The solution to the problem is input into the multimodal model. By comparing the answer analysis with the student's answer process, the specific steps in the student's solution where the error occurred are located. The detailed error reasons are obtained by combining the ability tag and knowledge point path tag of the current problem.

[0040] The beneficial effects of this invention are as follows:

[0041] This invention employs different approaches to different subjects and question types, allowing for personalized analysis, and obtains a knowledge point prediction and labeling model through training a large model.

[0042] This invention uses a refined prompt injection guidance model to deeply analyze students' incorrect results and provide detailed and accurate reasons for the errors.

[0043] This invention trains an SFT model based on annotated data, allowing the model to be labeled with specified atlas knowledge points according to the textbook version and grade level.

[0044] This invention employs a combination of a question border detection model, a handwritten OCR recognition model, and a multimodal large model to achieve efficient OCR recognition for different question types and formats.

[0045] This invention utilizes a high-quality prompt model that continuously iterates to describe the causes of students' mistakes and identify the core pain points of the errors. It then provides corresponding teaching suggestions and records students' frequently missed questions and knowledge points, storing them in a database to gradually form user profiles. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a schematic diagram of a student error analysis test paper system based on an AI large model, according to an embodiment of the present invention.

[0048] Figure 2 This is a schematic diagram of the question ability labeling module in an embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram of the knowledge point tagging module in an embodiment of the present invention;

[0050] Figure 4 This is a schematic diagram of the OCR recognition module according to an embodiment of the present invention;

[0051] Figure 5 This is a schematic diagram of the student error analysis module in an embodiment of the present invention. Detailed Implementation

[0052] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] like Figure 1 As shown in the figure, this embodiment discloses a student error summary test paper analysis system based on an AI large model, including: an image acquisition module for acquiring question images; a question localization module for processing the question images and extracting the question frames from the preprocessed question images; an information extraction module for extracting MLLM information and OCR information from the question frames, tagging the questions in the MLLM information with question abilities and knowledge points, generating ability tags and knowledge point path tags for the current question, fusing the handwritten answers in the OCR information and MLLM information, and grading the results by combining the questions and grading status in the MLLM information; and an error analysis module for performing error analysis based on the grading results and the ability tags and knowledge point path tags for the current question to obtain error analysis results.

[0055] Specifically, this embodiment discloses a student error analysis test paper system based on an AI large-scale model, including:

[0056] The question ability tagging module defines several ability tags for the four subjects based on their abilities and qualities, such as data analysis ability and intuitive phenomenon ability in mathematics, physical concept ability and scientific thinking ability in physics. The module calls LLM through the prompt method to indicate which abilities and qualities are tested in the question based on the question stem, answer and analysis.

[0057] The question knowledge point labeling module uses a training model to label the knowledge points of questions based on different knowledge graphs of the existing K12 knowledge system. The input is the question stem and answer analysis, plus the subject-level graph version and the old and new curriculum standards. The output is the lowest level knowledge points of the grade.

[0058] The OCR recognition module first determines the type of the input question and adopts different OCR recognition strategies for different question types. The main questions that need to be OCR recognized are fill-in-the-blank questions and problem-solving questions. For fill-in-the-blank questions with relatively standard format, a trained text detection model and OCR recognition model are used to perform handwriting circling and font recognition. For problem-solving questions with relatively messy format, a prompt method is used to call a multimodal LLM to recognize the student's handwritten solution process.

[0059] The student error analysis module also employs different analysis strategies for different question types. For multiple-choice questions, there is no handwritten solution process, and the knowledge points tested are relatively fixed. Therefore, a pre-set analysis method is used to directly analyze the reasons for incorrect selections and provide improvement suggestions based on the question stem. For fill-in-the-blank questions, incorrect answers resulting from calculation errors or misunderstandings are generally fixed among a few incorrect results. Therefore, LLM pre-selection analysis is used to identify possible incorrect answers from students, and the lack of mastery of the knowledge points involved in the question is used as a fallback reason for the error. For problem-solving questions, because students' solution processes can vary greatly, multimodal LLM is directly used to analyze the error causes of students' handwritten solution steps.

[0060] The question-based ability labeling module first generates a quantifiable, calculable, and evolving ability assessment system based on a massive question bank across four subjects. Its core architecture is as follows: Figure 2 As shown. The main process is as follows:

[0061] The construction of the cognitive atomic operation library begins with building an interdisciplinary, fine-grained library of "cognitive atomic operations." This is not a macroscopic "ability," but rather the most microscopic thought process steps in problem-solving. It's similar to a step-by-step, elegant analysis. For example:

[0062] OP_001: Identify and extract known conditions; OP_002: Perform conversions between different unit systems; OP_003: Apply Newton's second law (F=ma); OP_004: Perform logical syllogistic reasoning; OP_005: Abstract mathematical patterns from concrete examples; OP_006: Identify auxiliary lines in complex graphs; OP_007: Evaluate the reasonableness of a conclusion (can contain hundreds of such atomic operations).

[0063] This atomic library is not entirely manually defined, but rather formed through supervised fine-tuning training on a massive amount of question analysis texts, and then calibrated by educational experts, thus possessing the ability to self-evolve.

[0064] The system deconstructs the problem-solving process and generates atomic operation sequences. For a new problem, the system does not directly determine its ability label, but first automatically generates its complete problem-solving process, i.e., a step-by-step, refined analysis. This includes:

[0065] Standard deconstruction: The standard answer and analysis of the question are decomposed into a step-by-step sequence of "cognitive atomic operations" (i.e., progressively refined analysis) through a finely tuned sequence generation model (this model refers to the atomic operation library that is finely tuned through a large number of question analysis texts).

[0066] Input: [Problem text], [Answer and explanation], Output: [0P_002], [0P_001], [0P_007], [0P_004] (a sequence of atomic operations).

[0067] Multi-path deconstruction: For open-ended problems, the model generates multiple possible and reasonable solution paths (i.e., multiple sequences of atomic operations).

[0068] The system maintains an "atomic operation-macro-ability" mapping matrix for the generation and calculation of dynamic ability vectors. This matrix defines the contribution weight of each "cognitive atomic operation" to various macro-abilities (such as logical reasoning, mathematical operations, etc.). The calculation of the dynamic ability vector involves taking all atomic operations appearing in the deconstructed "cognitive atomic operation sequences" of a given problem, weighted, summed, and normalized according to their weights in the mapping matrix. This ultimately generates a fixed-dimensional, quantified vector, referred to as the "dynamic ability vector" of that problem.

[0069] For example, a question might be characterized as [Logical Reasoning: 0.85, Mathematical Operations: 0.45, Intuitive Imagination: 0.10, ...]. Then, by setting a threshold, the ability components above the threshold are defined as the ability labels that the question possesses.

[0070] The knowledge point labeling module: The main process of the knowledge point labeling system is as follows. First, training data needs to be constructed for different textbook versions. Specific knowledge point labels can be assigned for a given subject, grade level, version, and whether the curriculum standards are new or old. (Here, the Milvus vector database specifies the version, subject, grade level, and curriculum standards when storing knowledge point information and example questions from the knowledge graph. The filter information provided in the data for the subject, grade level, version, and curriculum standards can directly recall the candidate knowledge point range for the specified subject, grade level, version, and curriculum standards.) Second, a training task is constructed, the core of which is an anchoring framework that integrates graph neural networks and multimodal reasoning. Finally, the output structure is defined, and the knowledge point labels for the specified version are obtained through the parsing model output. The core of the system is to construct a dual-channel graph neural network, the architecture of which is as follows: Figure 3 As shown.

[0071] Dynamic knowledge topology graph structure modeling: Here, the knowledge graph is no longer viewed as a static set of nodes, but rather as a dynamic, relation-rich graph structure. Each knowledge point is a node, and the edges between nodes represent various relationships:

[0072] Prerequisite relationship: Knowledge point A must be mastered before learning knowledge point B. Parallel relationship: Knowledge points A and B are usually learned in the same stage and are closely related. Instantiation relationship: Knowledge point A is a specific application or special case of knowledge point B. This graph structure is pre-trained but will be dynamically learned and adjusted during model training.

[0073] Dual-channel question encoding and topology awareness: For each question, it will be input into two channels for parsing. Before parsing, a massive amount of question information and knowledge point information is stored in the Milvus vector database. Based on the similarity of question embedding and knowledge point embedding, the top k similar questions and similar knowledge points from various textbook versions are retrieved first.

[0074] Content channels: 1) Input: Text of the question (question stem, answer, explanation) (images have been modified) and images (charts, formulas); 2) Encoder: Uses a multimodal model to obtain the deep representation vector of the question content.

[0075] Topology Channels: 1) Input: A series of candidate knowledge point nodes retrieved from the Milvus vector database. 2) Encoder: These candidate knowledge point nodes and their neighbor relationships (local topology) in the knowledge graph are input into a graph neural network. 3) Role of GNN: The GNN aggregates information about each candidate knowledge point and its neighbors, generating a "topological context-enhanced" representation vector for each candidate knowledge point. This vector contains not only the meaning of the knowledge point itself, but also its position and role in the knowledge network.

[0076] Topology-aware cross-attention matching calculates the matching degree between the question content vector and each topology-enhanced candidate knowledge point vector. A cross-attention mechanism is used here, allowing the question content to "query" each candidate knowledge point and its topological context. The model not only determines "whether this question resembles this knowledge point," but also "whether the ability structure tested by this question matches the position of this knowledge point in the network (such as required prerequisite knowledge and derived subsequent knowledge)."

[0077] For example, a complex mechanics problem may be directly similar to "Newton's second law", but through topological perception, the model will find that it is more accurately anchored on the node of "the application of Newton's second law in inclined plane problems", because it also strongly depends on the prerequisite knowledge point of "force decomposition".

[0078] The hierarchical path generation and output identifies the node with the highest matching score as the core knowledge point. The system doesn't simply output an isolated, bottom-level knowledge point; instead, it automatically generates a complete path from the root node to that knowledge point. For example, "Middle School - Mechanics - Newton's Laws of Motion - Newton's Second Law - Inclined Plane Applications" is not merely string concatenation, but rather a deep reasoning result from the model's understanding of the question's location within the knowledge system.

[0079] To prevent predicted knowledge points from not existing in a specified version, a post-recovery logic is set up. For abnormal outputs, the embedding vector of the knowledge tree for the specified version is recalled based on the model prediction results to find the most similar knowledge point for the specified version and correct the result.

[0080] like Figure 4 As shown, the OCR recognition module: This module constructs an OCR system capable of understanding the student's problem-solving cognitive process. It uses a spatiotemporal-semantic dual-channel reasoning framework to infer the "writing flow" and reconstruct the authentic answer content. This framework primarily employs a dual-channel architecture, simultaneously processing the physical attributes and semantic intent of the handwriting.

[0081] Handwriting element deconstruction and multimodal decoding: First, advanced visual models (such as YOLO) are used to perform fine-grained analysis of the answer area beyond the text box: Text fragment detection: Identifying all coherent blocks of handwritten text. Graphic symbol detection: Specifically detecting non-text markers such as arrows, crosses, caret, underlines, and circles. Mathematical formula detection: Locating and isolating areas of mathematical expressions. Spatial relationship decoding: Recording the absolute and relative positions between all elements.

[0082] The reasoning and reconstruction of cognitive writing flow introduces a key problem-solving logic reasoning module. This module is based on the semantic information of the question itself, and its working principle is as follows:

[0083] Input: The text content of the question stem, the solution steps of the standard answer, and all handwriting elements extracted from the steps.

[0084] Inference process: (This can be mainly broken down into the following steps, using prompts to facilitate collaboration among agents from multiple large models:)

[0085] Constructing a problem-solving logic graph: The system does not reason from scratch. It pre-analyzes the problem itself using a large model, automatically breaking down the standard answer into a structured "standard problem-solving logic graph." This graph includes key steps, logical relationships between steps (such as "dependencies" and "derivations"), and the expected mathematical expressions and conclusions for each step.

[0086] Alignment of handwriting with logic: When recognizing student handwriting, the system not only records "what it is," but also "where it is written" and "in what order." For example, it recognizes a crossed-out formula F=mg in region A and an arrow pointing to F=ma in region B pointing to region C.

[0087] Hypothesis generation and verification: The reasoning module will propose the hypothesis "Did the student initially use the gravity formula incorrectly (region A), then realize the error, and correct it to Newton's second law (region C) via the arrow (region B)?"

[0088] Next, the system verifies this hypothesis: semantic consistency: F=mg and F=ma are both valid formulas in physics problems, but their contexts differ. Logical coherence: the "therefore" relationship indicated by the arrows; whether there is a corresponding logical edge from "force analysis" to "application of Newton's laws" in the standard problem-solving diagram; spatial evidence: the crossed-out content is usually a priori incorrect attempts, which is consistent with the hypothesis.

[0089] Generate a cognitive writing sequence: Based on the most reasonable assumptions, the system no longer outputs [Region A: F=mg]->[Region B]->[Region C: F=ma] according to physical location, but instead reconstructs a logical sequence: [Incorrect attempt: Applying the gravity formula F=mg]->[Key derivation: Applying Newton's second law F=ma]. Each element in this sequence is assigned a semantic role.

[0090] The system will simulate the thinking of a "problem-solving mentor" and ask a series of questions to deduce the student's writing flow: "To find the final answer, what do you usually need to calculate first?" (the order of reasoning steps), "Is this crossed-out formula an incorrect method?", "Does this arrow indicate a logical derivation relationship of 'therefore' and 'so'?", "This isolated calculation is a draft for which main step?"

[0091] Writing Flow Reconstruction: Based on the above reasoning, the system no longer simply sorts text by spatial location, but reconstructs a "cognitive writing sequence" that conforms to the problem-solving logic. For example, it organizes the content on a messy answer sheet as: [Step 1: Set the unknown] - [Draft: Try formula A (crossed out)] - [Step 2: Apply formula B] - [Step 3: Substitute and calculate] - [Final answer].

[0092] Intent-guided collaborative recognition differs from traditional OCR, which recognizes text in isolation. In this system, the recognition process is guided and constrained by a "cognitive writing flow." Each element in the reconstructed "cognitive writing sequence," along with its inferred "semantic role" (such as "key derivation steps," "auxiliary calculations," and "final conclusion"), is input into a specially trained multimodal LLM.

[0093] Specialized training methods for multimodal LLM include:

[0094] Building the Training Data Set: A large-scale, high-quality dataset of handwritten problem-solving images and text pairs will be artificially constructed. Data construction methods include: Real-world collection: Extensive collection of handwritten math, physics, chemistry, and biology problem-solving images from students. Synthetic generation: Using handwritten fonts and a formula rendering engine, images with various handwriting styles, corrections, and markings are synthesized according to standard problem-solving steps. Data annotation: The annotation for each image goes beyond simply transcribing text; it is annotated in a "cognitive writing sequence" format. For example, a data sample might be:

[0095] The labeled text reads: "Step 1, Define the unknown: Let the mass of the object be m. Step 2, Apply the core theorem: According to Newton's second law, F=ma. Step 3, Substitute and calculate: Substituting the data, we get 10=5a. Final answer: a=2." Additionally, the semantic meaning of non-text elements such as corrections and arrows must be clearly marked in the data.

[0096] SFT Task: During training, the instructions given to the model are carefully designed, such as: "You are a physics tutor. Please identify the following handwritten problem-solving process and reconstruct the student's writing flow according to the logical steps. Note: Crossed-out content represents incorrect attempts, and arrows indicate derivation relationships." The training goal of the model is to learn to output a "cognitive writing sequence" in the above format based on the image and instructions.

[0097] Iterative Optimization and Alignment: Using a method similar to Direct Preference Optimization (DPO), educational experts rank the model's multiple outputs, for example, which output is more logical and which accurately identifies the key formula. Through this training, the model learns to prioritize outputs that are logically sound and conform to problem-solving standards, rather than simply those that look most visually similar, i.e., a student's handwritten solution to a math problem. Its semantic role is "key derivation step." Please accurately identify the handwritten text and formulas, paying particular attention to the rigor of the derivation. If the handwriting is blurry, please make reasonable inferences based on the problem-solving logic.

[0098] This intention guidance enables the model to prioritize characters that conform to mathematical logic when recognizing ambiguous characters. For example, it may recognize ambiguous symbols as "=" instead of "-", and illegible characters as "✖" instead of multiplication signs.

[0099] The system employs dynamic confidence assessment and interactive repair. For each identified segment, the system outputs a confidence score that integrates visual clarity and consistency with the reasoning logic. For segments with low confidence scores but at critical reasoning steps, the system activates an interactive repair mechanism: automatically generating several most likely candidate recognition results. The reasoning module is then invoked again to determine which candidate result best fits the overall problem-solving logic flow and is selected first.

[0100] The input to the grading module is the question stem text, the question itself extracted from the MLLM information, and the student's answer, which is the final answer text processed by the OCR recognition module and fused with the handwritten MLLM answer information. The grading status serves as a preset status indicator to indicate the rigor or focus of the grading (e.g., "rigorous mode" strictly examines the steps, while "results-oriented mode" focuses more on the final answer). The specific process is as follows:

[0101] Multimodal semantic encoding and alignment: Pre-trained language models (such as BERT and DeBERTa) are used to encode the question stem, answer, and explanation, as well as the student's response, respectively, resulting in their respective semantic representation vectors. To accurately compare the differences between the student's response and the standard answer, the module employs a cross-attention mechanism. Specifically, the vector representation of the student's response is used as the query to query (attend to) the vector representation of the explanation (as the key and value), thus calculating a semantic alignment representation. This representation focuses on the parts of the student's response that are most relevant to the explanation.

[0102] Rule-based and reasoning-based grading decisions: Result correctness judgment: For multiple-choice and fill-in-the-blank questions, the student's text is directly compared to the answer key. For problem-solving questions, the correctness of the final answer is judged by calculating the cosine similarity between the semantic vector of the student's final answer and the standard answer vector, combined with keyword matching. Process rationality analysis: This is crucial for grading problem-solving questions. The module inputs the student's semantically aligned representation into a classifier, which, after training, can identify whether the answer contains key problem-solving steps, whether there are conceptual errors, and whether the logic is coherent. Simultaneously, the module calls a lightweight symbolic reasoner or mathematical engine to perform identity transformation verification or numerical verification on the mathematical expressions in the student's answer to judge the correctness of the calculation process. Step-by-step score mapping: The system pre-stores the score allocation scheme for each step of the question. Based on the results of the process analysis, the grading module locates the specific step where the student's error occurred and automatically deducts the corresponding score for that step, generating a step score.

[0103] Confidence Calculation and Output: The grading module comprehensively considers semantic matching, rule matching, and inference uncertainty to output a confidence score for this grading. The final grading result is structured data, including question scores and student scores. This result will then be passed to the error analysis module for in-depth error analysis and attribution.

[0104] The student error analysis module combines knowledge points and abilities as factors in the analysis of problem-solving questions, thereby determining which knowledge points students lack and which abilities they are deficient in. This module also employs different processing strategies for different question types, as follows: Figure 5 As shown.

[0105] Through the previous three systems, we can obtain the question information (question stem, answer, explanation, knowledge points, ability level, and score), the student's answers, and the student's score sheet (output from the grading interface). This module will first compare the student's score with the question's score to determine if the student answered the question correctly. If correct, it will output an empty error message; if incorrect, it will call the model to analyze the error cause. Specifically:

[0106] Firstly, for multiple-choice questions, since the options are fixed and students will not answer more than four options, students who choose the same incorrect option likely have the same reason for their mistake. Therefore, we pre-generate the reason for each incorrect option based on the question information and insert it into the MongoDB database. When analyzing students' error reasons in batches, we directly query the database based on the students' incorrect options and then match the corresponding error reason results. The prompt for the multiple-choice questions is as follows:

[0107] Your task is to analyze the error types of each option in a math multiple-choice question, and for multiple-answer questions, you also need to analyze why the correct options were not selected.

[0108] Please follow these steps:

[0109] 1. Read and understand the given questions, answers, and explanations.

[0110] 2. Analyze each incorrect option, put yourself in the student's shoes, think about why the incorrect option was selected or the correct option was missed, and then associate it with the most likely type of error.

[0111] 3. The error type must be specific and cannot be vague such as "option error" or "selection error". Only the most likely error type needs to be given.

[0112] 4. For multiple-choice questions, analyze each option. For correct options, analyze why students might miss them, and for incorrect options, analyze why students might choose them incorrectly. For single-choice questions, only analyze the incorrect options.

[0113] 5. Mathematical Formatting: All mathematical formulas, symbols, and expressions appearing in the question content must be transcribed in LaTeX format and enclosed in a single dollar sign.

[0114] 6. Error types must be described in Chinese; English is strictly prohibited.

[0115] Answer format: Please answer using the following format:

[0116] Option: A;

[0117] Error type: Error type (no need to specify whether it is an incorrect selection or an omission, just give the type);

[0118] Description: A description of the error type;

[0119] Recommendation: Recommendations for different error types.

[0120] Question: <|question|>, Answer: <|answer|>, Analysis: <|analysis|>;

[0121] Be sure to follow the given output format, enclosing each key field with ##, and do not add any other characters;

[0122] For multiple-choice questions, analyze options A, B, C, and D. For correct options, analyze why you missed selecting them; for incorrect options, analyze why you chose them incorrectly.

[0123] For multiple-choice questions, only the three incorrect options are analyzed;

[0124] If the question asks which of the following statements are correct, and the answer is AC, note that this is a multiple-choice question. You need to analyze the four options: A, B, C, and D. For AC, analyze why you missed selecting them, and for BD, analyze why you selected them incorrectly.

[0125] If the question asks which of the following statements is incorrect ( ) or which statement is wrong ( ), and the answer is B, you should analyze why the student incorrectly chose A, C, or D, and conclude that their statements are incorrect.

[0126] The processing logic for single-choice and multiple-choice questions is different here. For single-choice questions, only the three incorrect options are analyzed, while for multiple-choice questions, all four options must be analyzed. For incorrect options, we analyze why they were chosen incorrectly, and for correct options, we analyze why they were missed.

[0127] Secondly, since the knowledge points tested in fill-in-the-blank questions are relatively fixed, students' incorrect answers mostly fall into a few fixed categories. Therefore, a pre-set approach is also adopted, prompting the LLM to analyze the question information to identify possible incorrect answers from students and then insert the pre-set results into the MongoDB database. In addition, a fallback reason for error is set up, such as a lack of mastery of the knowledge points involved in the question. When the student's actual answer does not match all the incorrect answers given by the model, the fallback reason is given by default. The prompt for fill-in-the-blank questions is as follows:

[0128] Your task is to analyze the possible types of errors and possible incorrect answers in math fill-in-the-blank questions.

[0129] Please follow these steps:

[0130] 1. Read and understand the given questions, answers, explanations, and relevant knowledge points.

[0131] 2. Analyze the possible wrong answers students may give based on the content of the question, the answer, and the knowledge points being tested. Put yourself in the students' shoes and think about the wrong ways students might get results that are similar to but inconsistent with the correct answer, and then associate them with the most likely error types.

[0132] 3. For possible incorrect answers, the error type must be specific and cannot be vague such as "calculation error" or "logical error". Only the most likely error type needs to be given.

[0133] 4. In addition, you should provide a fallback error analysis, that is, based on the knowledge points in the question, give the error type such as "weak in xx knowledge point", and give corresponding guidance and suggestions.

[0134] 5. Mathematical Formatting: All mathematical formulas, symbols, and expressions appearing in the question content must be transcribed in LaTeX format and enclosed in a single dollar sign.

[0135] 6. Error types must be described in Chinese; English is strictly prohibited.

[0136] Be sure to provide as many possible wrong answers as possible, and provide a fallback error type that can be attributed to a weak understanding of the relevant knowledge point.

[0137] Answer format:

[0138] Please answer in the following format:

[0139] Incorrect answer: 1;

[0140] Error type: Error type 1 (details of the error type are given directly);

[0141] Description: Description of the error type 1;

[0142] Recommendation: Recommendation 1 for the error type;

[0143] Incorrect answer: frac{1}{2};

[0144] Error type: Error type 2 (details of the error type are given directly);

[0145] Description: Description 2 for error types;

[0146] Recommendation: Recommendation 2 for the error type;

[0147] The format of the fallback result:

[0148] Incorrect answer: <|error|>;

[0149] Error type: Error type (directly provide an error type such as "weak in xx knowledge point");

[0150] Description: A description of the error type;

[0151] Recommendations: Recommendations based on error type;

[0152] Question, Answer, Analysis, Key Concepts;

[0153] Note that incorrect answers should only provide the result of the incorrect answer, without any other extra textual descriptions;

[0154] Note that some fill-in-the-blank questions have multiple blanks. You need to analyze each blank according to the number of blanks that can be filled with the correct answer, and give some possible wrong answers for each blank.

[0155] Fill-in-the-blank questions may have multiple blanks. For multiple blanks, a preset error result will be given for each blank. Similarly, when processing students' OCR answers, the format of students' handwritten answers will be split to distinguish each student's answer for each blank. However, sometimes students do not answer or their answers are too messy to be recognized. In this case, the system will give a "not recognized answer" result to prevent the output from being empty.

[0156] Finally, for the problem-solving questions, since students may use various methods to solve them, a multimodal large model is used to directly analyze the students' answer sheet images. By comparing the answer explanations with the students' answering processes, LLM inference is used to pinpoint the specific problem-solving steps where students made mistakes, and detailed reasons for the errors are provided. The prompt for the problem-solving questions is as follows:

[0157] Your task is to analyze the types of errors students make in solving math problems.

[0158] Please follow these steps:

[0159] 1. Correlation analysis of scores:

[0160] Determine the severity of errors based on student scores:

[0161] A score of ≥90% (out of 5): Focus on minor errors (such as calculation errors / omission of units, failure to perform verification, etc.).

[0162] 50% ≤ Score < 90%: Analyze key steps for errors and knowledge gaps;

[0163] Score <50%: Systematic lack of understanding of basic concepts and problem-solving methods;

[0164] If a student scores almost full marks but loses only one or two points and the result is correct, note that the student may have forgotten to check their work or some steps may be missing, or the solution steps may be non-standard or skipped.

[0165] 2. Check the answers to the sub-questions. If the answer to a sub-question is correct, then that sub-question does not need to be analyzed.

[0166] 3. Identify the errors: You can only find the actual errors based on what the student wrote down; you cannot guess. Furthermore, you must analyze each student's incorrect answers by combining the question's ability tags and the relevant knowledge points.

[0167] Compare the student's answers with the correct answers to identify where the student made mistakes. The mistakes should be detailed, such as incorrect application of formula (specifically how the formula was applied incorrectly), flawed problem-solving approach (e.g., the student used the system of equations instead of establishing a coordinate system), or incorrect calculation of value (where the calculation started incorrectly).

[0168] 4. For each error point, summarize the cause of the error, classify the error type, and provide corresponding suggestions. The error type must be described in Chinese and must not contain English. The student's answer may be confusing due to OCR recognition errors. Pay attention to distinguishing between the reasonableness of the student's actual answer and the situation of OCR recognition error.

[0169] 5. Mathematical Formatting: All mathematical formulas, symbols, and expressions appearing in the question content must be transcribed in LaTeX format and enclosed in a single dollar sign.

[0170] Formatting requirements:

[0171] Error type: Error type 1;

[0172] Description: Description 1 for error type 1;

[0173] Recommendation: Recommendation 1 for error type 1;

[0174] Error type: Error type 2;

[0175] Description: Description 2 for error type 2;

[0176] Recommendation: Recommendation 2 for error type 2;

[0177] Question, Answer, Analysis, Key Concepts, Ability;

[0178] Full marks for the question:

[0179] Student score <|student_score|>;

[0180] The error type should be concise and not too long, and should be kept within 20 characters. Detailed error descriptions should be provided in the description.

[0181] If a sub-question is not answered, or only the sub-question number, the word "solution", or the formula is listed but no calculation or derivation is performed, directly give the error type as "unanswered (2)" or "unanswered (1)", and do not give any other form of error explanation.

[0182] Reordering: After providing all the error causes, the final list of error causes must be reordered.

[0183] The ranking principle is as follows: knowledge-based and conceptual errors > errors in problem-solving methods and procedures > calculation and operational errors.

[0184] Ultimate goal: Ensure that error type 1 in the output represents the most fundamental and critical knowledge or conceptual errors made by the student.

[0185] Note that students' answer sheets may have scribbled or altered parts. Be careful to distinguish them and do not forcibly identify and include the content that students have scribbled out in the analysis.

[0186] Similarly, some students may leave their answers blank or only write the word "Solution" or the question number. In these cases, a judgment logic will be added to directly output "Answer not recognized". In addition, many difficult problem-solving questions have multiple sub-questions. Here, the LLM will also be prompted to pay attention to the student's answers to the sub-questions. The student's error may only appear in the details of a certain sub-question. The error reason will be given in real time based on the student's score sheet to avoid the situation where the student loses 1-2 points due to details, but the model gives multiple error reasons. This will enable accurate analysis of the student's test paper.

[0187] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An AI large model-based student mistake reason induction test paper analysis system, characterized by, The method comprises the following steps: An image acquisition module is configured to acquire a question image; A question positioning module is configured to process the question image and extract a question frame of the preprocessed question image; An information extraction module is configured to extract MLLM information and OCR information of the question frame, mark a question ability and a question knowledge point in the MLLM information, generate an ability label and a knowledge point path label of a current question, fuse the OCR information and handwritten answer content of the MLLM information, and correct the fusion result in combination with the question and correction state of the MLLM information; The information extraction module comprises: An information extraction submodule is configured to extract MLLM information of the question frame; An ability marking submodule is configured to decompose a standard answer and analysis of the question in the MLLM information into a cognitive atomic operation sequence, mark a question ability of the cognitive atomic operation sequence, and generate an ability label of the current question; The ability marking submodule comprises: An ability marking unit is configured to map a cognitive atomic operation in the cognitive atomic operation sequence to a mapping matrix of atomic operation-macroscopic ability, generate a contribution weight of each cognitive atomic operation to each macroscopic ability, weight and summarize the contribution weight, normalize the contribution weight, compare the ability vector with a preset ability threshold, and take a macroscopic ability corresponding to the ability vector higher than the preset ability threshold as the ability label of the current question; A knowledge point marking submodule is configured to mark a knowledge point of a specified subject section of a textbook version to which the question in the MLLM information is directed, and generate the knowledge point path label; The knowledge point marking submodule comprises: An encoding and topology perception unit is configured to recall a preset number of similar questions and similar knowledge points of each textbook version according to the similarity of the question to knowledge points in a Milvus vector database; A deep representation vector of the question content in the text and image of the question is extracted, and a candidate knowledge point node is recalled from the Milvus vector database; The candidate knowledge point node and its neighbor relationship in a knowledge graph are input into a graph neural network model to obtain a topology enhanced representation vector; A cross-attention mechanism is used to calculate the matching degree between the deep representation vector and each topology enhanced representation vector, the node with the highest matching degree is taken as a core knowledge point, and the core knowledge point is taken as a knowledge point core to generate a complete path from a root node to a terminal knowledge point, and the complete path is taken as the knowledge point path label of the current question; An OCR recognition submodule is configured to adopt different OCR recognition strategies for different question types to perform OCR recognition on the question frame, and generate the OCR information; 2.The AI-large model-based student mistake reason induction test paper analysis system of claim 1, wherein, A mistake cause analysis module is configured to perform mistake cause analysis according to a correction result in combination with the ability label and the knowledge point path label of the current question, and obtain a mistake cause result. The information extraction module comprises: A correction submodule is configured to fuse the OCR information and the handwritten answer content of the MLLM information, and correct the fusion result in combination with the question and correction state of the MLLM information. 3.The AI-large model based student mistake reason induction test paper analysis system of claim 2, wherein, The ability marking submodule comprises: An atomic operation generation unit configured to decompose a standard answer and a solution of a question in the MLLM information into a sequence of cognitive atomic operations by using a cognitive atomic operation library, the cognitive atomic operation library containing the most microscopic thinking steps in a problem solving process. 4.The AI-large model based student mistake reason induction test paper analysis system of claim 2, wherein, The knowledge point marking submodule comprises: A knowledge topology unit configured to construct a graph structure of a dynamic knowledge topology to generate a knowledge graph, the knowledge graph comprising nodes corresponding to each knowledge point and edges generated by various relationships between the nodes, the relationships comprising prerequisite relationships, parallel relationships, and instantiation relationships. 5.The AI-large model based student mistake reason induction test paper analysis system of claim 4, wherein, The knowledge point marking submodule further comprises: A result correction unit configured to set a recall post-repair logic if a knowledge point does not exist in a specified textbook version of a subject, perform embedding vector recall on a knowledge tree of the textbook version for the complete path, determine a most similar knowledge point of the specified textbook version, and correct an output result of the coding and topology perception unit. 6.The AI-large model based student mistake reason induction test paper analysis system of claim 2, wherein, The OCR recognition submodule comprises: A first OCR recognition unit configured to input fill-in-the-blank questions and answer questions of the question frame into a text detection model and an OCR recognition model respectively to obtain handwriting circle drawing and font recognition results, the text detection model and the OCR recognition model being trained by using a training set; A second OCR recognition unit configured to call a multi-modal large language model to recognize a student handwriting solution process in the answer question of the question frame. 7.The AI-large model based student mistake reason induction test paper analysis system of claim 6, wherein, The second OCR recognition unit comprises: A handwriting element deconstruction subunit configured to recognize all coherent handwriting text blocks and non-text markers, isolate mathematical expression regions, and further record absolute positions and relative positions between all handwriting elements, i.e., absolute positions and relative positions between the handwriting text blocks, the non-text markers, and the mathematical expression regions; A cognitive reasoning and reconstruction subunit configured to perform problem solving reasoning by using text content of the question in the question frame, problem solving steps of the standard answer, and all handwriting elements; A writing flow reconstruction subunit configured to reconstruct a cognitive writing sequence in accordance with problem solving logic based on the problem solving reasoning result; An OCR recognition subunit configured to input the cognitive writing sequence into the multi-modal large language model to obtain a handwriting solution result. 8.The AI-large model based student mistake reason induction test paper analysis system of claim 1, wherein, The error cause analysis module comprises: An error cause analysis submodule configured to generate error causes of each incorrect option in a multiple-choice question and a fill-in-the-blank question in advance according to question information and insert the error causes into a MongoDB database, directly query the MongoDB database according to an incorrect option of a student when analyzing error causes of the student, and match a corresponding error cause result of the incorrect option; When analyzing the fill-in-the-blank question, a bottom error cause is set, i.e., when the student's fill-in-the-blank question cannot match all incorrect answers when actually answering, a bottom error cause is given by default; The solution content of the answer question is input into a multi-modal large model, the answer solution of the question and the student's answering process are compared, specific problem solving steps in which the student makes a mistake are located, and detailed error causes are obtained in combination with the ability label and the knowledge point path label of the current question.

Images