Test question repeatability detection method and system based on multi-modal large language model

Abstract

A test question repeatability detection method and system based on a multi-modal large language model, comprising: collecting multi-modal test question data as a question bank data set, and performing data cleaning on the question bank data set to obtain a cleaned data set; for the cleaned data set, generating a multi-modal high-dimensional semantic vector representation of the test question in combination with a multi-modal large language model, and obtaining a final vector representation of the test question after pooling the multi-modal high-dimensional semantic vector representation; constructing a Milvus collection and an HNSW index in combination with test question data exchange format JSON data for the final vector representation of the test question; and combining the multi-modal representation of the search set with the HNSW index of the Milvus to visually display the response result. The present application is based on the Milvus vector database and multi-modal representation technology, so that the present application can support similarity search scoring of a million-level question bank with millisecond-level more accurate response.

Description

Technical Field

[0001] This invention belongs to the field of multimodal retrieval in educational examination scenarios, specifically involving a method and system for detecting test question repetition based on a multimodal large language model. Background Technology

[0002] Question duplication detection technology aims to improve the efficiency of duplication detection and optimize the allocation of educational resources by comprehensively understanding multimodal information such as question text, formulas, and charts, thereby identifying questions with semantic or logical repetition. However, existing technologies face significant challenges in terms of modal coverage, depth of multimodal semantic understanding, and efficiency, specifically manifested in the following ways: 1. Insufficient modal coverage: Modal information such as formulas and charts requires manual annotation or separate processing, lacking a unified representation. Furthermore, it cannot effectively integrate the textual descriptions, diagrams, and knowledge point structures of test questions, resulting in insufficient cross-modal similarity detection capabilities.

[0003] 2. Weak depth of multimodal semantic understanding: Traditional models have limited ability to understand the semantics of multimodal information such as complex mathematical symbols and geometric figures.

[0004] 3. Inefficiency: Traditional relational databases such as MySQL are unable to support the real-time retrieval needs of massive amounts of test questions, resulting in slow response speeds. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for detecting test item repetition based on a multimodal large language model, so as to solve the above-mentioned problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a test item repetition detection method based on a multimodal large language model, comprising: Collect multimodal test item data as a question bank dataset, and clean the question bank dataset to obtain the cleaned dataset; For the cleaned dataset, a multimodal high-dimensional semantic vector representation of the test questions is generated by combining a multimodal large language model, and the final vector representation of the test questions is obtained by pooling the multimodal high-dimensional semantic vector representation. The final vector representation of the questions is used to construct a Milvus set and an HNSW index by combining the JSON data exchange format of the test questions; The multimodal representation of the query set, combined with Milvus's HNSW index, allows for the visualization of the response results.

[0007] Furthermore, the collection of multimodal test item data as a question bank dataset includes: The dataset for the question bank is compiled from subject textbooks and online question banks containing open-ended test questions. A tree-like index structure is constructed according to the subject knowledge points, and the raw data is stored in JSON format. .

[0008] Furthermore, the data cleaning of the question bank dataset to obtain the cleaned dataset includes: For the question dataset, regular expressions were used to remove HTML tags, and a LaTeX parsing tool was used to standardize the formula format and option number format, resulting in a cleaned question bank dataset. .

[0009] Furthermore, the step of generating multimodal high-dimensional semantic vector representations of test questions using a multimodal large language model for the cleaned dataset includes: The cleaned test data was split into three independent modal channels: text, image, and knowledge point. The text modality included the plain text content of the question stem and options; the image modality associated with diagrams or formula images in the test questions; and the knowledge point modality was mapped to the leaf node level in a tree-like index structure. A 7B-parameter multimodal large language model was used. The three modal data were input into the model through a preset instruction template, and the multimodal features were aligned and fused across modalities based on a multi-head attention mechanism to generate high-dimensional semantic vector representations. .

[0010] Furthermore, the final vector representation of the problem obtained after pooling the multimodal high-dimensional semantic vector representation includes: Representation of the generated high-dimensional semantic vector conduct The operation compresses the high-dimensional vector into a fixed-length low-dimensional vector to preserve the core semantic information, thus obtaining the final vector representation of the problem. .

[0011] Furthermore, the final vector representation of the questions, combined with the JSON data exchange format for the test questions, constructs a Milvus set and an HNSW index, including: Based on the pooled multimodal representation vector set A Milvus database collection was constructed by combining the JSON metadata of the test questions, with fields defined as test question ID, question stem text, image path, knowledge point, JSON metadata, and multimodal vector. The vector field was optimized for retrieval using the HNSW index, with the index parameters set as distance metric COSINE, index construction parameter number of adjacent nodes M=64, and search width.

[0012] Furthermore, the multimodal representation of the search set, combined with Milvus's HNSW index, visualizes the response results, including: Fast deduplication detection of multimodal test questions is achieved through vector cosine similarity calculation, based on the vector representation of the query set and the standardized vector. The system uses the HNSW index of the Milvus database, sets parameters for querying, and generates an interactive interface through a visualization engine to display the content of the top 100 similar questions.

[0013] Secondly, the present invention provides a test item repetition detection system based on a multimodal large language model, comprising: The data acquisition module is used to collect multimodal test question data as a question bank dataset, and to clean the question bank dataset to obtain a cleaned dataset. The cleaning and pooling module is used to generate multimodal high-dimensional semantic vector representations of test questions from the cleaned dataset by combining a multimodal large language model, and to obtain the final vector representation of the test questions by pooling the multimodal high-dimensional semantic vector representations. The optimization module is used to construct the final vector representation of the questions, and to build a Milvus set and an HNSW index by combining the JSON data exchange format of the test questions; The output module is used to visualize the response results by combining the multimodal representation of the data set with Milvus's HNSW index.

[0014] Thirdly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the test item repetition detection method based on a multimodal large language model.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the test item repetition detection method based on a multimodal large language model.

[0016] Compared with the prior art, the present invention has the following technical effects: The purpose of this invention is to provide a test question duplication detection method based on a multimodal large language model. By integrating multimodal data such as text, images, and knowledge points, and combining cross-modal semantic alignment and efficient indexing techniques, it achieves rapid duplication detection for large-scale question banks. First, the test question data undergoes multimodal cleaning. A 7B-level multimodal large language model is used to generate high-dimensional multimodal semantic vector representations of the questions, which are then compressed into low-dimensional vectors using mean pooling. Simultaneously, an HNSW index based on the Milvus database is constructed. Combined with similarity calculation, the method retrieves the Top 100 similar questions within milliseconds, overcoming the shortcomings of traditional methods that ignore image semantics and modal correlations, significantly improving the efficiency and accuracy of duplication question detection.

[0017] This invention utilizes a 7B-parameter multimodal large model to automatically generate high-dimensional semantic vectors, eliminating the need for manual rule design or lexicon maintenance, and adapting to formula variations and question type evolution. Based on the Milvus vector database and multimodal representation technology, this invention can support more accurate similarity retrieval scoring with millisecond-level responses for millions of questions. Attached Figure Description

[0018] Figure 1 This is a flowchart of the test item repetition detection method based on a multimodal large language model according to the present invention.

[0019] Figure 2 This is a flowchart of multimodal data crawling and cleaning.

[0020] Figure 3 This is a flowchart of multimodal data representation.

[0021] Figure 4 This is a flowchart of the pooling characterization process.

[0022] Figure 5 This is a flowchart of vector database index optimization.

[0023] Figure 6 This is a flowchart of multimodal similarity retrieval. Detailed Implementation

[0024] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and examples. It should be noted that the embodiments described herein are only for explaining the present invention and are not intended to limit the present invention. Furthermore, the technical features involved in the embodiments of the present invention can be combined with each other unless otherwise specified.

[0025] The specific implementation process of this invention includes multimodal test data acquisition, multimodal data cleaning, multimodal data representation, pooling representation, vector database index optimization, and multimodal similarity retrieval. Figure 1 This is a flowchart of the test item repetition detection method based on a multimodal large language model according to the present invention.

[0026] Multimodal test data acquisition process To build a high-quality multimodal test question bank, data needs to be collected from multiple channels, covering various formats such as text and images, while also acquiring structured data related to the test questions, such as difficulty level, year, and subject tags. First, textbooks for the subjects to which the open test questions belong are collected, covering all subjects from elementary to high school, including Chinese, mathematics, English, physics, chemistry, biology, history, geography, and politics, integrating the test question content from these textbooks. Next, web crawling technology is used to collect information from multiple data sources, including open-source subject question banks, educational resource databases, e-book databases, and academic resource databases. The data includes: Text: Question stem, options, explanation Images: function graphs, geometric graphs, etc. Structured data: difficulty level, year, region, knowledge point tags, etc.

[0027] The obtained test data is organized hierarchically by subject knowledge point, and a subject knowledge point tree index is constructed as follows: Mathematics (Root) ├─ Algebra and Elementary Functions │└─ Function Basics │├─ Linear function │└─ Quadratic Functions → Vertex Form / Graph Properties / Extremum Problems (Leaf Nodes) └─ Geometry and Analytic Geometry ├─ Plane Geometry └─ Solid Geometry → Spatial Vectors / Surface Area of ​​a Sphere (Leaf Nodes) The JSON format of each question is stored as follows: { "question_id": "0002091928132ebfc181a3299413df54", "paper_id": "1020#", "question_type": "Solve question", "source": "Autumn 2022, Zhangjiakou, Hebei, Senior Three Unified Examination Final Exam", "stem": "Given a square pyramid \(P-ABCD\). Find the cosine of the angle between plane \(BEC\) and plane \(PAD\)". "image": [], "choices": [], "url": "https: / / xxx.com / 11q17801498.html", "province": "Hebei", "md5": "f5ee32366c4be748ed21850d1fdca84a", "question_year": "2024", "knowledge": "Proving line-plane perpendicularity, proving line-line perpendicularity, vector method for finding plane-plane angle", "difficulty": "moderate" "difficulty_score": "0.65", "sub_questions": [{"stem": "(1) Prove the perpendicular relationship between a line and a plane;", "choices": []}], "sub_questions_count": 1, "knowledge_expanding": "Application of space vectors, vector method for spatial angles, vector method for dihedral angles, space vectors and solid geometry", "item_id": 84879 } The original data is stored in JSON format, and the original data fields include question ID, question stem, options, knowledge point tags, image paths, difficulty, year, etc., thus obtaining the original question bank dataset .

[0028] Multimodal data cleaning process In the data preprocessing stage, a series of cleaning operations are performed on the question dataset obtained in step 1 to ensure the quality and consistency of the data. First, regular expressions are used to match and remove HTML tags, and pure text content is extracted. Next, a LaTeX parsing tool is used to perform unified formatting processing on formulas, and a conversion function is defined , where S is the original text, is the standard text:

[0029] Rule set is as follows:

[0030] In terms of processing the option numbering format, unify the numbering format of all options to ensure their consistency. Unify numbers in different formats such as "1.", "(1)", "1.", etc. into the standard format "1.". The cleaned question bank dataset is obtained .

[0031] The above flow chart of multimodal data crawling and cleaning process is shown in Figure 2 Multimodal data representation process Split the cleaned question data into three independent modal channels: text, image, and knowledge point: Text modality: extract the pure text content of the question stem and options; Image modality: extract images related to the question, such as geometric schematic diagrams, function graphs, etc.; Knowledge point modality: extract knowledge point tags related to the question.

[0032] ​Subsequently, a multimodal standard prompt template was constructed, integrating text, images, and knowledge point information into a single prompt. Then, the images were first input into a 7B-level multimodal large language model and processed using an image-aligned ViT model to extract image features.

[0033] in, It is the input image. It is the ViT visual encoder of MLLM, which outputs visual features. Next, the visual features are projected onto the language embedding space through a linear layer:

[0034] here, It is a trainable projection matrix that maps visual features Transformed into visual tags aligned with the language model embedding space .

[0035] Embedding text and visual markers Together, they form a unified multimodal embedding:

[0036] The aggregated multimodal embedding is then performed. In the large language model that serves as the foundation for MLLM, multimodal information is aligned through self-attention and cross-attention mechanisms:

[0037] in, Through parameter matrix Multimodal information interaction. The specific prompt template is as follows: Observe the geometric figure in the image provided. Carefully analyze and describe all important geometric elements, including: 1. Shapes: Identify all geometric shapes such as circles, semicircles, triangles, quadrilaterals, etc. 2. Special points: Describe the positions of vertices, intersections,midpoints, and any other notable points. 3. Lines and angles: Discuss the relative positions of line segments,parallel or perpendicular lines, and the measures or relationships of angles. 4. Additional properties: Identify any symmetries, congruencies, orother notable geometric relationships. Question Context: <stem> Image: Knowledge: <knowledge> Based on the provided image, context, and knowledge points, provide adetailed analysis of the geometric figure. Your response should be structured and logically explain all observations in relation to the question. The flowchart of the above multimodal data representation process is as follows: Figure 3 As shown.

[0038] Pooling characterization process High-dimensional semantic vectors generated by a multimodal large language model (MLLM) It contains rich semantic information, but has high dimensionality and computational complexity. First, padding tokens are removed, retaining only valid semantic information. Then, [the following is used:] Pooling operations generate a fixed-length low-dimensional vector by calculating the element-wise average of a vector sequence. Specifically, it is defined as follows:

[0039] in, It is a high-dimensional semantic vector The One element, It is the dimension of the vector, after After the operation, the resulting low-dimensional vector This is the final vector representation of the problem.

[0040] The above pooling characterization process flowchart is as follows: Figure 4 As shown.

[0041] Vector database index optimization process The vector database index optimization process uses low-dimensional pooled representations as the core input and combines them with structured metadata of test questions to build an efficient retrieval system. First, the data storage structure is defined, and then the multimodal pooled representation vectors of the test question dataset are processed. A Milvus set is constructed by combining the test question JSON data. The fields defined include test question ID, question stem text, image path, knowledge points, JSON metadata, and multimodal representation vectors. Let the fields of the Milvus set be... For each question, a one-to-one storage unit is formed between the semantic vector and the metadata, satisfying the following mapping relationship:

[0042] To construct an efficient graph-based approximate nearest neighbor search algorithm, the HNSW index, we first achieve fast searching through a layered graph structure, specifying the number of layers each node is inserted into. It follows a geometric distribution, and the probability formula is:

[0043] in It is the interlayer attenuation factor (usually taken as...). ), The maximum number of neighbors per floor, in floor In this model, nodes are only connected to nodes at the same level and lower levels. Higher-level nodes have lower density but larger connection spans, while lower-level nodes contain all data and are densely connected. Inter-layer connection rules are then established. Subsequently, graph connectivity is optimized by dynamically adjusting neighbor relationships for newly inserted nodes. Select in layer l When there are 1 neighbor, the following condition is met:

[0044] The search process employs a greedy algorithm to implement multi-level navigation for new nodes. From the highest level Begin by searching for the nearest neighbor and inserting it layer by layer downwards:

[0045] During construction, the maximum number of neighbors per layer is set to M=64, the distance metric is cosine similarity (COSINE), and the candidate pool size during construction is controlled to be efConstruction=80.

[0046] The flowchart of the above vector database index optimization process is as follows: Figure 5 As shown.

[0047] Multimodal similarity retrieval process Obtain a new set of test questions, including multimodal data such as text, images, and knowledge points. Vectorize the test questions using the multimodal pooling representation method described above, converting them to the same format as the vectors stored in the database, and integrate them into a unified query vector. Utilize the HNSW index in Milvus, with the search parameter efSearch=200, to calculate the cosine similarity between the query vector and the vectors in the database.

[0048] By using efficient index queries, the top 100 most similar results were retrieved.

[0049] The flowchart of the above multimodal similarity retrieval process is as follows: Figure 6 As shown.

[0050] In another embodiment of the present invention, a test item repetition detection system based on a multimodal large language model is provided, which can be used to implement the above-mentioned test item repetition detection method based on a multimodal large language model. Specifically, the system includes: The data acquisition module is used to collect multimodal test question data as a question bank dataset, and to clean the question bank dataset to obtain a cleaned dataset. The cleaning and pooling module is used to generate multimodal high-dimensional semantic vector representations of test questions from the cleaned dataset by combining a multimodal large language model, and to obtain the final vector representation of the test questions by pooling the multimodal high-dimensional semantic vector representations. The optimization module is used to construct the final vector representation of the questions, and to build a Milvus set and an HNSW index by combining the JSON data exchange format of the test questions; The output module is used to visualize the response results by combining the multimodal representation of the data set with Milvus's HNSW index.

[0051] The module division in this embodiment of the invention is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in the various embodiments of the invention can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0052] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used for the operation of a test item repetition detection method based on a multimodal large language model.

[0053] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the test item repeatability detection method based on a multimodal large language model in the above embodiments.

[0054] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0055] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0056] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0057] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0058] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.< / knowledge> < / stem>

Claims

1. A test item repetition detection method based on a multimodal large language model, characterized in that, include: Collect multimodal test item data as a question bank dataset, and clean the question bank dataset to obtain the cleaned dataset; For the cleaned dataset, a multimodal high-dimensional semantic vector representation of the test questions is generated by combining a multimodal large language model, and the final vector representation of the test questions is obtained by pooling the multimodal high-dimensional semantic vector representation. The final vector representation of the questions is used to construct a Milvus set and an HNSW index by combining the JSON data exchange format of the test questions; The multimodal representation of the data set to be searched, combined with Milvus's HNSW index, allows for the visualization of the response results. The process of generating multimodal high-dimensional semantic vector representations of test questions using a multimodal large language model for the cleaned dataset includes: The cleaned test data was split into three independent modal channels: text, image, and knowledge point. The text modality contained the plain text content of the question stem and options; the image modality associated with diagrams or formula images in the test questions; and the knowledge point modality was mapped to the leaf node level in a tree-like index structure. A 7B-parameter multimodal large language model was used. The three modal data were input into the model through a preset instruction template, and the multimodal features were aligned and fused across modalities based on a multi-head attention mechanism to generate high-dimensional semantic vector representations. ; The final vector representation of the problem obtained after pooling the multimodal high-dimensional semantic vector representation includes: Representation of the generated high-dimensional semantic vector conduct The operation compresses the high-dimensional vector into a fixed-length low-dimensional vector to preserve the core semantic information, thus obtaining the final vector representation of the problem. ; The final vector representation of the questions, combined with the JSON data exchange format for the questions, constructs a Milvus set and an HNSW index, including: Based on the pooled multimodal representation vector set A Milvus database collection was constructed by combining the JSON metadata of the test questions, with fields defined as test question ID, question stem text, image path, knowledge point, JSON metadata, and multimodal vector. The vector field was optimized for retrieval using the HNSW index, with the index parameters set as distance metric COSINE, index construction parameter number of adjacent nodes M=64, and search width.

2. The test item repetition detection method based on a multimodal large language model according to claim 1, characterized in that, The collection of multimodal test item data, used as a question bank dataset, includes: The dataset for the question bank is compiled from subject textbooks and online question banks containing open-ended test questions. A tree-like index structure is constructed according to the subject knowledge points, and the raw data is stored in JSON format. .

3. The test item repetition detection method based on a multimodal large language model according to claim 1, characterized in that, The data cleaning process for the question bank dataset, resulting in a cleaned dataset, includes: For the question dataset, regular expressions were used to remove HTML tags, and a LaTeX parsing tool was used to standardize formula formatting and option numbering formatting, resulting in a cleaned question bank dataset. .

4. The test item repetition detection method based on a multimodal large language model according to claim 1, characterized in that, The multimodal representation of the search set, combined with Milvus's HNSW index, visualizes the response results, including: Fast deduplication detection of multimodal test questions is achieved through vector cosine similarity calculation, based on the vector representation of the query set and the vector set. The system uses the HNSW index of the Milvus database, sets parameters for querying, and generates an interactive interface through a visualization engine to display the content of the top 100 similar questions.

5. A test item repetition detection system based on a multimodal large language model for implementing the test item repetition detection method based on a multimodal large language model as described in claim 1, characterized in that, include: The data acquisition module is used to collect multimodal test question data as a question bank dataset, and to clean the question bank dataset to obtain a cleaned dataset. The cleaning and pooling module is used to generate multimodal high-dimensional semantic vector representations of test questions from the cleaned dataset by combining a multimodal large language model, and to obtain the final vector representation of the test questions by pooling the multimodal high-dimensional semantic vector representations. The optimization module is used to construct the final vector representation of the questions, and to build a Milvus set and an HNSW index by combining the JSON data exchange format of the test questions; The output module is used to visualize the response results by combining the multimodal representation of the data set with Milvus's HNSW index.

6. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the test item repetition detection method based on a multimodal large language model as described in any one of claims 1 to 4.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the test item repetition detection method based on a multimodal large language model as described in any one of claims 1 to 4.

Images