Job question type intelligent editing system based on large language model

Abstract

The invention discloses a homework question type intelligent editing system based on a large language model, and the system comprises a natural language input module which is used for receiving a natural language instruction inputted by a user; the preprocessing module is used for preprocessing natural language input; the semantic understanding module is used for inputting the semantic feature vectors into the structure perception grouping capsule neural network; the field generation module is used for generating structured question type content; the structural consistency checking module is used for performing integrity checking on the structured question type content; the visual interaction module is used for rendering the structured question type content into a visual editing interface; the instruction reinjection module is used for capturing an incremental natural language instruction of the user; and the standardized output module is used for outputting the final structured question type content according to a predefined format. According to the invention, a structure-aware packet capsule neural network and natural language analysis are adopted to realize question type structure generation and intelligent editing.

Description

Technical Field

[0001] This invention relates to the field of intelligent educational content generation and processing technology, and in particular to an intelligent editing system for homework questions based on a large language model. Background Technology

[0002] With the advancement of digital education, homework question generation and editing systems are becoming increasingly common in teaching platforms. However, existing systems typically rely on fixed templates or manual input to combine question stems, options, answers, and explanations, lacking the intelligent understanding of natural language editing instructions and struggling to adapt to diverse and dynamic question generation needs. While some systems incorporate language models to generate question content, the generation process often lacks field-level structural control, easily leading to issues such as missing options, field confusion, or mismatched answer explanations. Furthermore, existing technologies generally lack mechanisms to verify the structural consistency of generated content, failing to correct logical conflicts or complete missing fields during automatic generation, thus affecting question quality and subsequent usability.

[0003] At the same time, the existing system also has shortcomings in terms of user interaction. Content modification is mostly done through interface selection or form input, failing to achieve automatic linkage between natural language interaction and structured content, resulting in low efficiency and insufficient intelligence in the editing process.

[0004] Therefore, how to provide an intelligent editing system for homework questions based on a large language model is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent homework question editing system based on a large language model. This invention combines a structure-aware grouped capsule neural network with natural language processing technology. By parsing user natural language editing commands, it performs field-level structural modeling and dynamic generation of the question stem, options, answer, and parsing fields, constructing complete structured question content. A structural consistency verification mechanism is introduced to automatically complete missing fields and correct structural conflicts. This invention possesses advantages such as natural and user-friendly interaction, reliable question structure, and high editing efficiency, effectively improving the automation level of homework question generation and editing. It is suitable for constructing multi-round interactive teaching resources in intelligent education scenarios.

[0006] According to an embodiment of the present invention, an intelligent homework question editing system based on a large language model includes: The natural language input module is used to receive natural language commands input by the user; The preprocessing module is used to clean, segment, and embed word vectors of natural language input, and extract semantic feature vector representations. The semantic understanding module is used to input semantic feature vectors into the structure-aware grouped capsule neural network to generate a set of field structure representation vectors; The field generation module is used to generate structured question content based on a set of field structure representation vectors. The structural consistency verification module is used to check the completeness of structured question content, fill in missing fields, correct structural conflicts, and generate structural optimization suggestions. The visual interaction module is used to render structured question content into a visual editing interface, and provides a natural language interaction interface for users to modify the question content; The instruction feedback module is used to capture incremental natural language instructions from the user in the editing interface and reprocess them as input for a new round of editing tasks; The standardized output module is used to output the final structured question content in a predefined format, enabling the standardized generation, export, or deployment of assignment questions.

[0007] Optionally, modules can be integrated using the following methods: It receives user input instructions for editing homework questions, encodes the instructions using a pre-trained large language model, and generates semantic representation vectors for the instructions. Based on semantic representation vectors, multiple action semantic units are extracted, and a set of action capsules corresponding to the action semantic units is constructed; Input the action capsule set into the structure-aware grouped capsule neural network to construct multiple field capsule groups and generate a set of field structure representation vectors. Input the set of field structure representation vectors into the field generation module to generate structured question content; Perform structural consistency checks on the structured question content to determine whether the structured question content conforms to the predefined question structure specifications. If there are missing fields, they will be automatically filled in. If there are structural conflicts, correction suggestions will be output. The system renders structured question content into a visual editing interface, receives new natural language editing commands from users, performs field-level updates on the structured question content, and outputs a structured question data file.

[0008] Optionally, the step of receiving user input for editing homework questions, and encoding the editing instructions using a pre-trained large language model to generate a semantic representation vector of the instructions, specifically includes: The system receives user-inputted homework question type editing instructions through a user interface. The editing instructions are input in the form of natural language text, and the content is used to indicate the modification requirements or generation requirements of the target homework question type. The natural language text is stored as the raw input data in the system's pending instruction cache area. The instruction text in natural language form is segmented into words according to semantic or grammatical boundaries. Each word unit is then vectorized and encoded into a numerical word vector, forming a word vector sequence composed of multiple word vectors. The word vector sequence is input into the encoding layer of the pre-trained large language model. The encoding layer performs context association modeling on each word vector in the word vector sequence to generate a context vector sequence corresponding to each word vector. The context vector sequence is weighted and aggregated, specifically: according to the importance of each context vector in the semantic expression of the instruction, a corresponding semantic importance weight is assigned to each context vector in the context vector sequence. The semantic importance weight is a non-negative real number, and the sum of all semantic importance weights is 1. After the weight allocation is completed, the context vectors are summed according to their corresponding semantic importance weights to generate the instruction semantic representation vector.

[0009] Optionally, the step of extracting multiple action semantic units based on semantic representation vectors and constructing a set of action capsules corresponding to the action semantic units specifically includes: Obtain the instruction semantic representation vector and the context vector sequence, and calculate the similarity between the instruction semantic representation vector and each context vector based on the relationship between the instruction semantic representation vector and the context vector sequence. Specifically, the instruction semantic representation vector is multiplied by a preset first mapping matrix to generate a mapped instruction semantic vector. At the same time, each context vector is multiplied by a preset second mapping matrix to generate a mapped context vector. Then, the mapped instruction semantic vector and each mapped context vector are subjected to a vector inner product operation to obtain a score value representing the degree of correlation between the instruction semantic representation vector and each context vector. All scores are then arranged in the order of the corresponding context vectors to form a score sequence. The score sequence is normalized by performing an exponential operation on each score value in the score sequence and dividing the result by the sum of the exponential operation results of all score values ​​to obtain the attention weight value corresponding to each context vector. Arrange all the attention weight values ​​generated above in the order of their corresponding context vectors in the sequence to form an attention weight sequence; Constructing a query vector set based on instruction semantic representation vectors involves: performing matrix multiplication operations on the instruction semantic representation vectors and the corresponding query mapping matrices according to the preset number of queries to generate query vectors, with each query vector corresponding to an independent semantic focus direction; For each query vector, semantic relevance is calculated with each context vector in the context vector sequence. By linearly mapping the query vector to each context vector and performing inner product operations in sequence, a set of unnormalized relevance scores is obtained. The relevance scores are indexed and normalized to generate attention weight values ​​that correspond one-to-one with each context vector. The attention weight values ​​are arranged in the order of their corresponding context vectors to obtain the attention weight sequence corresponding to the current query vector. The above steps are performed on all query vectors in sequence to form a multi-head attention weight set containing multiple attention weight sequences. Based on the set of query vectors and the attention weight sequence corresponding to each query vector, the context vector sequence is weighted and summed to generate a set of action semantic units; Specifically, for each query vector, the attention weight values ​​corresponding to each context vector in the context vector sequence are sequentially extracted from the corresponding attention weight sequence. Each attention weight value is multiplied by its corresponding context vector, and all product results are added to obtain the action semantic unit vector corresponding to the query vector. All action semantic unit vectors are arranged in the order of their corresponding query vectors to generate an action semantic unit set. The action semantic unit set is converted into an action capsule set. Specifically, for each action semantic unit vector in the action semantic unit set, the action semantic unit vector is first multiplied by a preset capsule mapping matrix, and a bias vector with the same dimension is added to generate the corresponding pre-activation vector. Then, the squeezing function operation is performed on the pre-activated vector. Specifically, the square of the magnitude of the pre-activated vector is calculated, and the reciprocal of the addition with the unit value is obtained to get the scaling factor used to control the magnitude of the vector. Then, the pre-activated vector is normalized according to its own magnitude, and the normalized pre-activated vector is multiplied by the scaling factor to generate the action capsule vector. Finally, the vector magnitude of each generated action capsule vector is calculated, and the vector magnitude is used as the activation coefficient of the action capsule vector. All action capsule vectors and their corresponding activation coefficients are combined to form an action capsule set.

[0010] Optionally, the step of inputting the action capsule set into the structure-aware grouped capsule neural network, constructing multiple field capsule groups, and generating a set of field structure representation vectors specifically includes: Obtain a set of action capsules, which serves as the input to a structure-aware grouped capsule neural network; Construct a structure-aware grouped capsule neural network, and set independent field capsule groups for each structural field of the homework question type. Each field capsule group consists of multiple sub-capsules. The structural fields include, but are not limited to, question stem field, option field, answer field, and explanation field. Each field capsule group corresponds one-to-one with the corresponding field, forming a field-level subspace modeling structure. The subcapsule in each field capsule group is a vector with a fixed dimension, and its initial state can be set to a zero vector or a randomly initialized vector; In the structure-aware grouped capsule neural network, semantic fit is calculated between each action capsule and each field capsule group to obtain the dynamic routing coefficient between action and field. The routing coefficient represents the routing strength of the action capsule to the field. Based on the dynamic routing coefficients, each action capsule is weighted and passed along according to its semantic compatibility with the grouping of field capsules; For each field capsule group, perform a squeezing function operation on all subcapsule vectors whose state has been updated, convert the field capsule group into the corresponding field structure representation vector, and combine the field structure representation vectors of all field capsule groups in field order to form a set of field structure representation vectors.

[0011] Optionally, the step of inputting the set of field structure representation vectors into the field generation module to generate structured question content specifically includes: The system receives a set of field structure representation vectors output by a structure-aware grouped capsule neural network. The set of field structure representation vectors consists of multiple field structure representation vectors, each of which represents the semantic feature information of a specific field in the structure of the homework question. The fields include the question stem field, the option field, the answer field, and the explanation field. Each field is semantically independent and functionally clearly defined. Each field structure representation vector in the set of field structure representation vectors is input into the corresponding field generation module. Different fields correspond to different field generation modules, which are used to generate the question stem text, option content, reference answer or explanation. The field generation module receives a field structure representation vector and guides content generation by initializing a hidden state or embedding a vector. The field generation process adopts an autoregressive mechanism, that is, the starting marker is used as the first input, and at each time step, the output word of the current time step is predicted based on the generated historical word sequence and the field structure representation vector. The output word is then used as the input of the next time step. This process is iterated until a termination marker is generated or the preset maximum length limit is reached, thereby completing the text generation of the corresponding field content. During the field content generation process, for each field generation module, the word probability distribution of the output word at the current time step is calculated based on the context information of the current generation position and the corresponding field structure representation vector. Specifically, by utilizing the decoder structure within the field generation module, the generated historical word sequence and the field's structural representation vector are input together to obtain the hidden state representation corresponding to the current generation position; The hidden state representation is linearly transformed through the output mapping layer and then fed into a normalization function to calculate the word probability distribution on the complete vocabulary. After completing the word probability distribution calculation for the current time step, the word with the highest probability is selected from the vocabulary as the output word for the current time step according to the principle of maximizing probability. The output word is then used as the input for the next time step to continue executing the decoding operation in the field generation module. Repeat the process described above, performing calculations and selections at each time step based on the current context and field structure representation vector to gradually generate complete field text content; When a preset termination marker appears in the generated sequence, or when the number of generated words reaches the maximum length limit set by the system, the content generation process of the current field is stopped. The resulting word sequence is the structured question content corresponding to the field, including the question stem, option points, reference answer content or explanation, which is determined by the field type. The question stem, option list, reference answer, and explanation content output by the all-field generation module are combined to form a structured question type.

[0012] Optionally, the step of performing structural consistency verification on the structured question content, determining whether the structured question content conforms to the predefined question structure specifications, automatically filling in missing fields if any exist, and outputting correction suggestions if structural conflicts exist, specifically includes: Each field text in the structured question content is parsed one by one. First, the question content is formatted using a field identifier extractor to identify the delimiters and structural tags of various fields, including the question stem, options, reference answer and explanation. Based on the preset field identifier rule base, the text fragments of each field are matched and classified to determine the field type identifier to which it belongs. The parsed field type identifiers are used to construct a key-value mapping structure with the corresponding field generated content, forming a mapping set of field identifiers and field content; Based on the field type identifier information bound to each field in the structured question content during the generation stage, a field existence state vector is constructed according to the predefined field set order; Specifically, let the predefined structured question types include For each standard field, it iterates through the set of field identifiers parsed from the structured question content. For each standard field, if a corresponding identifier exists in the parsed set, then the field exists in the state vector at the 1st position. Each position is assigned a value of 1 to indicate that the field exists; otherwise, it is assigned a value of 0 to indicate that the field is missing, thus obtaining the field existence state vector. The field type identification information includes the field number, field category, question type structure level, and the field's position index in the question type structure. It is uniformly encoded in the form of an identifier vector, where the field number serves as the primary key and the other attributes form an extended attribute dictionary. A set of structured question templates is pre-built. Each structured question template is defined as a combination of fields that conform to teaching standards, and the existence and arrangement rules of the question stem, options, answer, and explanation fields are clearly specified. For the currently generated structured question content, construct the corresponding field existence state vector based on the field type identifier information bound during the field generation process, indicating whether each type of field appears in the current question type; A structural consistency matching function is used to calculate the structural consistency score, and the template with the highest score is selected as the best matching template for the current structured question type. Based on the set of fields contained in the best matching template, compare it with the actual set of fields identified by the field existence state vector in the current structured question content to identify all fields that should exist in the template but are missing in the current question content, forming a set of missing fields; For each field in the missing field set, extract the context information of the existing fields in the structured question content, construct a field completion hint vector, input the hint vector into the field generation module, and generate field completion text content that conforms to the structure specifications based on the context modeling capability of the large language model. Then, fill the generated completion text back into the corresponding position of the structured question content according to the field type to complete the automatic completion of missing fields. The existing field pairs in the structured question content are combined in pairs to construct a set of field pairs. The structural logic consistency is judged based on field type, semantic relationship and content dependency. For each field pair, the structural conflict score is calculated. If the structural conflict score exceeds the structural consistency threshold, it is determined that there is a structural conflict between the fields. At this time, the structural correction suggestion generation process is triggered, and the minimum structural change norm optimization strategy is adopted. The minimum structural change norm optimization strategy refers to constructing a structural representation vector and calculating the difference between the structure before and after the correction when correcting the structured question content. The norm value of the difference is used as the optimization target. Under the premise of satisfying the structural consistency constraint, the structural adjustment range is minimized, thereby reducing the disturbance to the overall structure caused by field modification, position change or insertion and deletion operations. Candidate correction schemes are generated by combining greedy search with rule base matching, and the final output includes correction suggestions such as field content replacement, insertion, deletion or order adjustment.

[0013] Optionally, the step of rendering the structured question content into a visual editing interface, receiving new natural language editing commands from the user, performing field-level updates on the structured question content, and outputting a structured question data file specifically includes: The structured question content is input into a question rendering engine built on a predefined front-end component library. The question rendering engine identifies the question stem text field, option content field, reference answer field, and explanation field according to the field types in the structured question content. The question stem text field is rendered as a question stem display component, the option content field is rendered as an option input component, the reference answer field is rendered as an answer selection component, and the explanation field is rendered as a text display component. All of these components are presented in a visual interface form. Through the above rendering operations, a visual editing interface is generated. In the visual editing interface, the interactive operation interface is initialized and configured. The interactive operation interface includes a command input box, a field highlight control, and an editing preview window, which are used to provide user input commands, command visual positioning, and real-time feedback functions. The system monitors user editing behavior, captures interactive events including clicks, hovers, text input, and voice input, and receives natural language editing instructions input by the user, including field modification instructions, field addition instructions, field deletion instructions, and field rearrangement instructions. The text cleaning operation is performed on the natural language editing instructions, including removing invalid symbols, standardizing punctuation, eliminating stop words, and regularizing grammatical structure to obtain standardized instruction text; The standardized instruction text is segmented into words using a segmentation dictionary consistent with the predefined structure fields, and then divided into word units with semantic boundaries. By combining a predefined set of operation types, semantic matching and intent recognition are performed on word units, and natural language editing instructions are parsed and converted into structured operation expressions, which include operation types, target field identifiers and field content change information. The target field identifiers and change information in the structured operation expression are formatted and standardized, and a new field update request is constructed based on the current structured question type content; Based on the field update request, determine the field type and its position index that need to be reconstructed, call the field generation module to perform the reconstruction operation on the target field, and generate the updated field content; The updated field content is embedded into the original structured question content, while maintaining the original logical order and dependency relationship between fields; Perform a completeness check and structural consistency analysis on the embedded structured question content to confirm that there are no conflicts in field completion, format alignment and structure, and obtain the final updated structured question content; Replace the current question type content with the updated structured question type content and refresh the visual editing interface to provide instant feedback on the editing results; The updated structured question content is formatted and filled into a standard output template to generate a structured question data file, thus completing the intelligent editing process for the assignment questions.

[0014] The beneficial effects of this invention are: This invention introduces a processing framework combining a structure-aware grouped capsule neural network with a large language model to achieve field-level structural modeling and semantic representation of homework question content, improving the consistency between field content and structure during question generation. The system can automatically parse user-input natural language editing commands, generate structured operation expressions, and drive dynamic updates to question content, thereby automatically generating, modifying, and completing the question stem, options, answers, and explanation fields, reducing manual intervention in field decomposition and organization. Through a structural consistency verification mechanism, the system can detect missing fields and correct logical conflicts in the generated content, improving the reliability and accuracy of the content structure while ensuring generation efficiency.

[0015] Furthermore, this invention, through the integration of a visual editing interface, enables structured question types to be presented to users intuitively and supports incremental editing interaction based on natural language. This not only optimizes the user experience but also enhances the system's adaptability to actual teaching tasks. Overall, this invention achieves an automated closed-loop processing flow from natural language editing command input to structured question type output, which helps improve the intelligence level of homework question generation, editing, and management processes, reduces the time cost for educators in resource compilation, and improves the consistency and quality stability of homework content generation. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0017] Figure 1 This is a flowchart of a method for an intelligent homework question editing system based on a large language model, as proposed in this invention. Figure 2 This is a schematic diagram of the structure-aware grouped capsule neural network in an intelligent homework question editing system based on a large language model proposed in this invention. Figure 3 This is a schematic diagram of the structure of an intelligent homework question editing system based on a large language model proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1-3 A homework question type intelligent editing system based on a large language model includes: The natural language input module is used to receive natural language commands input by the user; The preprocessing module is used to clean, segment, and embed word vectors of natural language input, and extract semantic feature vector representations. The semantic understanding module is used to input semantic feature vectors into the structure-aware grouped capsule neural network to generate a set of field structure representation vectors; The field generation module is used to generate structured question content based on a set of field structure representation vectors. The structural consistency verification module is used to check the completeness of structured question content, fill in missing fields, correct structural conflicts, and generate structural optimization suggestions. The visual interaction module is used to render structured question content into a visual editing interface, and provides a natural language interaction interface for users to modify the question content; The instruction feedback module is used to capture incremental natural language instructions from the user in the editing interface and reprocess them as input for a new round of editing tasks; The standardized output module is used to output the final structured question content in a predefined format, enabling the standardized generation, export, or deployment of assignment questions.

[0020] In this embodiment, the modules are interconnected using the following method: It receives user input instructions for editing homework questions, encodes the instructions using a pre-trained large language model, and generates semantic representation vectors for the instructions. Based on semantic representation vectors, multiple action semantic units are extracted, and a set of action capsules corresponding to the action semantic units is constructed; Input the action capsule set into the structure-aware grouped capsule neural network to construct multiple field capsule groups and generate a set of field structure representation vectors. Input the set of field structure representation vectors into the field generation module to generate structured question content; Perform structural consistency checks on the structured question content to determine whether the structured question content conforms to the predefined question structure specifications. If there are missing fields, they will be automatically filled in. If there are structural conflicts, correction suggestions will be output. The system renders structured question content into a visual editing interface, receives new natural language editing commands from users, performs field-level updates on the structured question content, and outputs a structured question data file.

[0021] In this embodiment, the step of receiving the user's input instruction for editing the homework question type, and encoding the editing instruction using a pre-trained large language model to generate a semantic representation vector of the instruction specifically includes: The system receives user-inputted homework question type editing instructions through a user interface. The editing instructions are input in the form of natural language text, and the content is used to indicate the modification requirements or generation requirements of the target homework question type. The natural language text is stored as the raw input data in the system's pending instruction cache area. The instruction text in natural language form is segmented into words according to semantic or grammatical boundaries. Each word unit is then vectorized and encoded into a numerical word vector, forming a word vector sequence composed of multiple word vectors. The number of word vectors contained in the word vector sequence corresponds to the number of word units divided in the instruction text, and each word vector is used to represent the semantic features of the corresponding word unit. The word vector sequence is input into the encoding layer of the pre-trained large language model. The encoding layer performs context association modeling on each word vector in the word vector sequence to generate a context vector sequence corresponding to each word vector. Each context vector represents the semantic representation of the corresponding word in the entire instruction text. The semantic representation takes into account the semantic relationship between the word and the words before and after it, so as to fully represent the contextual semantic information of the instruction text. The context vector sequence is weighted and aggregated, specifically: according to the importance of each context vector in the semantic expression of the instruction, a corresponding semantic importance weight is assigned to each context vector in the context vector sequence. The semantic importance weight is a non-negative real number, and the sum of all semantic importance weights is 1. After the weight allocation is completed, each context vector is weighted and summed according to its corresponding semantic importance weight to generate an instruction semantic representation vector that comprehensively represents the overall semantic features of the instruction text.

[0022] In this embodiment, the step of extracting multiple action semantic units based on semantic representation vectors and constructing an action capsule set corresponding to each action semantic unit specifically includes: Obtain the instruction semantic representation vector and the context vector sequence, and calculate the similarity between the instruction semantic representation vector and each context vector based on the relationship between the instruction semantic representation vector and the context vector sequence. Specifically, the instruction semantic representation vector is multiplied by a preset first mapping matrix to generate a mapped instruction semantic vector. At the same time, each context vector is multiplied by a preset second mapping matrix to generate a mapped context vector. Then, the mapped instruction semantic vector and each mapped context vector are subjected to a vector inner product operation to obtain a score value that represents the degree of correlation between the instruction semantic representation vector and each context vector. All scores are arranged in the order of the corresponding context vectors to form a score sequence, which is used to represent the strength of semantic correlation between the instruction semantic representation vector and each context vector in the context vector sequence. The score sequence is normalized by performing an exponential operation on each score value in the score sequence and dividing the result by the sum of the exponential operation results of all score values ​​to obtain the attention weight value corresponding to each context vector, so that the value of each attention weight value is a non-negative real number and the sum of all attention weight values ​​is equal to 1. All the attention weight values ​​generated above are arranged in order of their corresponding context vectors in the sequence to form an attention weight sequence. The attention weight sequence is used to characterize the semantic contribution of each context vector in generating action semantic units. Constructing a set of query vectors for extracting different semantic features based on instruction semantic representation vectors: Specifically, according to the preset number of queries, perform matrix multiplication operations on the instruction semantic representation vectors and the corresponding query mapping matrices to generate query vectors, with each query vector corresponding to an independent semantic focus direction; Furthermore, for each query vector, semantic relevance is calculated with each context vector in the context vector sequence. By sequentially linearly mapping the query vector to each context vector and performing inner product operations, a set of unnormalized relevance scores is obtained. Subsequently, the relevance scores are indexed and normalized to generate attention weight values ​​that correspond one-to-one with each context vector. The attention weight values ​​represent the importance of each context vector relative to the current query vector. The attention weight values ​​are arranged in the order of their corresponding context vectors to obtain the attention weight sequence corresponding to the current query vector. The above steps are performed on all query vectors in sequence to form a multi-head attention weight set containing multiple attention weight sequences. Based on the set of query vectors and the attention weight sequence corresponding to each query vector, the context vector sequence is weighted and summed to generate a set of action semantic units; Specifically, for each query vector, the attention weight values ​​corresponding to each context vector in the context vector sequence are sequentially extracted from the corresponding attention weight sequence. Each attention weight value is multiplied by its corresponding context vector, and all product results are added to obtain the action semantic unit vector corresponding to the query vector. All action semantic unit vectors are arranged in the order of their corresponding query vectors to generate an action semantic unit set. Each action semantic unit vector represents the action feature information extracted for different semantic attention directions. The action semantic unit set is converted into an action capsule set. Specifically, for each action semantic unit vector in the action semantic unit set, the action semantic unit vector is first multiplied by a preset capsule mapping matrix, and a bias vector with the same dimension is added to generate the corresponding pre-activation vector. The preset capsule mapping matrix refers to the linear transformation parameter matrix pre-set for each type of capsule unit in the structure-aware grouped capsule neural network, which is used to perform linear transformation on the action semantic unit vector to generate a pre-activated vector consistent with the capsule dimension. Then, the squeezing function operation is performed on the pre-activated vector. Specifically, the square of the magnitude of the pre-activated vector is calculated, and the reciprocal of the addition with the unit value is obtained to get the scaling factor used to control the magnitude of the vector. Then, the pre-activated vector is normalized according to its own magnitude, and the normalized pre-activated vector is multiplied by the scaling factor to generate the action capsule vector. Finally, the vector magnitude of each generated action capsule vector is calculated, and the vector magnitude is used as the activation coefficient of the action capsule vector. All action capsule vectors and their corresponding activation coefficients are combined to form an action capsule set.

[0023] In this embodiment, the step of inputting the action capsule set into the structure-aware grouped capsule neural network, constructing multiple field capsule groups, and generating a set of field structure representation vectors specifically includes: Obtain a set of action capsules, which serves as the input to a structure-aware grouped capsule neural network; A structure-aware grouped capsule neural network is constructed, and independent field capsule groups are set for each structural field of the homework question type. Each field capsule group consists of multiple sub-capsules, which are used to represent the structural semantic representation of the field. The structural fields include, but are not limited to, question stem field, option field, answer field, and explanation field. Each field capsule group corresponds one-to-one with the corresponding field, forming a field-level subspace modeling structure. The subcapsule in each field capsule group is a vector with fixed dimensions. The initial state can be set to a zero vector or a randomly initialized vector to support independent semantic modeling and dynamic response of different fields. In a structure-aware grouped capsule neural network, semantic fit is calculated between each action capsule and each field capsule group to obtain dynamic routing coefficients between actions and fields. These routing coefficients represent the routing strength of an action capsule to a field. ; in, Indicates the first The action capsule and the first Dynamic routing coefficients between capsule groups of each field Represents the first action capsule in the set. One action capsule, This represents the routing mapping matrix corresponding to the field. The average subcapsule representation of a field, indicating a grouping of field capsules, is defined as the vector average of all subcapsule representations within the field. This indicates the total number of pre-defined fields in the assignment question type structure. It is the transpose symbol. This indicates the index number of the action capsule. Indicates the index number of the structure field. This is a placeholder index variable in the normalized summation, representing the index of any field in the capsule grouping of all fields; Based on the dynamic routing coefficients, each action capsule is weighted and passed on according to its semantic compatibility with the grouping of field capsules: ; in, Grouping the fields in capsules The first field The vector of sub-capsules whose states have been updated. Grouping the fields in capsules The first field One original subcapsule vector, This represents the total number of action capsules in the action capsule set. This is the index of the action capsule within the action capsule collection. For dynamic routing coefficients, In order to target the The first field Each subcapsule has a pre-defined projection matrix used to convert the action capsule vector into an augmented vector with the same dimensions as the subcapsule. This projection matrix is ​​a trainable parameter matrix. For the first Action capsule vectors for each action capsule; For each field capsule group, the squeezing function is performed on all subcapsule vectors whose state has been updated, and the field capsule group is converted into the corresponding field structure representation vector. Specifically, the subcapsule vectors in the field capsule group are aggregated to obtain an aggregate vector. The squeezing function is applied to the aggregate vector to compress the vector magnitude to a range no greater than 1, while maintaining the original direction information, forming a compressed vector with activation strength and directionality. The compressed vector is the field structure representation vector corresponding to the field. The field structure representation vectors of all field capsules are combined in the order of the fields to form a set of field structure representation vectors. Each structure representation vector corresponds to the question stem field, option field, answer field, or analysis field.

[0024] In this embodiment, the step of inputting the set of field structure representation vectors into the field generation module to generate structured question content specifically includes: The system receives a set of field structure representation vectors output by a structure-aware grouped capsule neural network. The set of field structure representation vectors consists of multiple field structure representation vectors, each of which represents the semantic feature information of a specific field in the structure of the homework question. The fields include the question stem field, the option field, the answer field, and the explanation field. Each field is semantically independent and functionally clearly defined. Each field structure representation vector in the set of field structure representation vectors is input into the corresponding field generation module. Different fields correspond to different field generation modules, which are used to generate the question stem text, option content, reference answer or explanation. The field generation module receives a field structure representation vector and guides content generation by initializing a hidden state or embedding a vector. The field generation process adopts an autoregressive mechanism, that is, the starting marker is used as the first input, and at each time step, the output word of the current time step is predicted based on the generated historical word sequence and the field structure representation vector. The output word is then used as the input of the next time step. This process is iterated until a termination marker is generated or the preset maximum length limit is reached, thereby completing the text generation of the corresponding field content. During the field content generation process, for each field generation module, the word probability distribution of the output word at the current time step is calculated based on the context information of the current generation position and the corresponding field structure representation vector. Specifically, by utilizing the decoder structure within the field generation module, the generated historical word sequence and the field's structural representation vector are input together to obtain the hidden state representation corresponding to the current generation position; The hidden state representation is linearly transformed through the output mapping layer and then fed into a normalization function to calculate the word probability distribution on the complete vocabulary. The word probability distribution is used to represent the probability of each candidate word being the output word at the current time step. After completing the word probability distribution calculation for the current time step, the word with the highest probability is selected from the vocabulary as the output word for the current time step according to the principle of maximizing probability. The output word is then used as the input for the next time step to continue executing the decoding operation in the field generation module. Repeat the process described above, performing calculations and selections at each time step based on the current context and field structure representation vector to gradually generate complete field text content; When a preset termination marker appears in the generated sequence, or when the number of generated words reaches the maximum length limit set by the system, the content generation process of the current field is stopped. The resulting word sequence is the structured question content corresponding to the field, including the question stem, option points, reference answer content or explanation, which is determined by the field type. The question stem, option list, reference answer, and explanation content output by the all-field generation module are combined to form a structured question type.

[0025] In this embodiment, the step of performing structural consistency verification on the structured question content, determining whether the structured question content conforms to the predefined question structure specifications, automatically filling in missing fields if any exist, and outputting correction suggestions if structural conflicts exist, specifically includes: Each field text in the structured question content is parsed one by one. First, the question content is formatted using a field identifier extractor to identify the delimiters and structural tags of various fields, including the question stem, options, reference answer and explanation. Based on the preset field identifier rule base, the text fragments of each field are matched and classified to determine the field type identifier to which it belongs. The preset field identification rule base refers to a set of rules pre-defined for the question stem field, option field, answer field, and parsing field to identify the category to which the field belongs. The rule set includes field name keywords, field structure marker symbols, option prefix format, answer marker format, and parsing trigger phrases. The parsed field type identifiers are used to construct a key-value mapping structure with the corresponding field generated content, forming a mapping set of field identifiers and field content; Based on the field type identifier information bound to each field in the structured question content during the generation stage, a field existence state vector is constructed according to the predefined field set order; Specifically, let the predefined structured question types include For each standard field, it iterates through the set of field identifiers parsed from the structured question content. For each standard field, if a corresponding identifier exists in the parsed set, then the field exists in the state vector at the 1st position. Each position is assigned a value of 1 to indicate that the field exists; otherwise, it is assigned a value of 0 to indicate that the field is missing, thus obtaining the field existence state vector. The field type identification information includes field number, field category, question type structure level, and the field's position index in the question type structure. It is uniformly encoded in the form of an identifier vector, where the field number serves as the primary key, and other attributes form an extended attribute dictionary to support the structure identification and position parsing of the field. A set of structured question templates is pre-built. Each structured question template is defined as a combination of fields that conform to teaching standards, and the existence and arrangement rules of the question stem, options, answer, and explanation fields are clearly specified. For the currently generated structured question content, construct the corresponding field existence state vector based on the field type identifier information bound during the field generation process, indicating whether each type of field appears in the current question type; The structural consistency matching function is used to compare the field existence state vector with all preset templates in turn, calculate the structural consistency score, and select the template with the highest score as the best matching template for the current structured question type. The structural consistency matching function is calculated based on field type consistency and field arrangement deviation. Field type consistency is the ratio obtained by matching the field existence state vector and the template field existence vector dimension by dimension. Field arrangement deviation is the sum of the absolute values ​​of the difference between the actual field number and the template number. The product of field type consistency and the first preset coefficient is used as the positive component of the structural consistency score, and the product of field arrangement deviation and the second preset coefficient is used as the negative component of the structural consistency score. The two are then added together to obtain the structural consistency score. Based on the set of fields contained in the best matching template, compare it with the actual set of fields identified by the field existence state vector in the current structured question content to identify all fields that should exist in the template but are missing in the current question content, forming a set of missing fields; For each field in the missing field set, extract the context information of the existing fields in the structured question content, construct a field completion hint vector, input the hint vector into the field generation module, and generate field completion text content that conforms to the structure specifications based on the context modeling capability of the large language model. Then, fill the generated completion text back into the corresponding position of the structured question content according to the field type to complete the automatic completion of missing fields. For each existing field pair in the structured question content, pairwise combinations are performed to construct a set of field pairs. Structural logical consistency is then judged based on field type, semantic relationship, and content dependency. For each field pair, a structural conflict score is calculated. ; in, , This indicates the first part of the structured question type content. The and the first One field, This indicates a score for structural conflicts in the field pairs. This is the semantic conflict weighting coefficient. For logical conflict weighting coefficients, For order conflict weighting coefficients, For fields With fields The semantic similarity score ranges from [0,1]. A higher value indicates a closer semantic similarity, which may constitute a conflict when the two options and answers are mutually compatible. For fields With fields The score for logical dependency conflict between the answers ranges from [0,1], indicating whether there is a logical inconsistency or contradiction. For example, the score increases when the answer content does not match the analytical reasoning. For fields With fields The relative order deviation score reflects the degree of deviation between the actual order of the fields and the standard order in the predefined template, with a value range of [0,1]. The semantic similarity score refers to calculating the semantic closeness between any two fields based on the semantic vector representation of the field content, using the cosine similarity method for calculation; The logical dependency conflict score refers to detecting whether there is a valid directed dependency path between two fields based on a predefined logical dependency graph between fields. If no path exists, a conflict score is set according to the conflict weight coefficient to reflect the degree of violation of the logical dependency relationship. The predefined logical dependency graph between fields refers to a directed graph structure pre-constructed for the logical reference relationships between fields in the question structure. Each node in the graph corresponds to a field, and each directed edge indicates a definite logical dependency relationship between the fields. For example, a directed edge from field A to field B indicates that the content of field B depends on field A to be valid. The logical dependencies include the content reference relationship between the question stem and the options, the subordinate relationship between the options and the answer, and the basis relationship between the answer and the analysis content. The relative arrangement order deviation score refers to the calculation of the degree of deviation between the position of the generated field in the structured question type and the preset order position in the standard structure. The larger the deviation, the higher the risk of structural conflict. The relative arrangement order deviation score is obtained by comparing the actual position of each field in the structured question content with the preset position of the field in the standard structure, calculating the absolute value of the difference between the two, summing the absolute values ​​of the difference between the numbers of all fields, and multiplying the sum by the preset order deviation weight coefficient to obtain the relative arrangement order deviation score used to characterize the degree of deviation of the field arrangement. If the structural conflict score exceeds the structural consistency threshold, it is determined that there is a structural conflict between the fields. At this time, the structural correction suggestion generation process is triggered, and the minimum structural change norm optimization strategy is adopted. The minimum structural change norm optimization strategy refers to constructing a structural representation vector and calculating the difference between the structure before and after the correction when correcting the structured question content. The norm value of the difference is used as the optimization target. Under the premise of satisfying the structural consistency constraint, the structural adjustment range is minimized, thereby reducing the disturbance to the overall structure caused by field modification, position change or insertion and deletion operations, and achieving the goal of restoring structural consistency with minimal modification. Candidate correction schemes are generated by combining greedy search with rule base matching, and the final output includes correction suggestions such as field content replacement, insertion, deletion or order adjustment.

[0026] In this embodiment, the process of rendering the structured question content into a visual editing interface, receiving new natural language editing commands from the user, performing field-level updates on the structured question content, and outputting a structured question data file specifically includes: The structured question content is input into a question rendering engine built on a predefined front-end component library. The question rendering engine identifies the question stem text field, option content field, reference answer field, and explanation field according to the field types in the structured question content. The question stem text field is rendered as a question stem display component, the option content field is rendered as an option input component, the reference answer field is rendered as an answer selection component, and the explanation field is rendered as a text display component. All of these components are presented in a visual interface form. Through the above rendering operations, a visual editing interface is generated. In the visual editing interface, the interactive operation interface is initialized and configured. The interactive operation interface includes a command input box, a field highlight control, and an editing preview window, which are used to provide user input commands, command visual positioning, and real-time feedback functions. The system monitors user editing behavior, captures interactive events including clicks, hovers, text input, and voice input, and receives natural language editing instructions input by the user, including field modification instructions, field addition instructions, field deletion instructions, and field rearrangement instructions. The text cleaning operation is performed on the natural language editing instructions, including removing invalid symbols, standardizing punctuation, eliminating stop words, and regularizing grammatical structure to obtain standardized instruction text; The standardized instruction text is segmented into words using a word segmentation dictionary consistent with the predefined structure fields. These words are divided into word units with semantic boundaries. Combined with a predefined set of operation types, the word units are semantically matched and their intent is identified. The natural language editing instructions are then parsed and converted into structured operation expressions, which include operation types, target field identifiers, and field content change information. The target field identifiers and change information in the structured operation expression are formatted and standardized, and a new field update request is constructed based on the current structured question type content; Based on the field update request, determine the field type and its position index that need to be reconstructed, call the field generation module to perform the reconstruction operation on the target field, and generate the updated field content; The updated field content is embedded into the original structured question content, while maintaining the original logical order and dependency relationship between fields; Perform a completeness check and structural consistency analysis on the embedded structured question content to confirm that there are no conflicts in field completion, format alignment and structure, and obtain the final updated structured question content; Replace the current question type content with the updated structured question type content and refresh the visual editing interface to provide instant feedback on the editing results; The updated structured question content is formatted and filled into a standard output template to generate a structured question data file for the homework system, thus completing the intelligent editing process of the homework questions.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to the scenario of constructing homework content on an online education platform. This platform needs to generate and manage large-scale objective question-based structured content for multiple grade levels for online practice, unit assessments, and final exams. In the traditional model, question creation mainly relies on manual input and template editing tools, resulting in low editing efficiency, poor content structure consistency, and a high rate of missing fields. Especially in scenarios involving multi-person collaboration and interdisciplinary question generation, discrepancies often arise between options and answers, or missing parsing fields, affecting subsequent automatic grading and practice record statistics.

[0028] Using the method of this invention, the system first generates an initial version of structured question content based on natural language instructions provided by the teacher, such as "generate a multiple-choice question about fraction addition." The system automatically generates the question stem, four options, and the correct answer field. The teacher then uses natural language instructions to further modify the question, such as "change option C to 13 / 15," "add an explanation to this question," or "change this question to a true / false question." The system can analyze the instruction intent in real time and automatically update the structured field content without requiring manual switching of input controls. The entire process is supported by a visual editing interface provided by the system, allowing users to intuitively see the changes in the question structure and content.

[0029] During system operation, to verify the technical effectiveness of this invention in ensuring structural consistency, improving field completion efficiency, and enhancing the natural interactive editing experience, 1000 homework questions, including true / false, single-choice, and multiple-choice questions, were selected as experimental samples. Some of these questions contained missing fields or structural errors. With the support of the method described in this invention, the system automatically identified and completed 94% of the questions with missing fields, and automatically generated correction suggestions for questions with logical conflicts in field content. Users can complete the construction and publication of a question type with an average of 1.7 commands through natural language interaction, reducing editing time by 68% compared to traditional template editing methods.

[0030] Table 1. Statistical data of experimental results

[0031] As can be seen from the table above, this invention demonstrates significant technical advantages in several key application indicators for generating and editing structured homework questions, particularly in terms of question structure consistency, automatic field completion rate, and editing efficiency, showing a marked improvement over traditional manual form entry or semi-template editing methods. Firstly, regarding structural consistency assurance, this invention achieves a pass rate exceeding 96% for all three question types: true / false, single-choice, and multiple-selection, with single-choice questions reaching 99.2%. This indicates that through the collaborative processing of field structure modeling and consistency verification modules, this invention can effectively detect and correct errors in question structure, reducing content quality issues caused by missing fields, improper arrangement, or semantic conflicts.

[0032] Secondly, regarding the automatic completion capability for missing fields, this invention achieved a completion rate exceeding 92% for all three question types, with the highest completion rate of 95.7% for multiple-choice questions. This indicates that the system can automatically detect and complete missing field content when users generate or edit questions for the first time, reducing the amount of manual field-level content addition by users and improving task continuity and editing smoothness. In terms of conflict correction suggestions, this invention achieved an identification and output rate of over 84% for logical field conflicts, reaching 89.3% for multiple-choice questions. This demonstrates that the system can dynamically analyze the semantic and structural relationships between fields, generating reasonable suggestions for potential content contradictions or format inconsistencies, ensuring that the generated questions are structured and logically consistent before being presented to students.

[0033] Furthermore, regarding editing interaction, the natural language intelligent editing interface introduced in this invention allows users to complete tasks such as question construction or modification with an average of only 1.7 instructions, significantly simplifying the tedious field location and manual input process of traditional operations. Combined with the average reduction in editing time shown in the table, it can be seen that, under the combined effect of this invention's system, the average time users spend editing questions is reduced by 68.2%, with a maximum reduction of 71.2%. This fully demonstrates that this invention not only reduces the human resource cost of content generation but also significantly improves the interactive efficiency and automation level of task construction.

[0034] In summary, this invention achieves high automation, high accuracy, and high user-friendliness throughout the entire process of intelligent editing of homework questions through structure-aware semantic modeling, dynamic field generation, and interactive consistency verification strategies. It provides reliable technical support for the construction of online education resources and distributed teaching applications, and has good promotional value.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A large language model-based intelligent editing system for homework types, characterized by, The application relates to a natural language input module for receiving user input natural language instructions, a preprocessing module for cleaning, word segmentation and word vector embedding of the natural language input, extracting semantic feature vector representation, a semantic understanding module for inputting the semantic feature vector into a structure perception grouping capsule neural network to generate a field structure representation vector set, a field generation module for generating structured question type content based on the field structure representation vector set, a structure consistency checking module for integrity checking of the structured question type content, completing missing fields, correcting structure conflicts and generating structure optimization suggestions, a visual interactive module for rendering the structured question type content into a visual editing interface and providing a natural language interactive interface for user modification of the question content, an instruction feedback module for capturing user incremental natural language instructions in the editing interface and reprocessing the same as a new round of editing task input, and a standardized output module for outputting the final structured question type content in a predefined format. The modules are connected through the following methods: Receiving user input homework question type editing instructions, using a pre-trained large language model to encode the editing instructions to generate semantic representation vectors of the instructions; Based on the semantic representation vectors, a plurality of action semantic units are extracted, and an action capsule set corresponding to the action semantic units is constructed; The action capsule set is input into a structure perception grouping capsule neural network to construct a plurality of field capsule groups and generate a field structure representation vector set; The field structure representation vector set is input into the field generation module to generate structured question type content; The structured question type content is checked for structure consistency, and it is determined whether the structured question type content conforms to the predefined question type structure specification. If there are missing fields, they are automatically completed. If there is a structure conflict, a correction suggestion is output; The structured question type content is rendered into a visual editing interface, and user-added natural language editing instructions are received. The structured question type content is updated at the field level, and a structured question type data file is output. The receiving of user input homework question type editing instructions and the encoding of the editing instructions by using a pre-trained large language model to generate semantic representation vectors of the instructions specifically include:

2. The large language model-based intelligent editing system for exercise questions according to claim 1, wherein, Receiving user input homework question type editing instructions through a user interactive interface, performing word segmentation processing on the instruction text in natural language form to form a word vector sequence; Inputting the word vector sequence into the encoding layer of the pre-trained large language model to model the context association of each word vector in the word vector sequence through the encoding layer to generate a context vector sequence corresponding to each word vector; Assigning a corresponding semantic importance weight to each context vector in the context vector sequence; After completing the weight assignment, the context vectors are weighted and summed according to the corresponding semantic importance weights to generate instruction semantic representation vectors. The extraction of a plurality of action semantic units based on the semantic representation vectors and the construction of an action capsule set corresponding to the action semantic units specifically include: ​ ​ 3. The large language model-based intelligent editing system for job questions according to claim 2, wherein, ​ ​ ​ ​ ​ 4. The large language model-based intelligent editing system for job questions according to claim 2, wherein, ​ The instruction semantic representation vector and the context vector sequence are obtained, similarity calculations are performed on the instruction semantic representation vector and each context vector based on an association relationship between the instruction semantic representation vector and the context vector sequence, and a score sequence is formed; The score sequence is normalized to obtain attention weight values, and all of the attention weight values are arranged in order according to the order of the corresponding context vectors in the sequence to form an attention weight sequence; A query vector set is constructed based on the instruction semantic representation vector, and corresponding multi-head attention weights are calculated to form a multi-head attention weight set; The context vector sequence is weighted and summed according to the multi-head attention weights to generate a set of action semantic units; For each action semantic unit vector in the set of action semantic units, the action semantic unit vector is multiplied by a preset capsule mapping matrix, and a bias vector with the same dimension is added to generate a corresponding pre-activation vector; The pre-activation vector is subjected to a squeeze function operation to generate an action capsule vector, the vector norm of each action capsule vector is calculated, the vector norm is used as the activation coefficient of the action capsule vector, and all action capsule vectors and corresponding activation coefficients are combined to form an action capsule set.

5. The large language model-based intelligent editing system for job questions according to claim 2, wherein, The action capsule set is input into a structure-aware grouping capsule neural network to construct multiple field capsule groups and generate a set of field structure representation vectors, specifically including: The action capsule set is obtained and input into the structure-aware grouping capsule neural network; The structure-aware grouping capsule neural network is constructed, and independent field capsule groups are set for each structure field of the exercise type, each field capsule group being composed of multiple sub-capsules; The sub-capsules in each field capsule group are vectors with fixed dimensions, and the initial state can be a zero vector or a randomly initialized vector; In the structure-aware grouping capsule neural network, a semantic adaptation degree is calculated between each action capsule and each field capsule group to obtain a dynamic routing coefficient between the action and the field; According to the dynamic routing coefficient, each action capsule is weighted and transmitted according to the semantic adaptation degree between the action capsule and each field capsule group, and the state of each sub-capsule in all field capsule groups is updated; The squeeze function is performed on each updated sub-capsule vector in each field capsule group to convert the field capsule group into a corresponding field structure representation vector; All field structure representation vectors are combined in field order to form a set of field structure representation vectors.

6. The large language model-based intelligent editing system for job questions according to claim 2, wherein, The set of field structure representation vectors is input into a field generation module to generate structured question content, specifically including: The set of field structure representation vectors is received, and each field structure representation vector in the set of field structure representation vectors is input into a corresponding field generation module, different fields corresponding to different field generation modules; The field generation module implements a language generation process of field content based on a trainable decoder structure, and uses an autoregressive mechanism to generate a text sequence step by step from an initial symbol; During the field content generation process, for each field generation module, a word probability distribution of a current time step output word is calculated according to context information at a current generation position and a corresponding field structure representation vector; According to the principle of probability maximization, the word with the highest probability is selected from the vocabulary as the output word of the current time step, and the output word is taken as the input of the next time step, and the decoding operation in the field generation module is continued to be executed; According to the principle of probability maximization, the word with the highest probability is selected from the output vocabulary as the current step output in each generation step, and the above process is repeated until the field generation termination condition is met, and the structured text content corresponding to the field is obtained; The stem text, option list, reference answer and analysis content output by all field generation modules are combined to form structured question type content.

7. The large language model-based intelligent editing system for job questions according to claim 2, wherein, The structure consistency of the structured question type content is verified, whether the structured question type content conforms to the pre-defined question type structure specification is judged, if there is a missing field, it is automatically completed, if there is a structure conflict, a modification suggestion is output, which specifically includes: Each field text in the structured question type content is parsed, the corresponding field type identifier and generated content are extracted, and a mapping set of field identifier and content is constructed; Based on the field type identifier information bound by each field in the structured question type content in the generation stage, a field existence state vector is constructed according to the pre-defined field set order; A set of structured question type templates is constructed in advance, each structured question type template is defined as a group of field combinations conforming to the teaching specification, the structure consistency score between the current structured question type content and each template is calculated through a matching function, and the best matching template is determined; Based on the field existence state vector and the best matching template, the missing field set in the current structured question type content is identified, and the field generation module is used again for the completion of the missing field, and the completed field text is output; Each field pair in the structured question type content is combined, a field pair set is constructed, and the structure logical consistency is judged based on the field type, semantic relationship and content dependence. For each field pair, the structure conflict score of the field pair is calculated. If the structure conflict score exceeds the structure consistency threshold, it is determined that the field pair has structure conflict, the structure modification suggestion generation process is triggered, and the minimum structure change norm optimization strategy is adopted to generate the modification suggestion.

8. The large language model-based intelligent editing system for job questions according to claim 2, wherein, The structured question type content is rendered into a visual editing interface, natural language editing instructions input by a user are received, field-level updates are performed on the structured question type content, and a structured question type data file is output, which specifically includes: The structured question type content is input into a question type rendering engine based on a pre-defined front-end component library, and is rendered into a stem display component, an option input component, a reference answer component and an analysis text component in sequence according to the field category, to generate a visual editing interface; In the visual editing interface, an interactive operation interface is initialized and configured, user editing behaviors are listened to, and natural language editing instructions input by a user are received; The natural language editing instructions are subjected to text cleaning and word segmentation processing, and are converted into structured operation expressions in combination with a pre-defined operation type set; The target field identifier and change information in the structured operation expression are subjected to format standardization processing, and a new field update request is constructed in combination with the current structured question type content. According to the field update request, the field type and position index to be reconstructed are determined, a field generation module is called to perform a reconstruction operation on the target field, and updated field content is generated; The updated field content is embedded into the original structured question type content, and the embedded structured question type content is subjected to integrity check and structure consistency analysis to obtain updated structured question type content; The updated structured question type content is replaced with the current question type content, and the visual editing interface is refreshed; The updated structured question type content is structured and laid out, and filled into a standard format output template to generate a structured question type data file.

Images