An intelligent homework correction method based on multi-modal analysis and hierarchical degradation scoring

Abstract

The application provides an intelligent homework correction method based on multi-modal analysis and hierarchical degradation scoring, relates to the technical fields of artificial intelligence, intelligent education and multi-modal document analysis, and the method is concurrent analysis on a homework file to be corrected and a standard answer file to obtain homework text content and standard answer text content; an initial correction item is obtained by using a question structuring and matching alignment strategy; based on the homework text content and the initial correction item, missing score completion is performed to obtain a completed correction item and batch parallel scheduling is performed to obtain a distributed correction task; a second large language model is called to score to obtain a single question correction result; based on multiple single question correction results, data is summarized to obtain an overall correction result. The application solves the problems of insufficient complex layout homework file analysis capability and low question and answer structured matching accuracy in the existing intelligent correction technology.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, smart education, and multimodal document parsing technology, and in particular to an intelligent homework grading method based on multimodal parsing and hierarchical degraded scoring. Background Technology

[0002] With the rapid popularization of educational informatization and artificial intelligence technologies, homework correction and student learning analysis have gradually become key components of smart education systems. In traditional teaching scenarios, teachers typically need to manually correct a large number of student assignments. Given the large number of students in a class, the high frequency of assignments, and the complexity of question types, manual correction consumes a significant amount of after-school time. Furthermore, manual grading is easily affected by factors such as the teacher's subjective state, differences in experience, and time pressure, leading to low correction efficiency, excessively long feedback cycles, and insufficient consistency in grading standards.

[0003] Existing automated grading systems mostly focus on objective questions or questions with fixed formats. For multiple-choice and true / false questions, the system can score by comparing options; for some fill-in-the-blank questions, it can score by string matching. However, assignments in real teaching scenarios include open-ended questions such as short-answer questions, essay questions, and calculation problems. The answers to these subjective questions are diverse in expression and semantically complex. Simply relying on rule matching or traditional machine learning methods makes it difficult to accurately determine the semantic consistency, logical completeness, and coverage of scoring points between the answers and the standard answers.

[0004] The sources and formats of actual assignment documents are highly complex. Assignments often exist in the form of electronic PDFs, scanned documents, images, or paper copies containing handwritten content, mixing formulas, tables, graphs, and continuous content spanning multiple pages. Traditional OCR technology uses an image recognition-text extraction pipeline, which is extremely unstable in recognizing mathematical formulas, complex layouts, and illegible handwriting. This easily leads to garbled formulas, lost question numbers, and broken content spanning multiple pages, directly compromising the reliability of subsequent question segmentation and answer matching.

[0005] Furthermore, existing systems rely on fixed templates for structured question processing, making them prone to segmentation errors when question number formats change or questions span multiple pages. In the answer matching stage, they depend solely on question number or sequence alignment, making mismatches highly likely when answer formats are not standardized. Score recognition often requires manual configuration, lacking automatic extraction and completion mechanisms, thus reducing the system's automation rate.

[0006] In recent years, large language models have demonstrated excellent performance in natural language understanding, providing a new approach for subjective question scoring. However, directly applying large models for grading suffers from a fatal flaw: unstable output. Models frequently exhibit anomalies such as non-JSON formatted responses, missing fields, and out-of-bounds scores. A single output anomaly can cause grading failure for a single question, interrupting the grading process. Furthermore, grading large batches of assignments relies on serial processing, with time increasing linearly with the number of questions. The lack of unified state management, batch control, and anomaly degradation mechanisms makes it difficult to meet the business requirements for rapid feedback in online assignments. Summary of the Invention

[0007] To address the aforementioned shortcomings in existing technologies, this invention provides an intelligent homework grading method based on multimodal parsing and hierarchical degradation scoring. This method solves the problems of insufficient parsing capabilities for complex layout homework files, low accuracy of structured matching of questions and answers, reliance on manual configuration for missing scores, unstable scoring output from large language models that can easily lead to process interruptions, and lack of unified state scheduling for batch concurrent grading tasks.

[0008] To achieve the aforementioned objectives, the technical solution adopted by this invention is: an intelligent homework grading method based on multimodal analysis and hierarchical degradation scoring, comprising: S1: Obtain the grading task request. Based on the homework file to be graded and the standard answer file contained in the grading task request, perform concurrent parsing using a multimodal parsing method containing the first major language model to obtain the homework text content and the standard answer text content. S2: Based on the homework text content and the standard answer text content, the initial grading items are obtained by using the question structuring and matching alignment strategy; S3: Based on the homework text content and the initial grading items, extract the question type information and score information and fill in the missing scores to obtain the completed grading items; S4: Based on the completed correction entries, construct a correction task list using a stateful workflow and perform batch parallel scheduling to obtain the distributed correction tasks. S5: Based on the distributed grading task, call the second language model to score, and perform hierarchical downgrade scoring when the second language model outputs an anomaly to obtain the grading result for a single question. S6: Based on the multiple individual question grading results, the data is summarized to obtain the overall grading result, and the intelligent homework grading is completed.

[0009] Further, S1 includes: Get the grading task request; Based on the homework file to be graded and the standard answer file included in the grading task request, pagination rendering is performed and the layout structure features and position association information of adjacent pages are extracted to obtain the corresponding cross-page association area data. Based on the cross-page association region data, a cross-page joint context input structure is constructed and the first large language model is used to perform continuous content recognition to obtain the page parsing results of each page; Based on the page parsing results, the page order is spliced ​​and merged to obtain the homework text content and the standard answer text content, respectively.

[0010] Further, S2 includes: Based on the content of the assignment text, question boundary identification and information extraction are performed to obtain candidate question units; Based on the candidate question units and standard answer text content, semantic and attribute features are used for matching and the matching success rate is calculated to obtain preliminary matching results; When the matching success rate of the preliminary matching results is lower than a preset threshold, the interval mapping is re-mapped based on the question order features using a position alignment strategy to obtain the initial grading entries.

[0011] Further, S3 includes: Based on the content of the assignment text and the initial grading items, multi-dimensional score data is extracted and a mapping relationship is established to obtain a question type score mapping table; Based on the question type score mapping table, identify the score missing status in the initial grading items to obtain the missing score items; Based on the missing score entries, the score data of the corresponding level is matched sequentially according to the priority rules of the preset score source for automatic completion processing, resulting in the completed correction entries.

[0012] Further, S4 includes: Based on the completed revise entries, the task is divided using a preset maximum parallel number parameter to obtain multiple revise task batches. Based on the multiple batches of batch modification tasks, state data containing the current batch identifier and execution exception records is constructed in the stateful workflow; Based on the state data, the current batch of tasks is issued and the corresponding state is updated synchronously using the parallel execution unit of the stateful workflow to obtain the distributed batch modification tasks.

[0013] Further, S5 includes: Based on the distributed grading task, the structured output mode of the second language model is used for scoring and format anomaly detection is performed to obtain the first scoring status data. When the first rating status data indicates an anomaly, text parsing and field integrity verification are performed based on the original text returned by the second large language model to obtain the second rating status data; When the second scoring status data indicates an anomaly, a fallback calculation is performed based on the preset scoring strategy, and the score validity constraint is applied to obtain the single-question grading result.

[0014] Furthermore, the single-question grading result includes the actual score, and the expression for the actual score is: ; in, This represents the final actual score after constraints. This indicates the first language model obtained from the second largest language model. The original score of the question. This represents the function that takes the maximum value. This represents the function that takes the minimum value. This indicates the number obtained according to the question type score mapping table. The full marks for each question.

[0015] The beneficial effects of this invention are as follows: This invention provides an intelligent homework grading method based on multimodal parsing and hierarchical degraded scoring. By utilizing a multimodal parsing method that includes a first major language model to perform concurrent parsing of homework and standard answers, it breaks through the technical bottleneck of traditional single OCR and effectively handles handwritten content and complex structures in images and scanned documents. By implementing question structuring and matching alignment strategies based on the homework text content and standard answer text content, mismatches and omissions are reduced, providing high-confidence comparative benchmark data for subsequent scoring. By extracting question type information and score information and completing missing scores, automatic adaptation of scoring rules under different test paper formats is achieved, avoiding manual entry of scores for each question. By constructing a grading task list using a stateful workflow based on the completed grading entries and performing batch parallel scheduling, the performance bottleneck of traditional serial processing is solved, and the status monitoring and traceable management of concurrent tasks are realized. By performing hierarchical degraded scoring when calling the second major language model for scoring and the model output is abnormal, grading failures caused by single model output format errors or network jitter are avoided, ensuring the continuous high availability of the system when facing complex subjective and objective mixed question types. Attached Figure Description

[0016] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is an exemplary flowchart illustrating an intelligent job grading method based on multimodal analysis and hierarchical degradation scoring, according to some embodiments of this specification. Detailed Implementation

[0017] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0018] Example Figure 1 This is an exemplary flowchart illustrating an intelligent job grading method based on multimodal analysis and hierarchical degradation scoring, according to some embodiments of this specification. Figure 1 As shown, the process includes the following steps. In some embodiments, the process may be executed by a processor.

[0019] S1: Obtain the grading task request. Based on the homework file to be graded and the standard answer file contained in the grading task request, perform concurrent parsing using a multimodal parsing method containing the first major language model to obtain the homework text content and the standard answer text content.

[0020] A grading task request is the initial set of instructions and data used to trigger the entire intelligent homework grading system to execute the automated grading process. For example, a grading task request may include the homework file to be graded uploaded by the user's front end, the standard answer file, and grading configuration parameters used to control the grading logic, such as default score settings, parsing model identifier, scoring model identifier, maximum parallelism limit, and synchronous or asynchronous grading mode.

[0021] In some embodiments, the processor can obtain a batch modification task request by receiving a network request message sent by the client and validating its file format, null values, and necessary parameters.

[0022] The assignment files to be graded are the basic documents containing students' answers and serve as the raw input for the system's subsequent multimodal analysis and grading. For example, assignment files to be graded may include electronic PDF files, multi-page scanned files, image files, photographs, or digitized copies of paper assignments containing illegible handwriting, complex formulas, and charts.

[0023] In some embodiments, the processor can obtain the job file to be graded by parsing the file upload field in the received grading task request or by extracting a file stream from a preset cloud object storage server.

[0024] A standard answer key is a basic document containing the correct answers and reference scoring criteria. It serves as a benchmark during the grading process to determine the correctness and completeness of the answers. For example, a standard answer key can include text in electronic document format, paragraph text organized by question number or question type, or a reference document containing explanations and scoring points.

[0025] In some embodiments, the processor can obtain a standard answer file containing the standard answer text content by parsing the corresponding parameter fields in the grading task request or by calling the interface of the question bank system.

[0026] The first major language model is a low-level artificial intelligence algorithm model specifically designed for deep analysis of multimodal documents and extraction of complex structural information. For example, the first major language model may include a large multimodal model with powerful image recognition, complex mathematical and chemical formula analysis, page layout reconstruction, and accurate handwritten text recognition capabilities to handle issues such as page breaks and missing pages.

[0027] In some embodiments, the processor can construct a network request to invoke a preset multimodal parsing service interface deployed in the cloud or locally, and pass in a cross-page federated context to obtain the first major language model.

[0028] The assignment text content is a complete electronic plain text sequence data obtained by multimodal concurrent parsing and structural reorganization of the original assignment file to be graded. For example, the assignment text content may include a mixture of document information such as main question titles, sub-question numbers, question stem descriptions, text converted from complex formulas after cleaning, and strings converted from users' handwritten answers.

[0029] In some embodiments, the processor can obtain the job text content by rendering the job file to be graded in pages, inputting it into a multimodal large language model, and splicing and merging the parsing results of each page according to the original page order.

[0030] The standard answer text content is a sequence of plain text used for comparison, obtained by processing the standard answer file through a text parsing engine and removing irrelevant formatting information. For example, the standard answer text content may include a structured sequence of standard question numbers, answer body paragraphs, scoring criteria for each scoring step, and explanatory fields for teachers' reference but not for direct comparison.

[0031] In some embodiments, the processor can obtain the standard answer text content by performing preprocessing operations such as text cleaning, question number recognition, paragraph segmentation, and title filtering on the standard answer file to be graded.

[0032] In some embodiments, the processor can acquire a grading task request; based on the homework file to be graded and the standard answer file contained in the grading task request, perform pagination rendering and extract the layout structure features and positional association information of adjacent pages to obtain corresponding cross-page association area data; based on the cross-page association area data, construct a cross-page joint context input structure and use the first large language model to perform continuous content recognition to obtain the page parsing results of each page; based on the page parsing results, perform page sequence splicing and merging processing to obtain the homework text content and the standard answer text content, respectively.

[0033] Page layout features are the underlying topological data representations that describe the physical arrangement, logical distribution, and spatial relationships of various elements on a document page. For example, page layout features may include geometric structure information such as unfinished question stems near the bottom of the current page, unclosed mathematical formulas, incomplete tables, and continuation content areas at the top of adjacent pages.

[0034] In some embodiments, the processor can obtain page structure features by applying computer vision layout analysis algorithms and object detection algorithms to the rendered current page image.

[0035] Location association information is coordinate constraints and index data that characterizes the spatial continuity and logical connection possibilities between potentially broken cross-page elements in a document. For example, location association information may include numerical continuity features based on question numbers, semantic continuity features at the breaks in text sentences, and relative coordinates and weight parameters of cross-page association areas jointly established by the spatial continuity at the bottom and top of the page.

[0036] In some embodiments, the processor can obtain positional association information by extracting and comparing the layout structure features and text truncation markers of adjacent page edge regions.

[0037] Cross-page related region data is a joint input image and feature data structure constructed to solve the problem of questions or answers breaking across pages, containing local and global contextual information. For example, cross-page related region data may include a global complete page image of the current page, a local image of the potentially truncated area at the end of the current page, a local image of the potentially continuous area at the beginning of the adjacent page, and corresponding area position indexes and association weight information.

[0038] In some embodiments, the processor can obtain cross-page associated region data by integrating the extracted page structure features and position association information, and cropping and combining the corresponding image regions.

[0039] Page parsing results are a set of structured text or feature sequences output from a single document page after multimodal model recognition and inference processing. For example, page parsing results may include the plain text content extracted from the current page, the parsed and restored formula source code sequence, the page layout structure hierarchy information, and the recognized and converted handwritten answer text paragraphs.

[0040] In some embodiments, the processor can obtain page parsing results by inputting cross-page association region data containing the joint context and a single-page image into a first large language model for processing.

[0041] In some embodiments, constructing the cross-page joint context input structure includes calculating the association weight of the cross-page associated regions, whereby the expression for the association weight is: ; in, Indicates the cross-page region association weight. This represents the semantic similarity calculation function. This indicates the text characteristics at the bottom of the current page. This indicates the text features at the beginning of adjacent pages. This represents the sum of the absolute physical distances from the bottom edge of the page element to the top edge of the page element. This represents the preset page spacing normalization constant. This is a Boolean flag indicating whether the question numbers are consecutive; it is 1 if the question numbers are consecutive, and 0 otherwise. This indicates the preset adjustment coefficient. The processor compares... Based on the preset cross-page trigger threshold, determine whether to perform cross-page joint parsing.

[0042] S2: Based on the assignment text content and the standard answer text content, the initial grading items are obtained by using the question structuring and matching alignment strategy.

[0043] The initial grading entries are a baseline intermediate data structure that establishes a preliminary correspondence between the assignment answers and the standard answers after structured segmentation. For example, the initial grading entries may include the question number, question type identifier, question stem text, original text of the user's answer, the standard answer fragments retrieved through matching, and information fields indicating the matching status and confidence level.

[0044] In some embodiments, the processor can segment the assignment text content by identifying question boundaries and align it with the standard answer text content based on semantic and question number rules to obtain initial grading entries. Specifically, when multiple questions are misaligned simultaneously, continuous interval alignment is performed on the unmatched areas according to the relative positional relationship of the matched questions before and after; the candidate standard answer intervals are reordered based on the question sequence position, question type consistency, and text semantic similarity, and the standard answer with the highest matching score is selected as the mapping result of the current question unit.

[0045] In some embodiments, the selection of the standard answer with the highest matching score as the mapping result of the current question unit is defined as follows: ; in, Indicates the first The candidate topic unit and the first The matching score of each standard answer interval. This represents the semantic similarity between the question and the answer, calculated using a text embedding model. This represents the normalized position index of the problem in the sequence of assignments. This represents the normalized position index of the answer within the sequence of standard answers. This indicates the total number of questions. This represents the feature weight parameters.

[0046] In some embodiments, the expression for the success rate of answer retrieval in the initial grading entries, obtained by utilizing question structuring and matching alignment strategies, is as follows: ; in, Indicates the success rate of answer retrieval. This represents the total number of candidate question units obtained through question boundary identification and deduplication merging. This indicates the number of candidate question units that successfully matched the text content of the standard answer.

[0047] In some embodiments, the processor can perform question boundary recognition and information extraction based on the task text content to obtain candidate question units; based on the candidate question units and the standard answer text content, it can perform matching using semantic and attribute features and calculate the matching success rate to obtain preliminary matching results; when the matching success rate of the preliminary matching results is lower than a preset threshold, it can perform re-interval mapping based on the question order features using a position alignment strategy to obtain initial grading entries.

[0048] Candidate question units are independent, basic data structures of question blocks to be processed, obtained by physically and logically segmenting long, unstructured assignment texts. For example, a candidate question unit may include the unique question number of a single question identified by paragraph separators and a rule engine, the specific question type, the extracted question stem content, and the user's answer content extracted from the corresponding area.

[0049] In some embodiments, the processor can traverse the assignment text content using a sliding window approach and perform question boundary recognition and deduplication merging algorithms to obtain candidate question units. Specifically, when the length of the assignment text content exceeds a preset length threshold, the assignment text content is segmented using a sliding window approach; during the segmentation process, paragraph separators, line breaks, or question number identifiers are preferentially used as segmentation positions to avoid truncating the question content, and then boundary recognition is performed on each text block.

[0050] Semantic and attribute features are multidimensional data vectors used to accurately measure the similarity, logical consistency, and matching correlation between question and answer texts in complex layouts. For example, semantic and attribute features may include extracted question number consistency classification identifiers, question type classification discrete labels, and deep continuous semantic feature text vector matrices transformed from question stems and answer content using natural language processing techniques.

[0051] In some embodiments, the processor can process text blocks to obtain semantic and attribute features by invoking a pre-trained text embedding model or a natural language feature extraction engine.

[0052] Matching success rate is a quantified percentage metric that reflects the accuracy and overall quality of the automatic correspondence between the entire sequence of task questions and the sequence of standard answers. For example, matching success rate can be the percentage ratio of the number of questions that successfully and with high confidence map to the standard answer segment to the total number of identified candidate question units in a full matching task.

[0053] In some embodiments, the processor can obtain the matching success rate by counting the number of successfully matched questions with a confidence level exceeding a set baseline in the preliminary matching results and dividing it by the total number of questions.

[0054] The preliminary matching results are a temporary list of question-answer correspondences calculated based on the initial feature vector similarity, but which have not yet undergone final threshold downgrading verification. For example, the preliminary matching results may include the index identifier record of the recommended standard answer fragment mapped to each candidate question unit, as well as the corresponding semantic similarity score or rule matching confidence score calculated by the system's underlying layer.

[0055] In some embodiments, the processor can obtain preliminary matching results by calculating the semantic and attribute similarity between the candidate question unit features and the standard answer fragments, and sorting and filtering them.

[0056] Preset thresholds are static benchmark reference values ​​set within the system to determine the status of a certain algorithm process step or to trigger specific degradation or rollback strategies. For example, preset thresholds may include percentage limit values ​​configured to determine whether the overall answer retrieval success rate is acceptable, in order to decide whether to roll back to the position alignment strategy, such as the system-set lower limit threshold θ parameter for matching success rate.

[0057] In some embodiments, the processor can obtain a preset threshold by reading parameters from the underlying system configuration file or by parsing the custom batching configuration carried in the received batching task request.

[0058] S3: Based on the content of the assignment text and the initial grading items, extract the question type information and score information and fill in the missing scores to obtain the completed grading items.

[0059] The completed grading entries are data structure records that, after undergoing score logic checks and intelligent completion, possess the complete scoring dimension parameters required for subsequent intelligent scoring. For example, the completed grading entries may include the question number, question type, question stem content, user's answer, aligned standard answer, and a numerical field containing the exact full score information after rule priority derivation.

[0060] In some embodiments, the processor can obtain a completed correction entry by performing auto-completion on the initial correction entry with missing score fields according to the default score or the priority rule for calculating the score.

[0061] In some embodiments, the processor can extract multi-dimensional score data and establish a mapping relationship based on the assignment text content and initial grading entries to obtain a question type score mapping table; based on the question type score mapping table, identify the score missing status in the initial grading entries to obtain missing score entries; based on the missing score entries, automatically complete the score data of the corresponding level according to the priority rules of the preset score source to obtain the completed grading entries.

[0062] The question type score mapping table is a structured data dictionary that records the score allocation benchmark rules and global conversion relationships for each question type in the entire test paper. For example, the question type score mapping table may include the fixed basic score for each multiple-choice question extracted from the test paper text, the arithmetic conversion relationship between the total score of the major questions and the total number of questions, and exception score dictionary entries for individual specific questions.

[0063] In some embodiments, the processor can extract information from the question type titles and score description paragraphs of the assignment text content using regular expression matching technology and construct an association to obtain a question type score mapping table.

[0064] Missing score entries are abnormal or incomplete question data that could not be clearly identified from the surrounding text of the question during the initial structured extraction stage. For example, missing score entries may include the initial grading record set where the score description is not marked in the title of the major question and the question itself does not have a separate score label.

[0065] In some embodiments, the processor can obtain missing score entries by traversing the generated initial set of correction entries and verifying the integrity and validity of the score field in each entry record.

[0066] S4: Based on the completed correction entries, construct a correction task list using a stateful workflow and perform batch parallel scheduling to obtain the distributed correction tasks.

[0067] Stateful workflows are distributed control engine frameworks used to manage the lifecycle of complex concurrent tasks, schedule task batch flow, record intermediate results, and uniformly monitor abnormal states. For example, a stateful workflow may include configuration parsing nodes, planning nodes responsible for task splitting, distribution nodes responsible for task dispatching, batching calculation nodes for performing single-question scoring, and result auditing data flow nodes that record task execution status and summarize the results.

[0068] In some embodiments, the processor can obtain a stateful workflow by instantiating a pre-defined distributed task scheduling framework code or a microservice orchestration component.

[0069] The grading task list is a sequence of standardized, independent, and fully informational grading task objects to be distributed to the underlying computational execution units. For example, the grading task list may include dozens of independent, standardized grading task units. Each unit strictly encapsulates the question type, question stem text, user answer text, standard answer text fragment, and the final determined full score parameter required for execution and scoring.

[0070] In some embodiments, the processor can obtain a list of correction tasks by performing specific JSON format conversion and data encapsulation on the completed correction entries and writing them sequentially into a queue.

[0071] The distributed batching task is a data packet containing task instances that have been assigned by the system scheduling center to specific computing execution nodes and are either preparing to or waiting to initiate a model invocation request. For example, a distributed batching task may include a text data packet structure containing the IP address or thread identifier of a specific parallel execution node, the current batch number record, and the input payload text packet structure required for specific performance evaluation.

[0072] In some embodiments, the processor can use a stateful workflow task distribution control module to push batches of batch modification tasks to the parallel execution unit to obtain the distributed batch modification tasks according to the queue index.

[0073] In some embodiments, the processor can divide the task into multiple batches of batching tasks based on the completed batching entries using a preset maximum parallelism parameter; based on the multiple batching task batches, state data containing the current execution batch identifier and execution exception record is constructed in the stateful workflow; based on the state data, the current batch of tasks is issued and the corresponding state is updated synchronously using the parallel execution unit of the stateful workflow to obtain the distributed batching tasks.

[0074] The purpose of batching grading tasks is to control the overall concurrent request pressure of the system by dividing the overall grading task into several independent task execution groups according to a specific maximum concurrency strategy. For example, a batch of grading tasks may include a data sequence of multiple task subsets calculated based on a preset maximum number of concurrent threads, as well as a unique batch number hash index used to track the execution progress and status lifecycle of the batch throughout the workflow.

[0075] In some embodiments, the processor can obtain batches of batch tasks by performing a list sharding algorithm on the overall set of batch task lists according to a preset maximum parallelism parameter.

[0076] Status data is a dynamic, structured set of monitoring information recorded by the workflow engine during continuous execution, including the overall grading process, the status of task sub-nodes, and the environment context. For example, status data may include overall grading configuration information, an index of remaining pending tasks, the batch identifier currently being executed, real-time feedback on the success or failure of each individual task's grading, and detailed execution error log information.

[0077] In some embodiments, the processor can obtain state data by listening to network callback signals of task execution nodes in a stateful workflow framework and persisting them to a cache.

[0078] A parallel execution unit is an underlying system computing resource or logical thread pool used to concurrently process multiple single-question scoring reasoning calculations, initiate model calls, or execute rule-based fallback calculations. For example, a parallel execution unit may include multiple independent execution container instances deployed on a cloud server cluster, or a background worker thread group structure in a local computer system responsible for initiating asynchronous network calls.

[0079] In some embodiments, the processor can dynamically allocate CPU computing cores or instantiate independent network request threads to obtain parallel execution units by calling the resource scheduling interface of the underlying operating system.

[0080] S5: Based on the distributed grading task, call the second language model for scoring, and perform hierarchical downgrading scoring when the second language model outputs an anomaly to obtain the grading result for a single question.

[0081] The second major language model is a reasoning model specifically optimized for understanding the reasoning logic of complex questions, comparing semantic consistency, and generating final numerical scores and detailed evaluation reasons. For example, the second major language model may include a generative AI network model that, after receiving structured prompt input, can return specific numerical scores and detailed textual evaluation reasons in a specified structured JSON object format.

[0082] In some embodiments, the processor can obtain a second large language model by calling a pre-configured external large language model service API network interface or by loading local inference weights.

[0083] The single-question grading result is the final closed-loop data structure generated for a specific question in the assignment text after completing all model inference judgments and possible fallback calculations. For example, the single-question grading result may include a question type classification field, the final calculated score, the generated scoring reasoning text, the recorded error code, and a fallback level identifier indicating whether a rule-based fallback operation was performed.

[0084] In some embodiments, the processor can obtain the single-question grading result by parsing the structured output response of the second language model, or by executing a hierarchical rule fallback strategy and performing calculations when an exception occurs.

[0085] In some embodiments, the processor can score and perform format anomaly detection based on the structured output mode of the second language model based on the distributed grading task to obtain first scoring status data; when the first scoring status data indicates an anomaly, text parsing and field integrity verification are performed based on the original text returned by the second language model to obtain second scoring status data; when the second scoring status data indicates an anomaly, fallback calculation processing is performed based on the preset rule scoring strategy and score validity constraints are executed to obtain the single-question grading result.

[0086] In some embodiments, the fallback calculation process based on the preset rule scoring strategy includes: when the question type is a multiple choice question, comparing the consistency between the user's answer and the standard answer; when the question type is a true / false question, converting the user's answer and the standard answer into a sequence of true / false values ​​and then comparing them; when the question type is a fill-in-the-blank question, performing normalized text matching on the user's answer and the standard answer and calculating the similarity.

[0087] The first scoring status data is a low-level legality check report used to characterize whether the structured message initially output by the large language model conforms to the expected parsing specifications and system field constraints. For example, the first scoring status data may include low-level anomaly flags indicating whether JSON data structure deserialization failed, whether necessary core scoring fields are missing, whether field data type mapping is abnormal, and whether the actual score exceeds the preset full score range.

[0088] In some embodiments, the processor can obtain the first scoring status data by performing format validation rules on the structured output string results directly returned by the second language model.

[0089] The raw text is the underlying raw character stream sequence data returned by the second-largest language model during network response generation, which has not been filtered by the system's structured logic, extracted by regular expressions, or validated by data. For example, the raw text may include non-standard punctuation marks directly output by the model, incomplete JSON message fragments due to instability in the output process, and mixed strings of actual numerical scores and textual explanations.

[0090] In some embodiments, the processor can obtain the raw text by capturing and extracting the content of the underlying HTTP protocol response packet body returned by the second largest language model API network call.

[0091] The second scoring status data represents the detection result of the field data integrity after a secondary extraction operation on the unstructured raw text returned by the model following a failure in the structured parsing path. For example, the second scoring status data may include Boolean flags indicating whether the key actual score value field extracted by regular expressions has been effectively obtained, and whether the evaluation reason field has been successfully located and truncated, etc., to verify the validity of field extraction.

[0092] In some embodiments, the processor can obtain second scoring status data by applying regular expression matching or keyword semantic analysis algorithms to the original text data and then performing numerical verification.

[0093] Score validity constraints are mathematical restrictions or procedural operations that ensure the final score for each question falls within a reasonable range according to actual business logic. For example, score validity constraints may include the mathematical constraint process of forcibly truncating any score result that exceeds the theoretical range and using comparison logic to limit it to between zero and the maximum score marked for that question.

[0094] In some embodiments, the processor can obtain score validity constraints by inputting the initially obtained numerical score value into the preset score range comparison algorithm module code for maximum and minimum boundary comparison correction.

[0095] In some embodiments, the grading result for a single question includes a score, the expression for which is: ; in, This represents the final actual score after constraints. This indicates the first language model obtained from the second largest language model. The original score of the question. This represents the function that takes the maximum value. This represents the function that takes the minimum value. This indicates the number obtained according to the question type score mapping table. The full marks for each question.

[0096] S6: Based on the multiple individual question grading results, the data is summarized to obtain the overall grading result, and the intelligent homework grading is completed.

[0097] The overall grading result is a comprehensive statistical data set encompassing all aspects of a complete exam grading task, used for final front-end presentation or persistent archiving in the system, and includes detailed results for each question. For example, the overall grading result may include a JSON list of detailed grading results for each question arranged in the original layout of the exam paper, the total score calculated by summing all actual scores, the theoretical maximum score, the calculated overall score percentage, and statistical data on abnormal situations.

[0098] In some embodiments, the processor can obtain the overall grading result by traversing, sorting, accumulating, and performing comprehensive division on the aggregated data structure of multiple individual question grading results. The grading results are statistically analyzed in Table 1.

[0099] Table 1. Statistical Table of Correction Results by Item

[0100] In some embodiments, the expression for the total score in the data aggregation to obtain the overall grading result is: ; in, This indicates the total score corresponding to the current grading task request. This indicates the total number of individual question corrections included in the overall correction results. This indicates the first value obtained based on the score validity constraint. The actual score for each question. This represents the function for summing.

[0101] In some embodiments, the expression for the total score in the data aggregation to obtain the overall grading result is: ; in, This indicates the total score corresponding to the current grading task request. This indicates the total number of individual question corrections included in the overall correction results. This indicates the number obtained according to the question type score mapping table. The full marks for this question, This represents the function for summing.

[0102] In some embodiments, the expression for the score rate in the overall grading result obtained by summarizing the data is: ; in, This indicates the overall score rate corresponding to the current grading task request. This represents the total score obtained from the above calculation. This represents the total score obtained from the above calculation.

[0103] In synchronous grading mode, grading progress and individual question grading results can be pushed in real time by sending events from the server; in asynchronous grading mode, structured grading results can be stored in the database, and grading progress events can be stored in the cache medium for users to query based on the grading task identifier.

[0104] In some embodiments, to verify the effectiveness and stability of the intelligent homework grading method provided by the present invention, real course exam papers containing multiple-choice questions, fill-in-the-blank questions, true / false questions, calculation questions, and essay questions were used for testing, and the grading time for each question type was statistically analyzed. The grading time for each question type is shown in Table 2.

[0105] Table 2. Statistics on the time spent grading individual questions of different question types

[0106] Test results show that the present invention can automatically parse complex mixed-type assignments, process the questions in a structured manner, match standard answers, and provide intelligent scoring. In batch grading scenarios, batch parallel scheduling can effectively reduce the overall grading time. Even when the output of the large language model is abnormal, the hierarchical degradation scoring mechanism can still generate valid scoring results, thereby improving overall stability and fault tolerance.

Claims

1. An intelligent homework grading method based on multi-modal analysis and hierarchical degradation scoring, characterized in that, include: S1: Obtain the grading task request. Based on the homework file to be graded and the standard answer file contained in the grading task request, perform concurrent parsing using a multimodal parsing method containing the first major language model to obtain the homework text content and the standard answer text content. S2: Based on the homework text content and the standard answer text content, the initial grading items are obtained by using the question structuring and matching alignment strategy; S3: Based on the homework text content and the initial grading items, extract the question type information and score information and fill in the missing scores to obtain the completed grading items; S4: Based on the completed correction entries, construct a correction task list using a stateful workflow and perform batch parallel scheduling to obtain the distributed correction tasks. S5: Based on the distributed grading task, call the second language model to score, and perform hierarchical downgrade scoring when the second language model outputs an anomaly to obtain the grading result for a single question. S6: Based on the multiple individual question grading results, the data is summarized to obtain the overall grading result, and the intelligent homework grading is completed.

2. The intelligent homework grading method based on multi-modal analysis and hierarchical degradation scoring according to claim 1, characterized in that, S1 includes: Get the grading task request; Based on the homework file to be graded and the standard answer file included in the grading task request, pagination rendering is performed and the layout structure features and position association information of adjacent pages are extracted to obtain the corresponding cross-page association area data. Based on the cross-page association region data, a cross-page joint context input structure is constructed and the first large language model is used to perform continuous content recognition to obtain the page parsing results of each page; Based on the page parsing results, the page order is spliced ​​and merged to obtain the homework text content and the standard answer text content, respectively.

3. The intelligent homework grading method based on multimodal analysis and hierarchical degraded scoring according to claim 1, characterized in that, S2 includes: Based on the content of the assignment text, question boundary identification and information extraction are performed to obtain candidate question units; Based on the candidate question units and standard answer text content, semantic and attribute features are used for matching and the matching success rate is calculated to obtain preliminary matching results; When the matching success rate of the preliminary matching results is lower than a preset threshold, the interval mapping is re-mapped based on the question order features using a position alignment strategy to obtain the initial grading entries.

4. The intelligent assignment grading method based on multi-modal analysis and hierarchical degrading scoring according to claim 1, characterized in that, S3 includes: Based on the content of the assignment text and the initial grading items, multi-dimensional score data is extracted and a mapping relationship is established to obtain a question type score mapping table; Based on the question type score mapping table, identify the score missing status in the initial grading items to obtain the missing score items; Based on the missing score entries, the score data of the corresponding level is matched sequentially according to the priority rules of the preset score source for automatic completion processing, resulting in the completed correction entries.

5. The intelligent assignment grading method based on multi-modal analysis and hierarchical degrading scoring according to claim 1, characterized in that, S4 includes: Based on the completed revise entries, the task is divided using a preset maximum parallel number parameter to obtain multiple revise task batches. Based on the multiple batches of batch modification tasks, state data containing the current batch identifier and execution exception records is constructed in the stateful workflow; Based on the state data, the current batch of tasks is issued and the corresponding state is updated synchronously using the parallel execution unit of the stateful workflow to obtain the distributed batch modification tasks.

6. The intelligent assignment grading method based on multi-modal analysis and hierarchical degrading scoring according to claim 1, characterized in that, S5 includes: Based on the distributed grading task, the structured output mode of the second language model is used for scoring and format anomaly detection is performed to obtain the first scoring status data. When the first rating status data indicates an anomaly, text parsing and field integrity verification are performed based on the original text returned by the second large language model to obtain the second rating status data; When the second scoring status data indicates an anomaly, a fallback calculation is performed based on the preset scoring strategy, and the score validity constraint is applied to obtain the single-question grading result.

7. The intelligent homework grading method based on multi-modal analysis and hierarchical degradation scoring according to claim 6, characterized in that, The individual question grading result includes the actual score, and the expression for the actual score is: ； in, This represents the final actual score after constraints. This indicates the first language model obtained from the second largest language model. The original score of the question. This represents the function that takes the maximum value. This represents the function that takes the minimum value. This indicates the number obtained according to the question type score mapping table. The full marks for each question.

Images