Intelligent paper marking method based on knowledge graph fusion large model

Abstract

The invention discloses an intelligent paper marking method based on a knowledge graph fusion large model. The intelligent paper marking method comprises the following steps: step a, inputting test questions to be marked and corresponding student answers; b, calling a pre-constructed subject knowledge graph, wherein the knowledge graph comprises knowledge points related to the test questions and a semantic association relationship of the knowledge points; c, performing semantic analysis on the student answers by using a deep learning large model, extracting key information and matching the key information with knowledge points in a knowledge graph; step d, based on a semantic matching result and weight distribution in the knowledge graph, scoring the answers through a multi-dimensional scoring model; e, generating a feedback report containing an interpretable scoring basis according to a scoring result; step f, receiving an intervention instruction of the teacher for the scoring result, dynamically adjusting the scoring parameter and recalculating the score; and g, outputting the final scoring result and the feedback report to a user interaction interface. According to the invention, the paper marking efficiency, accuracy and intelligent level can be obviously improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent examination marking technology, and in particular to an intelligent examination marking method based on a knowledge graph fusion large model. Background Art

[0002] The traditional manual marking method currently used in educational assessment suffers from significant drawbacks. First, for large-scale exams, manual review can take weeks, making it inefficient. For example, the average marking time for a single question in the Gaokao Chinese language essay section is less than 90 seconds. Second, subjective differences in examiner perception lead to grading standard deviations as high as 15%-20%. Third, scoring consistency for complex question types such as open-ended and essay questions is insufficient, with cross-group scoring deviations exceeding 30%. Existing automated marking technologies often rely on rule-based template-based keyword matching, which has the following limitations: 1. Difficulty understanding semantics: Students' answers vary widely, making it difficult for traditional algorithms to accurately match semantics. 2. Difficulty analyzing logic: Open-ended questions require evaluation of answers' completeness, logic, and innovation. 3. Difficulty dynamically adjusting scoring standards: Scoring weights for different questions must be adaptively adjusted to ensure fairness. Therefore, an intelligent marking solution that integrates deep semantic understanding with a dynamic scoring mechanism is urgently needed. Summary of the Invention

[0003] The purpose of the present invention is to provide an intelligent examination marking method based on a knowledge graph fusion model. The present invention can significantly improve the efficiency, accuracy and intelligence level of examination marking.

[0004] The technical solution of the present invention is an intelligent examination marking method based on a knowledge graph fusion model, comprising the following steps:

[0005] Step a: Enter the test questions to be reviewed and the corresponding student answers;

[0006] Step b: calling a pre-built subject knowledge graph, wherein the knowledge graph includes test question-related knowledge points and their semantic associations;

[0007] Step c: Use a deep learning model to perform semantic analysis on the student's answer, extract key information, and match it with knowledge points in the knowledge graph;

[0008] Step d: Based on the semantic matching results and the weight distribution in the knowledge graph, the answers are scored using a multi-dimensional scoring model;

[0009] Step e: Generate a feedback report containing the basis for interpretability scoring based on the scoring results;

[0010] Step f: receiving the teacher's intervention instructions on the scoring results, dynamically adjusting the scoring parameters and recalculating the scores;

[0011] Step g: Output the final scoring results and feedback report to the user interaction interface.

[0012] In the above-mentioned intelligent marking method based on the knowledge graph fusion large model, the subject knowledge graph construction process in step b specifically includes:

[0013] Step b1: extract knowledge point entity relationship multi-tuples by establishing a subject corpus, perform knowledge fusion, and store them in the model layer and data layer of the knowledge base;

[0014] Step b2: performing semantic dependency analysis on the multi-tuple using a bidirectional long short-term memory network to construct a semantic dependency graph;

[0015] Step b3: Calculate the semantic similarity between the entity and the candidate entity through the graph attention network and select the optimal entity link result;

[0016] Step b4: Store the structured knowledge graph in the Neo4j database.

[0017] In the aforementioned intelligent examination marking method based on the knowledge graph fusion large model, the semantic parsing and matching process of step c includes:

[0018] Step c1: Use the BERT or GPT model to perform word segmentation and semantic analysis on the answer text;

[0019] Step c2: extracting the logical structure of the answer based on dependency syntactic analysis and generating a syntactic dependency tree;

[0020] Step c3: Match the key information obtained by parsing with the knowledge graph nodes, and calculate the semantic matching degree using the cosine similarity algorithm.

[0021] In the aforementioned intelligent marking method based on the knowledge graph fusion model, the multi-dimensional scoring model in step d includes the following scoring dimensions that work together:

[0022] The correctness scoring module calculates the score based on the matching degree between the key information of the answer and the knowledge graph nodes, combined with the preset weights;

[0023] The logic scoring module evaluates the structural integrity of the answer by analyzing the syntactic dependency tree generated in step c2;

[0024] An innovative scoring module detects new viewpoints in answers that go beyond the preset scope of the knowledge graph but are consistent with the logic of the subject;

[0025] The language expression scoring module evaluates text coherence and grammatical correctness based on a pre-trained language model.

[0026] In the aforementioned intelligent examination marking method based on the knowledge graph fusion model, the feedback report generation process of step e specifically includes:

[0027] Step e1: Use the LIME algorithm to perform interpretability analysis on the scoring results and visually annotate key scoring segments;

[0028] Step e2: Generate a topological diagram of the knowledge graph matching path, showing the association path between the answer and the standard knowledge point;

[0029] Step e3: Present the weight distribution and contribution ratio of each scoring dimension in the report.

[0030] In the aforementioned intelligent examination marking method based on the knowledge graph fusion large model, the dynamic adjustment operation of step f specifically includes:

[0031] Step f1, responding to the teacher's instructions to modify the weight distribution coefficient of a specific node in the knowledge graph;

[0032] Step f2: Adjust the weight of each scoring dimension in the multi-dimensional scoring model;

[0033] Step f3: Update the semantic matching threshold parameters of the deep learning large model and re-execute steps c to d.

[0034] The aforementioned intelligent marking method based on the knowledge graph fusion model also includes an adaptive scoring mechanism during the execution of step d:

[0035] Monitor the characteristic distribution of students' answers in real time. When abnormal answer patterns are detected, the scoring rules are automatically triggered to adjust dynamically and the scoring parameters are adjusted according to the preset strategy.

[0036] In the aforementioned intelligent examination marking method based on the knowledge graph fusion model, the user interaction interface of step g provides the following functional modules:

[0037] Input module, used to support batch uploading and version management of test papers and reference answers;

[0038] Visual analysis module, used to compare intelligent scoring and manual scoring results;

[0039] Statistical chart generation module, used to display statistical charts of historical review data trends;

[0040] The operation module is used to provide interactive operations for collaborative editing and annotation of scoring results.

[0041] The aforementioned intelligent examination marking method based on the knowledge graph fusion large model, the knowledge graph construction process also includes the keyword vectorization processing step:

[0042] Step b5: Obtain vector representations of knowledge point keywords from the pre-trained word vector model;

[0043] Step b6: Calculate the average word vectors VX, VY, and VZ of the knowledge points and establish a mapping relationship between the keyword set and the knowledge graph nodes;

[0044] Step b7: Input the vector representation as an auxiliary feature for semantic matching into the parsing process of step c.

[0045] Compared with existing technologies, the present invention deeply integrates the structured knowledge associations of knowledge graphs with the contextual reasoning capabilities of deep learning models, addressing the technical issues of fragmented semantic understanding and one-sided logical analysis in open-ended questions. The present invention constructs a dynamic scoring model based on multi-level semantic matching, achieving a multi-dimensional comprehensive evaluation from single keyword matching to "knowledge point relevance, logical integrity, and innovative value." Through a dynamic knowledge graph node weight allocation algorithm and a real-time detection mechanism for abnormal answer patterns, the present invention allows the system to adaptively adjust the scoring criteria for questions of different disciplines and difficulty levels. Simultaneously, combined with interpretability technology, the scoring path is traced for knowledge topology and key sentence attribution analysis, making the AI scoring basis transparent and verifiable. The present invention can also support teachers' dynamic modification of scoring rules, knowledge graph weights, and semantic matching thresholds, forming a collaborative review model with bidirectional optimization between humans and machines, significantly improving the efficiency of grading complex questions while ensuring the objectivity of scoring. The present invention can rapidly expand subject knowledge graphs and adapt to differentiated assessment needs, providing a unified technical framework for critical analysis of liberal arts essay questions and rigorous step evaluation of science deduction questions, promoting the development of intelligent education assessment to a deeper level and multiple scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 The figure is a schematic diagram of the application process of the intelligent marking method of the embodiment. DETAILED DESCRIPTION

[0047] The present invention will be further described below with reference to the accompanying drawings and examples, but they are not intended to limit the present invention.

[0048] Example: An intelligent examination marking method based on knowledge graph fusion large model, comprising the following steps:

[0049] Step a: Enter the test questions to be reviewed and the corresponding student answers;

[0050] Step b: calling a pre-built subject knowledge graph, wherein the knowledge graph includes test question-related knowledge points and their semantic associations;

[0051] Step c: Use a deep learning model to perform semantic analysis on the student's answer, extract key information, and match it with knowledge points in the knowledge graph;

[0052] Step d: Based on the semantic matching results and the weight distribution in the knowledge graph, the answers are scored using a multi-dimensional scoring model;

[0053] Step e: Generate a feedback report containing the basis for interpretability scoring based on the scoring results;

[0054] Step f: receiving the teacher's intervention instructions on the scoring results, dynamically adjusting the scoring parameters and recalculating the scores;

[0055] Step g: Output the final scoring results and feedback report to the user interaction interface.

[0056] In specific implementation, the construction of subject knowledge graphs can be achieved in a variety of ways, such as extracting knowledge point entity relationship multi-tuples through subject corpora and then performing knowledge fusion, or using graph databases for storage and management. The semantic parsing process can be implemented using pre-trained models such as BERT and GPT, where the word segmentation and semantic parsing of the answer text can be completed through the encoding layer of these models. The multi-dimensional scoring model can include scoring dimensions such as correctness, logic, innovation, and language expression, and the weight distribution of each dimension can be dynamically adjusted according to the specific question type. During the feedback report generation process, key scoring sentences can be visually annotated using the LIME algorithm, and the knowledge graph matching path can be displayed through a topological diagram. During the teacher intervention phase, the weight distribution coefficient of a specific node can be directly modified through the interactive interface, and the semantic matching threshold parameters can be updated based on new teaching needs.

[0057] This approach, through the synergy of knowledge graphs and large models, effectively addresses the technical challenges of traditional grading methods in terms of semantic understanding, logical analysis, and dynamic adjustment. The knowledge graph provides a structured subject knowledge system, making answer evaluation interpretable; the deep learning large model enables in-depth semantic analysis of student answers, overcoming the limitations of traditional keyword matching methods; and the multi-dimensional scoring model, combined with a dynamic adjustment mechanism, ensures adaptability and fairness in scoring across different question types. As a result, grading efficiency has been significantly improved, the lack of scoring consistency has been effectively addressed, and teachers have been provided with an intelligent, interventional grading tool.

[0058] Furthermore, in this embodiment, the subject knowledge graph construction process is as follows:

[0059] Step b1: extract knowledge point entity relationship multi-tuples by establishing a subject corpus, perform knowledge fusion, and store them in the model layer and data layer of the knowledge base;

[0060] Step b2: performing semantic dependency analysis on the multi-tuple using a bidirectional long short-term memory network to construct a semantic dependency graph;

[0061] Step b3: Calculate the semantic similarity between the entity and the candidate entity through the graph attention network and select the optimal entity link result;

[0062] Step b4: Store the structured knowledge graph in the Neo4j database.

[0063] Specifically, by establishing a subject corpus, knowledge point entity relationship tuples are extracted, and after knowledge fusion, they are stored in the model layer and data layer of the knowledge base. Among them, the model layer is used to define the type and relationship structure of knowledge points, and the data layer is used to store specific knowledge point instances and their associations. A bidirectional long short-term memory network is used to perform semantic dependency analysis on the tuples and construct a semantic dependency graph. As a preferred embodiment, the bidirectional long short-term memory network can adopt the BiLSTM-CRF model to identify the semantic dependency relationship between knowledge points by capturing contextual information. The semantic similarity between the entity and the candidate entity is calculated through the graph attention network, and the optimal entity link result is selected. For example, the graph attention network can adopt the GAT model to calculate the semantic association strength between entities through a multi-head attention mechanism. The structured knowledge graph is stored in the Neo4j database, and its graph database characteristics are used to achieve efficient knowledge point retrieval and relationship traversal.

[0064] Furthermore, the knowledge graph construction process also includes keyword vectorization processing steps, as follows:

[0065] Step b5: Obtain vector representations of knowledge point keywords from the pre-trained word vector model;

[0066] Step b6: Calculate the average word vectors VX, VY, and VZ of the knowledge points and establish a mapping relationship between the keyword set and the knowledge graph nodes;

[0067] Step b7: Input the vector representation as an auxiliary feature for semantic matching into the parsing process of step c.

[0068] Keyword vectorization can be achieved using pre-trained models such as Word2Vec, GloVe, or FastText to obtain word vectors. Average word vectors can be calculated using arithmetic averaging, weighted averaging, or dynamic weighting based on an attention mechanism. Mapping relationships can be established by creating an inverted index or constructing a vector similarity retrieval library. Auxiliary feature input can be fused into the semantic matching module through feature concatenation or an attention mechanism.

[0069] Thus, by constructing a structured subject knowledge graph, the problem of insufficient semantic understanding in traditional automatic marking technology is solved. Specifically, by combining a bidirectional long short-term memory network and a graph attention network, accurate modeling of semantic relationships between knowledge points is achieved. Compared with the existing technology, this solution can more accurately capture the semantic association between student answers and standard knowledge points, providing a reliable knowledge basis for subsequent intelligent scoring. By introducing word vector representation, the semantic representation ability of the knowledge graph is enhanced. Specifically, traditional knowledge graphs mainly rely on symbolic representation and have difficulty handling semantic association problems such as synonyms and near-synonyms. By converting knowledge point keywords into continuous vector space representations, knowledge points with similar semantics but different expressions can be quantitatively matched through vector distance. This not only improves the accuracy of the subsequent semantic parsing link, but also solves the problem of matching between students' diverse expressions and standard knowledge points in open-ended questions. Compared with the existing keyword matching method based on rule templates, the distributed representation obtained by deep learning in the present invention can better capture the similarity at the semantic level, thereby improving the consistency and objectivity of scoring.

[0070] Furthermore, this embodiment uses the BERT or GPT model to perform word segmentation and semantic analysis on the answer text, wherein the BERT model captures the deep semantic features of the text through a multi-layer Transformer structure, and the GPT model uses an autoregressive mechanism to generate a coherent semantic representation. The logical structure of the answer is extracted based on dependency syntactic analysis, and specifically a syntactic dependency tree is generated by the Stanford Core NLP tool. This tree structure can clearly show the modification relationship and logical hierarchy of each component in the answer. The key information obtained by the analysis is matched with the knowledge graph nodes, and the semantic matching degree is calculated by the cosine similarity algorithm, wherein the knowledge graph nodes are represented by pre-trained entity vectors, and the similarity threshold is set to 0.75 to achieve accurate screening. In addition, the semantic parsing stage can be implemented in the following variants: when processing mathematical questions, it can be replaced by the MathBERT model for formula parsing; for multimodal answers, a visual feature extractor can be combined to process mixed text and image content. Syntactic dependency analysis can be extended to a hierarchical parsing method based on graph neural networks to enhance the modeling ability of long text structures. The semantic matching link can introduce an adversarial training mechanism to improve the robustness of the matching model by generating an adversarial network. Therefore, the present invention effectively solves the semantic ambiguity problem of answers to open questions by combining deep semantic analysis with structured matching. Compared with traditional keyword matching methods, it can accurately identify the implicit logical relationships and innovative expressions in students' answers. Through the calculation of quantifiable semantic matching, it provides an objective basis for subsequent multi-dimensional scoring, while retaining the technical interface for manual review and adjustment of the scoring results.

[0071] Furthermore, the multi-dimensional scoring model of this embodiment includes the following scoring dimensions that work together:

[0072] Correctness scoring module: Calculates scores based on the matching degree between the key information of the answer and the knowledge graph nodes, combined with preset weights;

[0073] Logical scoring module: evaluates the structural integrity of the answer by analyzing the syntactic dependency tree;

[0074] Innovation Scoring Module: Detects new ideas in answers that go beyond the preset scope of the knowledge graph but are consistent with the subject logic;

[0075] Language expression scoring module: Evaluates text coherence and grammatical correctness based on a pre-trained language model.

[0076] Specifically, the correctness scoring module can use the cosine similarity algorithm to calculate the semantic matching degree between the student's answer and the standard knowledge point. The matching threshold can be set to the range of 0.7-0.9 and dynamically adjusted according to different question types. The logic scoring module scores by traversing the subject-verb-object structure integrity of the syntactic dependency tree, and deducts points proportionally when the core components are missing. The innovation scoring module is achieved by comparing the difference between the student's answer and the coverage of the knowledge graph. When valid viewpoints that are not included but verified by subject experts are detected, additional points are given. The language expression scoring module can use the probability distribution output by the masked language modeling task of the BERT model to deduct points for grammatical errors and semantically incoherent sentences. Therefore, the present invention solves the comprehensive problem that existing automatic marking technology is difficult to comprehensively evaluate the semantic accuracy, logical structure, innovation and language quality of answers by establishing a four-dimensional collaborative scoring mechanism. Compared with the traditional rule-based template scoring method, this solution ensures the objectivity of scoring through knowledge graph matching, improves the evaluation effect of open questions by combining syntactic analysis and innovation detection, and improves the accuracy of language expression judgment by using pre-training models. The weights of each module can be dynamically configured according to the characteristics of different question types. For example, essay questions can increase the weights of logic and innovation, while fill-in-the-blank questions focus on correctness scoring.

[0077] Furthermore, this embodiment also proposes a specific implementation scheme for the feedback report generation process, which is as follows:

[0078] Step e1: Use the LIME algorithm to perform interpretable analysis on the scoring results and visually annotate key scoring segments. The following implementation method is used: first, the output results of the deep learning large model are used as input, local explanation samples are generated by perturbing the input features, and the relationship between these samples and the model prediction results is fitted using a linear model. Finally, the key segments that have the greatest impact on the score are visually annotated in the answer text.

[0079] Step e2: Generate a topological diagram of the knowledge graph matching path, showing the path between the answer and the standard knowledge point. This topological diagram can be generated using the Graphviz tool. By parsing the node relationship data in the Neo4j database, a directed graph containing the matching paths between the standard knowledge points and the student's answers is automatically generated.

[0080] Step e3: Present the weight distribution and contribution ratio of each scoring dimension in the report. The weight distribution and contribution ratio of each scoring dimension can be presented through a combination of pie charts and bar charts, where the bar chart shows the raw score of each dimension and the pie chart shows the weight distribution ratio.

[0081] Specifically, the implementation of the LIME algorithm includes the following steps: performing word segmentation on the answer text and constructing a bag-of-words model; generating perturbation samples by randomly masking some words; recording changes in the prediction results of the large model on the perturbation samples; and training a local interpretable model to identify key words. Regarding knowledge graph visualization, the topology diagram generation process includes: extracting the knowledge point path that matches the student's answer from the knowledge graph; calculating the centrality index of each node in the path; adjusting the node's display size in the graph based on the centrality; and using different colors to distinguish core knowledge points from peripheral knowledge points. For scoring dimension analysis, the report generation module extracts the raw score data for each dimension from the scoring log, calculates the contribution ratio based on preset weights, and visualizes the data using the Matplotlib library. Thus, by combining interpretability algorithms with visualization technology, this invention effectively addresses the lack of transparency in traditional automated grading systems. Specifically, the LIME algorithm transforms the black-box model into an interpretable local linear model, making the scoring basis transparent. The knowledge graph visualization intuitively displays the path connecting the answer and the standard knowledge points, helping to understand the scoring logic. Multi-dimensional weight analysis clearly reveals the influence of each scoring factor on the final result.

[0082] Furthermore, this embodiment also proposes a technical solution that includes an adaptive scoring mechanism in the scoring process. By real-time monitoring the feature distribution of student answers, when an abnormal answer pattern is detected, the dynamic adjustment of the scoring rules is automatically triggered, and the scoring parameters are adjusted according to the preset strategy. Specifically, the adaptive scoring mechanism includes the following implementation methods: real-time cluster analysis of the answer text features is performed through statistical analysis methods to establish an answer pattern recognition model; the sliding window technology is used to dynamically calculate the deviation of the semantic feature distribution of the current batch of answers; a preset abnormal threshold trigger mechanism is set, and when it is detected that the feature distribution deviates from the baseline value by more than the preset threshold, the parameter adjustment process is automatically activated. As a preferred embodiment, the parameter adjustment may include: predicting the optimal scoring weight combination based on the regression model obtained by historical data training; dynamically optimizing the semantic matching threshold through the reinforcement learning algorithm; and updating the contribution ratio of each scoring dimension using the Bayesian network. As a result, the present invention solves the technical problem that the existing automatic marking system is difficult to adapt to abnormal answer scenarios. By introducing real-time monitoring and dynamic adjustment mechanisms, the scoring system can automatically optimize the scoring strategy for unconventional answer patterns, avoiding scoring deviations caused by abnormal answers. Compared to existing grading methods based on fixed rules, this invention significantly improves the grading system's adaptability to diverse answering patterns, ensuring the stability and reliability of grading results. Specifically, through data-driven adaptive adjustments, this mechanism effectively reduces grading fluctuations caused by differences in students' answering styles and reduces the frequency of manual intervention.

[0083] Furthermore, this embodiment also proposes specific implementations of the functional modules provided by the user interaction interface. The input module uses a distributed file storage system to enable batch uploading of test papers and reference answers, and manages different batches of test data through a version control algorithm, employing the Git version control principle to enable historical version tracing and difference comparison. The visualization analysis module integrates the D3.js framework to construct radar charts and heat maps, dynamically displaying the differences in sub-indicators between intelligent scoring and manual scoring. Specifically, the deviation between the two scoring results is calculated using the Euclidean distance algorithm and presented as a color gradient. The statistical chart generation module is implemented based on Apache ECharts and uses a time series prediction model to analyze historical review data trends. For example, the ARIMA algorithm is used to predict future changes in score distribution. The operation module uses collaborative editing technology, employing the OperationalTransformation algorithm to resolve conflicts between multiple users' concurrent annotations and synchronizing the modified status of scoring results in real time via the WebSocket protocol. Thus, the present invention addresses the poor interactivity of traditional grading systems by constructing a multifunctional integrated interface. The version management function ensures the consistency of scoring standards, the visual comparison function improves the transparency of the scoring process, the historical data analysis function assists in optimizing teaching strategies, and the collaborative editing function enables human-computer collaborative decision-making.

[0084] Based on the above-mentioned intelligent marking method, this embodiment also provides an application of the intelligent marking method. Figure 1 , the steps are as follows:

[0085] 1. Test Paper Upload: Teachers upload test papers in batches. This step provides the basic data source for the subsequent marking process. Uploading multiple test papers at once can improve test paper entry efficiency and reduce the tedious operation of uploading each test paper one by one.

[0086] 2. Format Check: The system determines whether the exam format meets the requirements. If not, the teacher must re-upload the exam. If it does, the system proceeds to the next step. This step prevents errors in the subsequent grading process caused by formatting issues and ensures smooth grading. For example, the system checks whether the file type and layout specifications meet the established requirements.

[0087] 3. Test paper processing: Automatically crop correctly formatted test papers to streamline their content. By removing excess blank areas and borders, the system can accurately identify question and answer areas, paving the way for subsequent intelligent grading.

[0088] 4. Set rules: Determine the marking rules that will serve as the basis for subsequent grading. Specifically set scoring rules and deduction criteria for different question types. These rules serve as the criteria for the intelligent grading system to determine the scores of students' answers.

[0089] 5. Marking: Knowledge graphs are used to identify key points and compare student answers with standard answers for scoring. Knowledge graphs store subject knowledge points and their relationships, accurately locating the key points corresponding to the questions and providing the knowledge support for accurate AI scoring. Answer comparison is a key operation for determining the correctness and completeness of student answers, thereby generating preliminary scoring results.

[0090] 6. Manual Review: Manually check the grading results. If errors are found, the rules can be changed and re-graded. This step ensures the accuracy and fairness of the grading results. If inaccurate scoring or misjudgment is found in the intelligent grading, it can be corrected in a timely manner.

[0091] 7. Student Upload: Wait for the student to upload a sample paper (you can choose to upload a sample paper or batch upload) and confirm the uploaded content. This step provides sample data for subsequent score report generation and can also be used to verify the accuracy of the marking rules and intelligent marking.

[0092] 8. Score Report Generation: Generate a score report based on the grading results. The score report includes the student's score for each question, total score, ranking, and other information, comprehensively presenting the student's exam performance.

[0093] 9. Teaching Suggestion Generation: Generate personalized teaching suggestions based on the score report, determining whether they are for the class or for the individual. If they are for the class, the class score and ranking are calculated. These personalized teaching suggestions are sent to students, helping them understand their own shortcomings and improve their learning outcomes in a targeted manner. The calculation of class scores and rankings allows teachers to understand the overall learning level of the class and the students' position within the class.

[0094] 10. Data storage: Data such as wrong questions and corresponding knowledge points are stored in a database. This facilitates subsequent query and analysis at any time, providing data support for teaching research, student progress tracking, etc.

[0095] In summary, this invention, by deeply integrating the structured knowledge associations of knowledge graphs with the contextual reasoning capabilities of deep learning models, can address the technical issues of fragmented semantic understanding and one-sided logical analysis of open-ended questions. This invention can significantly improve the efficiency, accuracy, and intelligence of exam marking.

Claims

1. An intelligent examination marking method based on a knowledge graph fusion model, characterized by: The steps include: Step a: Enter the test questions to be reviewed and the corresponding student answers; Step b: calling a pre-built subject knowledge graph, wherein the knowledge graph includes test question-related knowledge points and their semantic associations; Step c: Use a deep learning model to perform semantic analysis on the student's answer, extract key information, and match it with knowledge points in the knowledge graph; Step d: Based on the semantic matching results and the weight distribution in the knowledge graph, the answers are scored using a multi-dimensional scoring model; Step e: Generate a feedback report containing the basis for interpretability scoring based on the scoring results; Step f: receiving the teacher's intervention instructions on the scoring results, dynamically adjusting the scoring parameters and recalculating the scores; Step g: Output the final scoring results and feedback report to the user interaction interface.

2. The intelligent examination paper marking method based on the knowledge graph fusion large model according to claim 1 is characterized in that: The subject knowledge graph construction process in step b specifically includes: Step b1: extract knowledge point entity relationship multi-tuples by establishing a subject corpus, perform knowledge fusion, and store them in the model layer and data layer of the knowledge base; Step b2: performing semantic dependency analysis on the multi-tuple using a bidirectional long short-term memory network to construct a semantic dependency graph; Step b3: Calculate the semantic similarity between the entity and the candidate entity through the graph attention network and select the optimal entity link result; Step b4: Store the structured knowledge graph in the Neo4j database.

3. The intelligent examination paper marking method based on the knowledge graph fusion large model according to claim 1 is characterized in that: The semantic parsing and matching process of step c includes: Step c1: Use the BERT or GPT model to perform word segmentation and semantic analysis on the answer text; Step c2: extracting the logical structure of the answer based on dependency syntactic analysis and generating a syntactic dependency tree; Step c3: Match the key information obtained by parsing with the knowledge graph nodes, and calculate the semantic matching degree using the cosine similarity algorithm.

4. The intelligent examination paper marking method based on the knowledge graph fusion large model according to claim 3 is characterized in that: The multi-dimensional scoring model in step d includes the following scoring dimensions that work together: The correctness scoring module calculates the score based on the matching degree between the key information of the answer and the knowledge graph nodes, combined with the preset weights; The logic scoring module evaluates the structural integrity of the answer by analyzing the syntactic dependency tree generated in step c2; An innovative scoring module detects new viewpoints in answers that go beyond the preset scope of the knowledge graph but are consistent with the logic of the subject; The language expression scoring module evaluates text coherence and grammatical correctness based on a pre-trained language model.

5. The intelligent examination paper marking method based on the knowledge graph fusion large model according to claim 1 is characterized in that: The feedback report generation process in step e specifically includes: Step e1: Use the LIME algorithm to perform interpretability analysis on the scoring results and visually annotate key scoring segments; Step e2: Generate a topological diagram of the knowledge graph matching path, showing the association path between the answer and the standard knowledge point; Step e3: Present the weight distribution and contribution ratio of each scoring dimension in the report.

6. The intelligent examination marking method based on the knowledge graph fusion large model according to claim 1 is characterized in that: The dynamic adjustment operation of step f specifically includes: Step f1, responding to the teacher's instructions to modify the weight distribution coefficient of a specific node in the knowledge graph; Step f2: Adjust the weight of each scoring dimension in the multi-dimensional scoring model; Step f3: Update the semantic matching threshold parameters of the deep learning large model and re-execute steps c to d.

7. The intelligent examination paper marking method based on the knowledge graph fusion large model according to claim 1 is characterized in that: The execution process of step d also includes an adaptive scoring mechanism: Monitor the characteristic distribution of students' answers in real time. When abnormal answer patterns are detected, the scoring rules are automatically triggered to adjust dynamically and the scoring parameters are adjusted according to the preset strategy.

8. The intelligent examination paper marking method based on the knowledge graph fusion large model according to claim 1 is characterized in that: The user interaction interface of step g provides the following functional modules: Input module, used to support batch uploading and version management of test papers and reference answers; Visual analysis module, used to compare intelligent scoring and manual scoring results; Statistical chart generation module, used to display statistical charts of historical review data trends; The operation module is used to provide interactive operations for collaborative editing and annotation of scoring results.

9. The intelligent examination paper marking method based on the knowledge graph fusion large model according to claim 2 is characterized in that: The knowledge graph construction process also includes the keyword vectorization processing step: Step b5: Obtain vector representations of knowledge point keywords from the pre-trained word vector model; Step b6: Calculate the average word vectors VX, VY, and VZ of the knowledge points and establish a mapping relationship between the keyword set and the knowledge graph nodes; Step b7: Input the vector representation as an auxiliary feature for semantic matching into the parsing process of step c.

Images