Online Examination Integrated Intelligent Grading Method and System

Abstract

This invention relates to the field of intelligent education technology, and discloses an integrated intelligent grading method and system for online examinations. The method includes: constructing a subject knowledge graph; assigning dynamic weight coefficients to entities in the subject knowledge graph; parsing the semantic paragraph set and keyword set from the answers to subjective questions; extracting entities and logical connectors corresponding to the subject knowledge graph from the semantic paragraph set; constructing a directed logic graph based on the semantic associations and logical connectors between entities; and generating a logical coherence score between the logical graph nodes of the directed logic graph using a graph attention network. This invention can improve the accuracy of integrated intelligent grading for online examinations.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, and in particular to an integrated intelligent marking method and system for online examinations. Background Technology

[0002] The level of intelligence in online exam marking systems directly impacts the fairness and scientific rigor of exam evaluation. Current marking technologies have significant shortcomings in the subjective question evaluation stage:

[0003] First, at the semantic understanding level, subjective question answers usually rely on keyword matching or simple semantic analysis, lacking in-depth analysis of the completeness of logical chains (such as causal and temporal relationships), which may result in answers that are "stuffed with keywords but logically chaotic" scoring inflated.

[0004] Secondly, in terms of knowledge representation, traditional methods rely excessively on word frequency statistics (such as TF-IDF) or literal string matching, making it difficult to handle semantic disambiguation issues related to subject-specific terminology. More significantly, existing systems generally employ static weight allocation strategies, failing to dynamically adjust the importance of test points based on the actual examination situation. Summary of the Invention

[0005] This invention provides an integrated intelligent marking method and system for online examinations, the main purpose of which is to solve the problem of low accuracy in integrated intelligent marking of online examinations.

[0006] To achieve the above objectives, the present invention provides an integrated intelligent marking method for online examinations, comprising:

[0007] S1. Construct a subject knowledge graph and assign dynamic weight coefficients to entities in the subject knowledge graph;

[0008] S2. Perform multi-granularity analysis on the answers to subjective questions based on semantic units to obtain the set of semantic paragraphs and the set of keywords for the answers to subjective questions;

[0009] S3. Extract entities and logical connectors corresponding to the subject knowledge graph from the semantic paragraph set, and construct a directed logic graph of the subjective question answer based on the semantic association between the entities and the logical connectors;

[0010] S4. Generate the relationship strength between the logical graph nodes of the directed logic graph through a graph attention network, and generate the logical coherence score of the subjective question answer based on the relationship strength;

[0011] S5. Combining the entity coverage score of the subject knowledge graph with the logical coherence score, calculate the final score of the subjective question answer based on the dynamic weighting coefficient.

[0012] Optionally, the construction of the subject knowledge graph includes:

[0013] Based on textbooks and past exam questions, the subject knowledge system is analyzed to obtain a set of knowledge points.

[0014] The logical relationships between the knowledge points are established through expert annotation.

[0015] A subject knowledge graph is constructed based on the entity set and the logical relationships.

[0016] Optionally, the step of performing multi-granularity parsing of subjective question answers based on semantic units to obtain a set of semantic paragraphs and a set of keywords for the subjective question answers includes:

[0017] Subjective question answers are divided into semantic paragraphs based on dependency syntax.

[0018] The BiLSTM-CRF model is used to identify each semantic paragraph in order to obtain the keyword set of the subjective question answer;

[0019] The keywords in the keyword set are semantically disambiguated using the matching strategy of the subject knowledge graph, and the hierarchical relationship between the semantic paragraphs is determined based on the semantically disambiguated keywords.

[0020] The set of semantic paragraphs for determining the answers to the subjective questions is determined based on the semantic paragraphs and the hierarchical relationships.

[0021] Optionally, constructing the directed logic graph of the subjective question answers based on the semantic relationships between the entities and the logical connectors includes:

[0022] Construct an initial logical graph using the entities as nodes and the logical connectors as edges;

[0023] The syntactic associations between the entities are verified based on the dependency syntax.

[0024] The graph pruning algorithm is used to remove the edges that fail the verification in the initial logic graph to obtain a directed logic graph with syntax verification.

[0025] Optionally, generating the relation strength between logical graph nodes of the directed logical graph through a graph attention network includes:

[0026] The nodes and edges in the directed logic graph are input into the graph attention network as initial data;

[0027] The initial attention score between each pair of entity nodes is calculated using the multi-head attention layer in the graph attention network.

[0028] The initial attention score is nonlinearly activated and normalized using the LeakyReLU function to obtain a normalized attention score.

[0029] The normalized attention scores are weighted and aggregated to obtain the aggregated node features between the logical graph nodes in the directed logical graph.

[0030] Numerical mapping is performed on the features of the aggregated nodes to obtain the relationship strength between the nodes of the logical graph.

[0031] Optionally, generating the logical coherence score for the subjective question answer based on the strength of the relationship includes:

[0032] Based on the relationship strength, the node features and edge weights in the directed logical graph are aggregated to obtain a global logical view;

[0033] The integrity score of the logical chain in the directed logical graph is generated based on the global logical view;

[0034] The completeness score is standardized to obtain the logical coherence score of the subjective question answer.

[0035] Optionally, the calculation steps for the entity coverage score are as follows:

[0036] Compare the entities corresponding to the subjective question answers with the entities corresponding to the standard answers;

[0037] Based on the comparison results, the Jaccard similarity between the entity corresponding to the subjective question answer and the entity of the standard answer is calculated, and the Jaccard similarity is used as the entity coverage score.

[0038] Optionally, the graph pruning algorithm is a pruning algorithm based on grammatical confidence scoring and semantic reasonableness verification.

[0039] Optionally, the allocation of the dynamic weighting coefficients is dynamically adjusted based on the question type and subject characteristics.

[0040] To address the aforementioned problems, the present invention also provides an integrated intelligent online examination grading system, characterized in that the system comprises:

[0041] The weight allocation module is used to construct a subject knowledge graph and assign dynamic weight coefficients to entities in the subject knowledge graph.

[0042] The multi-granularity parsing module is used to perform multi-granularity parsing on the answers to subjective questions based on semantic units, so as to obtain the set of semantic paragraphs and the set of keywords of the answers to subjective questions;

[0043] The logic graph construction module is used to extract entities and logical connectors corresponding to the subject knowledge graph from the semantic paragraph set, and construct a directed logic graph of the subjective question answer based on the semantic association between the entities and the logical connectors.

[0044] The coherence evaluation module is used to generate the relationship strength between the logic graph nodes of the directed logic graph through a graph attention network, and to generate a logical coherence score for the subjective question answer based on the relationship strength.

[0045] The final score generation module is used to combine the entity coverage score of the subject knowledge graph with the logical coherence score, and calculate the final score of the subjective question answer based on the dynamic weight coefficient.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. This invention is based on a multi-level logical relationship extraction model constructed using graph neural networks. It can automatically identify logical relationships such as cause and effect, transition, and progression in answer text, and quantify logical coherence through an attention mechanism. Furthermore, it uses a dynamically updated subject knowledge graph to achieve fine-grained differentiation of professional terms.

[0048] 2. This invention constructs a subject knowledge graph and assigns dynamic weight coefficients to entities. Through textbook analysis and statistics of past exam questions, it achieves adaptive adjustment of the importance of knowledge points, solving the problem that traditional static weights cannot reflect changes in exam focus and improving the relevance of entity coverage scoring.

[0049] 3. Dependency parsing and BiLSTM-CRF model are used to achieve multi-level semantic segmentation of subjective question answers. Keyword disambiguation is performed by combining subject knowledge graph. A directed logic graph is constructed based on entities and logical connectors. Invalid associations are filtered through graph pruning algorithm, which solves the problem of traditional keyword matching ignoring logical relationships and enhances the ability to analyze the logical structure of answers. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating an integrated intelligent marking method for online examinations provided in an embodiment of the present invention.

[0051] Figure 2 This is a functional module diagram of an integrated intelligent online examination and marking system provided in an embodiment of the present invention;

[0052] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0053] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0054] This application provides an integrated intelligent online examination marking method. The executing entity of the integrated intelligent online examination marking method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application embodiment: a server, a terminal, etc. In other words, the integrated intelligent online examination marking method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0055] Reference Figure 1 The diagram shown is a flowchart illustrating an integrated intelligent online examination grading method according to an embodiment of the present invention. In this embodiment, the integrated intelligent online examination grading method includes:

[0056] S1. Construct a subject knowledge graph and assign dynamic weight coefficients to the entities in the subject knowledge graph.

[0057] In this embodiment of the invention, the construction of the subject knowledge graph includes:

[0058] Based on textbooks and past exam questions, the subject knowledge system is analyzed to obtain a set of knowledge points.

[0059] The logical relationships between the knowledge points are established through expert annotation.

[0060] A subject knowledge graph is constructed based on the entity set and the logical relationships.

[0061] In detail, the analysis of a subject knowledge system is the process of extracting core knowledge points from textbook texts and past exam questions to form a structured set of entities.

[0062] In detail, textbooks are the authoritative carriers of subject knowledge, while past exam questions reflect the distribution and key points of the test points. The combination of the two can ensure the comprehensiveness and practicality of the entity set.

[0063] In detail, Natural Language Processing (NLP) technology was used to perform word segmentation and syntactic analysis on the textbook chapters to identify core terms (such as "photosynthesis" and "Xinhai Revolution").

[0064] In detail, high-frequency test points (such as "the impact of the three technological revolutions") are statistically analyzed using keyword extraction algorithms (such as TF-IDF) to supplement implicit knowledge points not explicitly mentioned in the textbook.

[0065] Example: In the subject of history, textbook analysis may extract entities such as "May Fourth Movement" and "New Culture Movement", while test question analysis will find the high-frequency association of "the connection between the May Fourth Movement and the spread of Marxism".

[0066] In detail, an entity is the smallest unit of knowledge in a discipline, including concepts, events, people, etc.

[0067] In detail, standardization is the unification of entity names and aliases (such as merging "World War I" and "World War I").

[0068] In detail, an entity dictionary is constructed, and entity types are defined (such as "event" and "theory"); a rule engine or fuzzy matching algorithm is used for disambiguation (such as distinguishing between "Newton (physicist)" and "Newton (place name)").

[0069] In detail, this avoids the same entity being counted repeatedly due to differences in representation, thus improving the accuracy of subsequent logical relationship annotations.

[0070] In detail, logical relationships are the semantic connections between knowledge points, such as causality (A causes B), temporality (A precedes B), inclusion (A belongs to B), etc.

[0071] In detail, manual annotation can compensate for the shortcomings of automatic extraction and ensure logical rigor (such as avoiding the erroneous association of "apples caused Newton").

[0072] In detail, the annotation tool can be an icon annotation platform (such as Protege) that supports drag-and-drop relationship definition;

[0073] In detail, subject matter experts and knowledge engineers jointly develop labeling rules (such as "causal relationships must be supported by clear evidence").

[0074] Example: In a historical knowledge graph, experts mark "Xinhai Revolution → Overthrow of the Imperial System" as a causal relationship and provide historical evidence.

[0075] In detail, dynamic weights are assigned to logical relationships, such as adjusting the strength of the relationship based on the frequency of the questions.

[0076] In detail, based on the number of times historical grading data is statistically referenced, the weight update formula is designed as follows: Where N is the number of citations, high-frequency logical relationships account for a higher proportion in the score, which is closer to the actual exam focus.

[0077] In detail, knowledge graphs store entities and their relationships in a graph structure (nodes + edges), supporting efficient querying and visualization.

[0078] In detail, Neo4j graph database is selected, and node attributes (such as entity type and weight) are defined using Cypher language; Gephi or D3.js is used to generate interactive graphs (such as using node size to represent weight and edge color to distinguish relationship types).

[0079] Example: In the historical graph, the "May Fourth Movement" node is larger (higher weight), and its edge with the "New Culture Movement" is red (strong causal relationship).

[0080] Furthermore, the weights of entities and relationships are dynamically updated based on exam trends to avoid the lag of static graphs and adapt to changes in subject focus. For example, the weight of "environmental protection" has increased in recent years, reflecting the requirements of the new curriculum standards.

[0081] In detail, after each exam, the entity coverage in the candidates' answers is fed back to the graph; if an entity is frequently missed in this exam, its weight is automatically increased (e.g., weight +5%).

[0082] In detail, the textbook analysis provides basic entities, expert annotations ensure logical rigor, and dynamic construction continuously optimizes the graph. Taking history as an example, from extracting the entity of "Xinhai Revolution" to annotating its causal logic with "the end of the imperial system", the weight of the event is dynamically adjusted in the graph, closely following the scoring requirements throughout the process.

[0083] In detail, invalid associations (such as "Newton → Xinhai Revolution") are filtered out through expert-annotated causal relationships to avoid candidates piling up irrelevant entities for points; if the answer piles up low-frequency entities (such as "a certain battle"), the coverage score will be greatly reduced because they have low weight in the graph.

[0084] S2. Perform multi-granularity analysis on the answers to subjective questions based on semantic units to obtain the set of semantic paragraphs and the set of keywords for the answers to subjective questions.

[0085] In this embodiment of the invention, the step of performing multi-granularity parsing of subjective question answers based on semantic units to obtain a set of semantic paragraphs and a set of keywords for the subjective question answers includes:

[0086] Subjective question answers are divided into semantic paragraphs based on dependency syntax.

[0087] The BiLSTM-CRF model is used to identify each semantic paragraph in order to obtain the keyword set of the subjective question answer;

[0088] The keywords in the keyword set are semantically disambiguated using the matching strategy of the subject knowledge graph, and the hierarchical relationship between the semantic paragraphs is determined based on the semantically disambiguated keywords.

[0089] The set of semantic paragraphs for determining the answers to the subjective questions is determined based on the semantic paragraphs and the hierarchical relationships.

[0090] In detail, dependency parsing is a technique that divides long texts into independent semantic segments by identifying grammatical dependencies (such as subject-verb, verb-object) between words in a sentence.

[0091] In detail, subjective question answers often contain multiple layers of logic (such as argument → evidence → conclusion), and directly extracting sentences can easily destroy semantic integrity. By using dependency parsing, we can ensure that each paragraph expresses a single, complete meaning.

[0092] In detail, the dependency tree structure of the answer sentences is parsed using the StanfordParser or Spacy libraries; centered on the core predicate, its dependent components are merged (e.g., "photosynthesis (core) → requires light and chloroplasts" is grouped into a single paragraph) to avoid excessively long paragraphs that could lead to semantic confusion, thus providing clear input for subsequent keyword extraction.

[0093] Example: The historical answer "The May Fourth Movement promoted the spread of Marxism and laid the foundation for the New Democratic Revolution" was split into two paragraphs:

[0094] ①The May Fourth Movement promoted the spread of Marxism;

[0095] ② It laid the foundation for the New Democratic Revolution.

[0096] In detail, BiLSTM-CRF is a sequence labeling model that combines a bidirectional long short-term memory network (capturing context) with a conditional random field (constraining label transfer); traditional TF-IDF only counts word frequencies and cannot identify technical terms (such as "separation of powers") or long-tail entities (such as "Monroe Doctrine").

[0097] In detail, entities (such as historical events and chemical concepts) in subject textbooks and test answers are annotated to build an annotated corpus; input answer text and output entity labels (such as "B-event" and "I-theory").

[0098] Example: Input paragraph ①, model labels "May Fourth Movement (B-Event), Marxism (B-Theory)".

[0099] In detail, semantic disambiguation uses knowledge graph matching to determine the specific meaning of polysemous words in context (e.g., does "apple" refer to fruit or company); test takers may misuse terms (e.g., writing "cell wall" as "cell membrane"), or use words with multiple meanings (e.g., "evolution" has different meanings in different disciplines).

[0100] In detail, keywords are aligned with map entities (e.g., "evolution" is matched with "species evolution" in biology rather than "cultural evolution" in sociology); based on the disambiguated entities, the relationships between paragraphs are inferred (e.g., paragraph ① "May Fourth Movement" and paragraph ② "New Democratic Revolution" form a causal chain).

[0101] Example: If the answer contains "cellular respiration requires mitochondria", after disambiguation, it is confirmed that "mitochondria" is a biological entity, not "mitochondrial landforms" in geography.

[0102] In detail, the disambiguated paragraphs are integrated into a tree or graph structure according to hierarchical relationships (general-specific, causal, progressive); flat paragraph sets cannot reflect the logical depth of the answer (such as whether the argument is adequately supported); predefined tags (such as "Support" and "Contrast") are used to mark the relationships between paragraphs; and structured data in JSON format is generated, including paragraph content, keyword list, and relationship tags.

[0103] In detail, through a four-step process of dependency parsing, BiLSTM-CRF annotation, knowledge graph disambiguation, and hierarchical relationship integration, this solution achieves multi-dimensional analysis of subjective question answers. For example, in a history exam, if a candidate's answer only lists event names (such as "May Fourth Movement, New Culture Movement") without causal explanation, the system will determine its logical weakness based on the completeness of the logical chain in the semantic paragraph set (such as the absence of "→ promote" relationship), and reduce its final score by combining dynamic weights. This design accurately quantifies the logical quality of the answer, completely solving the pain point of "keyword stuffing but logical confusion" in grading.

[0104] S3. Extract entities and logical connectors corresponding to the subject knowledge graph from the semantic paragraph set, and construct a directed logic graph of the subjective question answer based on the semantic association between the entities and the logical connectors.

[0105] In this embodiment of the invention, extracting entities and logical connectors corresponding to the subject knowledge graph from the semantic paragraph set refers to using a BiLSTM-CRF model or a pre-trained language model (such as BERT) to perform sequence labeling on the semantic paragraphs and identify subject-related entities (such as historical events and scientific concepts). For example, for the paragraph "The May Fourth Movement promoted the spread of Marxism", the model labels the entities "May Fourth Movement (event)" and "Marxism (theory)".

[0106] In detail, model training needs to be based on a well-annotated subject corpus (such as textbooks and test question analysis) to ensure comprehensive coverage of entity types (events, people, theories); and combine rule templates (such as "therefore" and "however") and dependency parsing to identify logical connectors that express causal, adversative, and progressive relationships.

[0107] Example: In the sentence "The May Fourth Movement promoted the spread of Marxism, thus laying the foundation for the New Democratic Revolution", extract the logical connectives "promoted" and "therefore".

[0108] In detail, the grammatical validity of conjunctions (such as whether "therefore" connects two independent clauses) is verified through dependency syntax trees.

[0109] In this embodiment of the invention, constructing the directed logic graph of the subjective question answer based on the semantic associations between the entities and the logical connectors includes:

[0110] Construct an initial logical graph using the entities as nodes and the logical connectors as edges;

[0111] The syntactic associations between the entities are verified based on the dependency syntax.

[0112] The graph pruning algorithm is used to remove the edges that fail the verification in the initial logic graph to obtain a directed logic graph with syntax verification.

[0113] In detail, the initial logic graph is a directed graph structure in which entities (such as historical events and scientific concepts) are nodes and logical connectors (such as "therefore" and "however") are edges, representing semantic relationships between entities.

[0114] In detail, the logical structure of the answer can be intuitively expressed through graph structure. For example, "A leads to B" can be abstracted as node A → edge "leads to" → node B.

[0115] In detail, obtain the list of disambiguated entities (such as "May Fourth Movement" and "Marxism") from the semantic paragraph set of the previous step.

[0116] In detail, logical connectors (such as "therefore" indicating cause and effect, and "however" indicating contrast) are extracted from the answer text based on rules or pre-trained models (such as BERT).

[0117] Example: In the historical answer paragraph "The May Fourth Movement promoted the spread of Marxism, thus laying the foundation for the New Democratic Revolution", extract the nodes "May Fourth Movement", "Marxism", and "New Democratic Revolution", and connect the nodes with "promoted" and "therefore".

[0118] In detail, directed edges represent logical directionality (e.g., "A causes B" is irreversible); the construction rules include edge type mapping (e.g., "therefore" is mapped to causal relationship) and node uniqueness verification.

[0119] In detail, predefine the relationship type for each logical connector (e.g., "therefore → causality", "but → contrast"); merge duplicate entity nodes (e.g., "New Culture Movement" retains only one node in different paragraphs) to avoid redundant nodes interfering with logical chain analysis and ensure the simplicity and accuracy of the graph.

[0120] In detail, dependency parsing verifies whether the logic between entities conforms to the grammatical norms of the answering language by identifying the grammatical dependencies of sentence components (such as subject-verb, verb-object). Test takers may misuse logical connectors (such as "because the sun is round, so 1+1=2"), and invalid associations need to be eliminated through grammatical verification.

[0121] In detail, a dependency tree is generated for sentences with logical connections (e.g., "May Fourth Movement (subject) → promote (predicate) → spread (object)"); the grammatical roles corresponding to the edges are checked to see if they are reasonable (e.g., "promote" should connect the subject and object, rather than an adverbial phrase).

[0122] Example: If the answer states "Marxism was spread because of the May Fourth Movement", dependency analysis finds that "because" modifies an adverbial phrase rather than the causal subject, and therefore the side is deemed invalid.

[0123] Specifically, edges that pass grammatical validation are assigned higher weights (e.g., grammatically compliant edges receive a 20% weight increase), while those that fail are assigned lower weights or are removed. The grammatical confidence scoring formula is designed as follows: If S is below the threshold (e.g., 0.6), the edge is marked as "to be pruned".

[0124] In detail, graph pruning involves removing edges from the initial graph that do not conform to grammatical or semantic rules using rules or machine learning models, while preserving valid logical chains.

[0125] Furthermore, the semantic rationality probability of edges is predicted by a graph neural network (GNN). The inputs of the graph neural network (GNN) include, but are not limited to, node features, connector word types, and historical pruning data. The output of the graph neural network (GNN) is the edge validity probability.

[0126] In detail, the initial diagram may contain invalid connections arbitrarily piled up by the candidate (such as "Newton → Xinhai Revolution"), which need to be purified by pruning the logical structure.

[0127] In detail, the graph pruning algorithm is a pruning algorithm based on grammatical confidence scoring and semantic reasonableness verification.

[0128] Furthermore, the graph pruning algorithm ensures the validity of the logical graph through a double-check mechanism.

[0129] In detail, rule-based pruning refers to directly deleting edges that fail syntax validation (e.g., confidence S < 0.6); probabilistic pruning refers to using graph neural networks (GNNs) to predict the rationality of edges and deleting low-probability edges (e.g., probability < 0.5).

[0130] Example: In the initial diagram, "Marxism → (therefore) → Reform and Opening Up" is pruned due to a chronological inconsistency (Reform and Opening Up occurred later than the spread of Marxism).

[0131] In detail, a directed logic graph with syntax verification is an optimized graph that retains only syntactically and semantically valid edges after pruning; optimization includes denoising, weight adjustment, and logic chain closure checks; merging parallel edges (such as multiple "therefore" connecting the same node pair, only retaining the one with the highest weight); checking whether the logic chain is closed (such as "A→B→C" must form a complete causal chain to avoid broken chains).

[0132] Effect: The historical answer "May Fourth Movement → Promoted → Marxism → Therefore → New Democratic Revolution" forms a complete causal chain, while the broken parts (such as not explaining "how Marxism influenced the revolution") are marked as logical weaknesses.

[0133] In detail, the candidate's answer contained three entities: "May Fourth Movement, New Culture Movement, and Reform and Opening Up," and forcibly connected them with "therefore."

[0134] In detail, the initial graph generates three nodes and two edges (May Fourth Movement → Therefore → New Culture Movement; New Culture Movement → Therefore → Reform and Opening Up); grammatical verification reveals that "therefore" has no reasonable grammatical dependency in the second instance (the New Culture Movement and Reform and Opening Up have no direct causal relationship); the pruning algorithm deletes the second edge, and the final logic graph retains only the valid part (May Fourth Movement → New Culture Movement); the logical coherence score drops from the initial 0.6 (two edges) to 0.3 (one edge), accurately reflecting the quality of the answer.

[0135] In summary, this solution achieves accurate modeling of the logical structure of subjective question answers through a three-step process: constructing an initial logic graph, dependency syntax verification, and graph pruning optimization. Taking history as an example, if a candidate piles up phrases like "May Fourth Movement" and "Reform and Opening Up" but misuses "therefore" connections, the system will identify invalid edges through syntax verification, prune the graph, retain only valid logical chains, and calculate the score based on the weight and completeness of the remaining edges. This design not only solves the problem of "keyword stuffing with chaotic logic" but also improves the scientific rigor and adaptability of the scoring by combining dynamic weights with grammatical rules.

[0136] S4. Generate the relationship strength between the logical graph nodes of the directed logic graph through a graph attention network, and generate the logical coherence score of the subjective question answer based on the relationship strength.

[0137] In this embodiment of the invention, generating the relationship strength between logical graph nodes of the directed logical graph through a graph attention network includes:

[0138] The nodes and edges in the directed logic graph are input into the graph attention network as initial data;

[0139] The initial attention score between each pair of entity nodes is calculated using the multi-head attention layer in the graph attention network.

[0140] The initial attention score is nonlinearly activated and normalized using the LeakyReLU function to obtain a normalized attention score.

[0141] The normalized attention scores are weighted and aggregated to obtain the aggregated node features between the logical graph nodes in the directed logical graph.

[0142] Numerical mapping is performed on the features of the aggregated nodes to obtain the relationship strength between the nodes of the logical graph.

[0143] In this embodiment of the invention, generating the logical coherence score for the subjective question answer based on the relation strength includes:

[0144] Based on the relationship strength, the node features and edge weights in the directed logical graph are aggregated to obtain a global logical view;

[0145] The integrity score of the logical chain in the directed logical graph is generated based on the global logical view;

[0146] The completeness score is standardized to obtain the logical coherence score of the subjective question answer.

[0147] In detail, entity nodes (such as "May Fourth Movement" and "Marxism") and logical connecting word edges (such as "therefore" and "promote") in the directed logic graph are transformed into input features of the graph attention network (GAT). Traditional graph neural networks (GNN) have difficulty distinguishing the importance of different logical relationships, while GAT can dynamically learn the strength of the association between nodes through the attention mechanism.

[0148] In detail, entity types (events, theories) are converted into vectors (e.g., "May Fourth Movement" → [0.8, 0.2, ...]); logical connectors (e.g., "therefore") are mapped to initial edge weights (e.g., initial weight of causal edge = 0.7).

[0149] Example: In the historical answer logic graph, the node features of "May Fourth Movement" are 0.9 and 0.1, "Marxism" is 0.7 and 0.3, and the edge "promote" is encoded as 0.6.

[0150] In detail, multi-head attention computes multiple sets of attention weights in parallel to capture semantic associations in different dimensions. A single attention mechanism may ignore complex logic (such as the simultaneous existence of causal and parallel relationships). Each attention head independently learns the association strength of node pairs (e.g., head 1 focuses on temporal sequence, head 2 focuses on causality).

[0151] In detail, the initial attention score The calculation formula is:

[0152]

[0153] in, and These are nodes and eigenvectors, It is a learnable attention parameter vector. It is the transpose of the learnable attention parameter vector. It is a learnable weight matrix. It is a splicing symbol, spliced ​​along the feature dimension. and , It is a modified linear unit activation function with leakage.

[0154] Example: In the edge “May Fourth Movement → Promote → Marxism”, multi-head attention calculates the causal strength (0.8) and temporal strength (0.6) respectively.

[0155] In detail, It is an improved activation function that solves the "neuron death" problem of ReLU; normalization (Softmax) transforms the scores into a probability distribution; directly using the original attention scores may lead to vanishing gradients or numerical instability.

[0156] In detail, attention score application (Slope in the negative interval = 0.2), retain weak negative signals; for nodes All neighbors calculate For example, with an initial attention score of 1.2, then -0.5, after... It becomes 1.2, −0.1, and after normalization, it becomes 0.85, 0.15.

[0157] In detail, aggregation is the weighted sum of the features of neighboring nodes according to attention weights, and the current node representation is updated; the aggregated features incorporate contextual logical information, and more comprehensively reflect the role of the node in the global context.

[0158] In detail, after aggregating the features of the "May Fourth Movement" into the "Marxism" node, its representation is updated from 0.7, 0.3 to 0.8, 0.4, 0.6, ..., containing causal and temporal information.

[0159] In detail, the aggregated features are mapped to relation strength scalars (such as 0-1) through fully connected layers. The relation strength needs to be quantized to support subsequent scoring (such as strength > 0.5 is considered valid logic).

[0160] In detail, the mapping formula is as follows: ,in, It is the logical graph node and The strength of the relationship between them; It is an activation function. It is a learnable weight vector. It is a node eigenvectors; It is a learnable bias term.

[0161] In detail, the formula works by applying the nodes... eigenvectors After performing a weighted summation (dot product operation) and adding a bias term, followed by processing with a sigmoid activation function, a score representing a certain association strength or probability is obtained. .

[0162] In detail, the effective strength threshold is adjusted according to the question type (e.g., the threshold for essay questions = 0.6, and for short answer questions = 0.4). For example, the strength of the relationship between "May Fourth Movement → Marxism" is 0.82, while the strength of "Marxism → Reform and Opening Up" is only 0.35 due to the temporal contradiction.

[0163] In detail, the global logical view is a comprehensive representation that integrates node features, edge weights, and relationship strength. A single dimension (such as node coverage) cannot fully evaluate the logical quality of the answer.

[0164] In detail, max pooling or average pooling is used to compress the full image features (e.g., historical answer image → pooled into a 128-dimensional vector).

[0165] In detail, logical chain integrity refers to whether key nodes form closed paths (e.g., "A→B→C→A" forming a cycle indicates a logical contradiction); use depth-first search (DFS) to traverse the logical graph and count the length of valid paths (e.g., the longest causal chain contains 3 nodes).

[0166] In detail, the completeness score is equal to the effective path length divided by the theoretical maximum length, and then multiplied by the mean relationship strength. For example, in the historical answer, "May Fourth Movement → Marxism → New Democratic Revolution" is a complete causal chain (length 3), and the score is 0.8; if there are no intermediate nodes, the score is 0.3.

[0167] In detail, the integrity score is normalized to a 0-1 rating to facilitate integration with other indicators (such as entity coverage); different question types have different requirements for the length of the logical chain (e.g., essay questions require a long chain, while short answer questions can have a short chain), and standardization eliminates differences between question types.

[0168] In detail, the threshold is dynamically adjusted based on the overall performance of the candidates (for example, if the average chain length in this exam is relatively short, the threshold will be lowered by 10%); the original integrity score of 0.75 is standardized to 0.68, which reflects that the logic of the answer is better than average but not excellent.

[0169] In detail, from the original answer to the calculation of relationship strength, the entire process is automated through GAT, avoiding the subjectivity of manual rules; taking historical answers as an example, GAT identifies the high-strength relationship "May Fourth Movement → Marxism", which improves the integrity score and the final logical score is significantly higher than that of the piled-up answer.

[0170] In detail, if a candidate's answer is "The May Fourth Movement thus promoted reform and opening up", GAT will identify that the relationship strength of "therefore" is low (time sequence error), the integrity score will be zero, and the logic score will be 0; while traditional methods may misjudge due to the high coverage of keywords "May Fourth Movement" and "reform and opening up".

[0171] In summary, this scheme achieves dynamic quantification of logical relationship strength through the multi-head attention mechanism, nonlinear activation, feature aggregation, and numerical mapping of graph attention networks, and generates a logical coherence score by combining a global view and path analysis. Taking historical answers as an example, if a candidate constructs a complete causal chain of "May Fourth Movement → Marxism → New Democratic Revolution," the system will assign a high-strength relationship (0.8+) and a high completeness score (0.75), resulting in a final logical score of 0.72 (out of 1.0); while answers that are simply piled up have a low relationship strength and a broken logical chain, resulting in a score below 0.3.

[0172] S5. Combining the entity coverage score of the subject knowledge graph with the logical coherence score, calculate the final score of the subjective question answer based on the dynamic weighting coefficient.

[0173] In this embodiment of the invention, the step of combining the entity coverage score of the subject knowledge graph with the logical coherence score to calculate the final score of the subjective question answer based on the dynamic weight coefficient refers to dynamically adjusting the weight ratio of coverage and logic according to the question type and subject characteristics. Specifically, the logical reasoning weight is high for essay questions (e.g., logic:coverage = 7:3), and the coverage weight is high for short answer questions (e.g., logic:coverage = 3:7). The final score of the subjective question answer is calculated based on the determined weight ratio.

[0174] In detail, the dynamic weight coefficients in the dynamic weight coefficients for assigning entities in the subject knowledge graph are determined based on the question type and subject characteristics.

[0175] In this embodiment of the invention, the calculation steps for the entity coverage score are as follows:

[0176] Compare the entities corresponding to the subjective question answers with the entities corresponding to the standard answers;

[0177] Based on the comparison results, the Jaccard similarity between the entity corresponding to the subjective question answer and the entity of the standard answer is calculated, and the Jaccard similarity is used as the entity coverage score.

[0178] In detail, entity extraction is the process of identifying subject knowledge points (such as historical events or chemical concepts) from candidates' answers and standard answers, and standardizing them into unified terminology. Candidates may use aliases (such as "World War I" instead of "World War I") or incorrect terms (such as "cell membrane" being mistakenly written as "cell wall"), which requires standardization to ensure the accuracy of the comparison.

[0179] In detail, the BiLSTM-CRF model is used to extract entities (such as "May Fourth Movement" and "New Culture Movement"), and semantic disambiguation is performed through knowledge graph matching (such as "cellular respiration" being mapped to a standard biological entity); a set of standard entities is obtained from a pre-constructed subject knowledge graph (such as the standard answer to a history question containing "May Fourth Movement", "spread of Marxism", and "New Democratic Revolution").

[0180] Example: A candidate's answer mentions "the New Culture Movement promoted the liberation of thought," which, after standardization, is mapped to "the New Culture Movement (event)" and "the liberation of thought (concept)."

[0181] In detail, entity comparison calculates the overlap (intersection) and difference (union-intersection) between the entities in the candidate's answer and the entities in the standard answer; simply counting the number of entities cannot reflect the coverage of core test points (e.g., piling up 10 minor entities is not as good as covering 3 core entities).

[0182] In detail, define the candidate entity set A and the standard entity set B, calculate the intersection A∩B and the union A∪B; record the standard entities that are not covered (such as candidates omitting "Marxist dissemination") and redundant entities (such as candidates mistakenly adding "Westernization Movement").

[0183] In detail, if the standard answer is required to include {May Fourth Movement, Marxism, New Democratic Revolution}, and the candidate's answer includes {May Fourth Movement, New Culture Movement}, then the intersection is {May Fourth Movement}, and the union is {May Fourth Movement, New Culture Movement, Marxism, New Democratic Revolution}.

[0184] In detail, Jaccard similarity can simultaneously measure coverage accuracy (avoiding interference from redundant entities) and recall (encouraging coverage of more standard entities), among which, In this process, based on statistics of past exam questions and expert annotations, each entity in the knowledge graph is assigned a weight. .

[0185] In detail, if the weights of "May Fourth Movement" = 0.9, "Marxism" = 0.8, "New Democratic Revolution" = 0.7, and "New Culture Movement" = 0.5, then... It is significantly higher than the traditional Jaccard (0.25).

[0186] In detail, the semantic accuracy of the candidate's entities is verified through a knowledge graph, and incorrect terms are penalized with weights; if a candidate mistakenly writes "cell membrane" as "cell wall", even if the word matches, it must be downgraded or removed.

[0187] In detail, the similarity of entities is calculated using graph embedding models such as TransE. If the similarity is less than the threshold (e.g., 0.6), it is considered an error and the weight is set to zero. The weight of the error entity is not included in the numerator (intersection), but it is included in the denominator (union).

[0188] Example: If a candidate mistakenly writes "Marxism" as "Marxist theory", the semantic similarity is 0.4 (below the threshold). Therefore, the intersection weight is 0, and the union weight is 0.8 (standard entity weight) + 0.3 (default weight of incorrect term) = 1.1.

[0189] like Figure 2 The diagram shown is a functional module diagram of an integrated intelligent online examination and marking system provided in an embodiment of the present invention.

[0190] The online examination integrated intelligent grading system 100 described in this invention can be installed in an electronic device. Depending on the functions implemented, the online examination integrated intelligent grading system 100 may include a weight allocation module 101, a multi-granularity parsing module 102, a logic diagram construction module 103, a coherence evaluation module 104, and a final score generation module 105. The module described in this invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0191] In this embodiment, the functions of each module / unit are as follows:

[0192] The weight allocation module 101 is used to construct a subject knowledge graph and assign dynamic weight coefficients to entities in the subject knowledge graph.

[0193] The multi-granularity parsing module 102 is used to perform multi-granularity parsing on the subjective question answers based on semantic units, so as to obtain the semantic paragraph set and keyword set of the subjective question answers;

[0194] The logic graph construction module 103 is used to extract entities and logical connectors corresponding to the subject knowledge graph from the semantic paragraph set, and construct a directed logic graph of the subjective question answer based on the semantic association between the entities and the logical connectors.

[0195] The coherence evaluation module 104 is used to generate the relationship strength between the logic graph nodes of the directed logic graph through a graph attention network, and generate the logical coherence score of the subjective question answer based on the relationship strength.

[0196] The final score generation module 105 is used to combine the entity coverage score of the subject knowledge graph with the logical coherence score, and calculate the final score of the subjective question answer based on the dynamic weight coefficient.

[0197] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0198] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0199] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0200] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0201] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0202] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An integrated intelligent marking method for online examinations, characterized in that, The method includes: S1. Construct a subject knowledge graph and assign dynamic weight coefficients to entities in the subject knowledge graph; S2. Perform multi-granularity analysis on the answers to subjective questions based on semantic units to obtain the set of semantic paragraphs and the set of keywords for the answers to subjective questions; S3. Extract entities and logical connectors corresponding to the subject knowledge graph from the semantic paragraph set, and construct a directed logic graph of the subjective question answer based on the semantic association between the entities and the logical connectors; S4. Generate the relation strength between logical graph nodes of the directed logical graph using a graph attention network, and generate a logical coherence score for the subjective question answer based on the relation strength, including: aggregating node features and edge weights in the directed logical graph based on the relation strength to obtain a global logical view; generating a completeness score for the logical chains in the directed logical graph based on the global logical view; and standardizing the completeness score to obtain a logical coherence score for the subjective question answer. S5. Combining the entity coverage score of the subject knowledge graph with the logical coherence score, calculate the final score of the subjective question answer based on the dynamic weighting coefficient.

2. The integrated intelligent marking method for online examinations as described in claim 1, characterized in that, The construction of the subject knowledge graph includes: Based on textbooks and past exam questions, the subject knowledge system is analyzed to obtain a set of knowledge points. The logical relationships between the knowledge points are established through expert annotation. A subject knowledge graph is constructed based on the entity set and the logical relationships.

3. The integrated intelligent marking method for online examinations as described in claim 1, characterized in that, The method of performing multi-granularity parsing of subjective question answers based on semantic units yields a set of semantic paragraphs and a set of keywords for the subjective question answers, including: Subjective question answers are divided into semantic paragraphs based on dependency syntax. The BiLSTM-CRF model is used to identify each semantic paragraph in order to obtain the keyword set of the subjective question answer; The keywords in the keyword set are semantically disambiguated using the matching strategy of the subject knowledge graph, and the hierarchical relationship between the semantic paragraphs is determined based on the semantically disambiguated keywords. The set of semantic paragraphs for determining the answers to the subjective questions is determined based on the semantic paragraphs and the hierarchical relationships.

4. The integrated intelligent marking method for online examinations as described in claim 3, characterized in that, The construction of the directed logic graph for the subjective question answers based on the semantic relationships between the entities and the logical connectors includes: Construct an initial logical graph using the entities as nodes and the logical connectors as edges; The syntactic associations between the entities are verified based on the dependency syntax. The graph pruning algorithm is used to remove the edges that fail the verification in the initial logic graph to obtain a directed logic graph with syntax verification.

5. The integrated intelligent marking method for online examinations as described in claim 4, characterized in that, The generation of the relation strength between nodes in the directed logical graph using a graph attention network includes: The nodes and edges in the directed logic graph are input into the graph attention network as initial data; The initial attention score between each pair of entity nodes is calculated using the multi-head attention layer in the graph attention network. The initial attention score is nonlinearly activated and normalized using the LeakyReLU function to obtain a normalized attention score. The normalized attention scores are weighted and aggregated to obtain the aggregated node features between the logical graph nodes in the directed logical graph. Numerical mapping is performed on the features of the aggregated nodes to obtain the relationship strength between the nodes of the logical graph.

6. The integrated intelligent marking method for online examinations as described in claim 1, characterized in that, The steps for calculating the entity coverage score are as follows: Compare the entities corresponding to the subjective question answers with the entities corresponding to the standard answers; Based on the comparison results, the Jaccard similarity between the entity corresponding to the subjective question answer and the entity of the standard answer is calculated, and the Jaccard similarity is used as the entity coverage score.

7. The integrated intelligent marking method for online examinations as described in claim 4, characterized in that, The graph pruning algorithm is a pruning algorithm based on syntactic confidence scoring and semantic reasonableness verification.

8. The online examination integrated intelligent marking method as described in any one of claims 1 to 7, characterized in that, The allocation of the dynamic weighting coefficients is dynamically adjusted based on the question type and subject characteristics.

9. An integrated intelligent online examination grading system, characterized in that, The system includes: The weight allocation module is used to construct a subject knowledge graph and assign dynamic weight coefficients to entities in the subject knowledge graph. The multi-granularity parsing module is used to perform multi-granularity parsing on the answers to subjective questions based on semantic units, so as to obtain the set of semantic paragraphs and the set of keywords of the answers to subjective questions; The logic graph construction module is used to extract entities and logical connectors corresponding to the subject knowledge graph from the semantic paragraph set, and construct a directed logic graph of the subjective question answer based on the semantic association between the entities and the logical connectors. A coherence evaluation module is used to generate the relation strength between nodes of the directed logic graph through a graph attention network, and to generate a logical coherence score for the subjective question answer based on the relation strength. This includes: aggregating node features and edge weights in the directed logic graph based on the relation strength to obtain a global logical view; generating a completeness score for the logical chains in the directed logic graph based on the global logical view; and standardizing the completeness score to obtain the logical coherence score for the subjective question answer. The final score generation module is used to combine the entity coverage score of the subject knowledge graph with the logical coherence score, and calculate the final score of the subjective question answer based on the dynamic weight coefficient.

Images