Online examination integrated intelligent paper marking method and system

Abstract

The invention relates to the technical field of intelligent education, and discloses an online examination integrated intelligent paper marking method and system, and the method comprises the steps: constructing a subject knowledge graph, and distributing dynamic weight coefficients for entities in the subject knowledge graph; analyzing a semantic paragraph set and a keyword set in the subjective question answers; entities and logic conjunctions corresponding to the subject knowledge graph are extracted from the semantic paragraph set, and a directed logic graph is constructed based on semantic association between the entities and the logic conjunctions; a logical coherence score between logical diagram nodes of the directed logical diagram is generated through the graph attention network. According to the invention, the accuracy of online examination integrated intelligent paper marking can be improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent education technology, and in particular to an integrated intelligent scoring method and system for online examinations. Background Art

[0002] The intelligence level of online exam marking systems directly affects the fairness and scientific nature of exam evaluation. Existing marking technologies have significant technical shortcomings in the subjective question assessment process:

[0003] First, in terms of semantic understanding, answers to subjective questions often rely on keyword matching or simple semantic analysis, lacking in-depth analysis of the integrity of the logical chain (such as causal and temporal relationships). This can lead to inflated scores for answers that are "keyword-heavy but logically confusing."

[0004] Secondly, in terms of knowledge representation, traditional methods overly rely on word frequency statistics (such as TF-IDF) or literal string matching, making it difficult to handle semantic disambiguation of subject-specific terminology. More importantly, existing systems generally use static weighting strategies, which are unable to dynamically adjust the importance of test points based on actual exam conditions. Summary of the Invention

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

[0006] To achieve the above objectives, the present invention provides an integrated intelligent scoring 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 the subjective questions based on semantic units to obtain a set of semantic paragraphs and a set of keywords for the answers to the subjective questions;

[0009] S3. Extracting entities and logical connectives corresponding to the subject knowledge graph from the semantic paragraph set, and constructing a directed logical graph of the answers to the subjective questions based on the semantic associations between the entities and the logical connectives;

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

[0011] S5. Combine the entity coverage score of the subject knowledge graph and the logical coherence score, and calculate the final score of the answer to the subjective question based on the dynamic weight coefficient.

[0012] Optionally, constructing a subject knowledge graph includes:

[0013] Analyze the subject knowledge system based on textbooks and past examination papers to obtain an entity set of knowledge points;

[0014] Establishing logical relationships between the knowledge points through expert annotation;

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

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

[0017] Divide the answers to subjective questions into semantic paragraphs based on dependency syntax;

[0018] Using a BiLSTM-CRF model to identify each of the semantic paragraphs to obtain a keyword set for the answer to the subjective question;

[0019] Performing semantic disambiguation on the keywords in the keyword set using the matching strategy of the subject knowledge graph, and determining the hierarchical relationship between the semantic paragraphs based on the semantically disambiguated keywords;

[0020] A semantic paragraph set of the answer to the subjective question is determined based on the semantic paragraphs and the hierarchical relationship.

[0021] Optionally, constructing a directed logical graph of the answers to the subjective questions based on the semantic associations between the entities and the logical connectives includes:

[0022] Constructing an initial logical graph by using the entities as nodes and the logical connectives as edges;

[0023] Verifying the grammatical association between the entities based on the dependency syntax;

[0024] The edges that fail to pass the verification in the initial logic graph are deleted through a graph pruning algorithm to obtain a directed logic graph with syntax verification.

[0025] Optionally, generating the relationship strength between the logic graph nodes of the directed logic graph by using the graph attention network includes:

[0026] Input the nodes and edges in the directed logical graph into the graph attention network as initial data;

[0027] Calculating an initial attention score between each pair of entity nodes using a 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] Performing weight distribution aggregation on the normalized attention scores to obtain aggregated node features between logic graph nodes in the directed logic graph;

[0030] Numerical mapping is performed on the aggregate node features to obtain the relationship strength between the logic graph nodes.

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

[0032] Aggregating node features and edge weights in the directed logical graph based on the relationship strength to obtain a global logical view;

[0033] generating a completeness score of a logic chain in the directed logic graph according to the global logic view;

[0034] The completeness score is normalized to obtain a logical coherence score for the answer to the subjective question.

[0035] Optionally, the entity coverage score is calculated as follows:

[0036] Comparing the entity corresponding to the answer to the subjective question with the standard answer entity;

[0037] The Jaccard similarity between the entity corresponding to the answer to the subjective question and the standard answer entity is calculated based on the comparison result, 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 rationality verification.

[0039] Optionally, the distribution of the dynamic weight coefficient is dynamically adjusted according to the question type and subject characteristics.

[0040] In order to solve the above problems, the present invention also provides an online examination integrated intelligent scoring system, characterized in that the system includes:

[0041] A weight assignment module is used to construct a subject knowledge graph and assign dynamic weight coefficients to entities in the subject knowledge graph;

[0042] A multi-granularity analysis module is used to perform multi-granularity analysis on the answers to the subjective questions based on semantic units to obtain a set of semantic paragraphs and a set of keywords for the answers to the subjective questions;

[0043] A logic graph construction module is used to extract entities and logical connectives corresponding to the subject knowledge graph from the semantic paragraph set, and construct a directed logic graph of the answers to the subjective questions based on the semantic associations between the entities and the logical connectives;

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

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

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

[0047] 1. The multi-level logical relationship extraction model constructed based on the graph neural network in this paper can automatically identify logical relationships such as causality, transitions, and progression in the answer text, and quantify logical coherence through the attention mechanism. Secondly, the dynamically updated subject knowledge graph is used to achieve fine-grained differentiation of professional terms;

[0048] 2. This invention constructs a subject knowledge graph and assigns dynamic weight coefficients to entities. It achieves adaptive adjustment of the importance of knowledge points through textbook analysis and statistics of previous exam questions. This solves the problem that traditional static weights cannot reflect changes in exam focus and improves the pertinence of entity coverage scoring.

[0049] 3. Dependency parsing and the BiLSTM-CRF model are used to achieve multi-level semantic segmentation of answers to subjective questions. Keyword disambiguation is performed in combination with subject knowledge graphs. A directed logic graph is constructed based on entities and logical connectives. Invalid associations are filtered through a graph pruning algorithm. This solves the problem of traditional keyword matching ignoring logical relationships and enhances the ability to parse the logical structure of answers. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A flow chart of an integrated intelligent scoring method for online examinations provided in one embodiment of the present invention;

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

[0052] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0053] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

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

[0055] Reference Figure 1 FIG. 1 is a flow chart of an integrated intelligent scoring method for online examinations according to an embodiment of the present invention. In this embodiment, the integrated intelligent scoring method for online examinations includes:

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

[0057] In an embodiment of the present invention, constructing a subject knowledge graph includes:

[0058] Analyze the subject knowledge system based on textbooks and past examination papers to obtain an entity set of knowledge points;

[0059] Establishing logical relationships between the knowledge points through expert annotation;

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

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

[0062] In detail, textbooks are the authoritative carriers of subject knowledge, and past examination papers reflect the distribution and key points of examination points. The combination of the two can ensure the comprehensiveness and practicality of the entity collection.

[0063] In detail, natural language processing (NLP) technology is used to perform word segmentation and syntactic analysis on textbook chapters to identify core terms (such as "photosynthesis" and "Xinhai Revolution").

[0064] In detail, we use keyword extraction algorithms (such as TF-IDF) to count high-frequency test points (such as "the impact of the three scientific and technological revolutions") to supplement implicit knowledge points that are not explicitly mentioned in the textbooks.

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

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

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

[0068] Specifically, build an entity dictionary and define entity types (such as "event" and "theory"); use a rule engine or fuzzy matching algorithm to resolve ambiguity (such as distinguishing between "Newton (physicist)" and "Newton (place name)").

[0069] Specifically, it avoids the same entity from being counted repeatedly due to differences in expression, and improves the accuracy of subsequent logical relationship annotation.

[0070] In detail, logical relations are semantic associations between knowledge points, such as causality (A causes B), temporal order (A precedes B), inclusion (A belongs to B), etc.

[0071] In detail, manual annotation can make up for the shortcomings of automatic extraction and ensure logical rigor (such as avoiding the wrong association of "apple leads to Newton").

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

[0073] In detail, subject matter experts and knowledge engineers jointly develop annotation rules (e.g., “causal relationships must be supported by clear evidence”).

[0074] Example: In the historical knowledge graph, experts marked "Xinhai Revolution → Overthrow of the Monarchy" as a causal relationship and attached historical evidence.

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

[0076] Specifically, based on the number of times the historical scoring data statistical relationship is cited, 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 test focus.

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

[0078] Specifically, the Neo4j graph database is used, and node properties (such as entity type and weight) are defined in Cypher. Gephi or D3.js are 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 large (high weight), and its edge with the New Culture Movement is red (strong causal relationship).

[0080] Furthermore, the entity and relationship weights are dynamically updated according to the examination 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 to reflect the requirements of the new curriculum standards.

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

[0082] In detail, textbook analysis provides basic entities, expert annotation ensures rigorous logic, and dynamic construction enables continuous optimization of the map; taking the history subject as an example, from extracting the "Xinhai Revolution" entity, to annotating its causal logic with the "end of the imperial system", and finally dynamically adjusting its weight in the map, the entire process closely adheres to the scoring requirements.

[0083] In detail, invalid associations (such as "Newton → Xinhai Revolution") are screened out through the causal relationships marked by experts to prevent candidates from piling up irrelevant entities to gain points; if the answer piles up low-frequency entities (such as "a minor battle"), the coverage score will be greatly reduced due to its low weight in the graph.

[0084] S2. Perform multi-granularity analysis on the answers to the subjective questions based on semantic units to obtain a semantic paragraph set and a keyword set of the answers to the subjective questions.

[0085] In an embodiment of the present invention, the multi-granularity analysis of the answers to the subjective questions based on semantic units to obtain the semantic paragraph set and keyword set of the answers to the subjective questions includes:

[0086] Divide the answers to subjective questions into semantic paragraphs based on dependency syntax;

[0087] Using a BiLSTM-CRF model to identify each of the semantic paragraphs to obtain a keyword set for the answer to the subjective question;

[0088] Performing semantic disambiguation on the keywords in the keyword set using the matching strategy of the subject knowledge graph, and determining the hierarchical relationship between the semantic paragraphs based on the semantically disambiguated keywords;

[0089] A semantic paragraph set of the answer to the subjective question is determined based on the semantic paragraphs and the hierarchical relationship.

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

[0091] Specifically, answers to subjective questions often contain multiple layers of logic (e.g., argument → evidence → conclusion). Directly truncating sentences can easily undermine semantic integrity. Dependency-based syntax segmentation ensures that each paragraph expresses a single, complete semantic meaning.

[0092] Specifically, use StanfordParser or Spacy to parse the dependency tree structure of the answer sentence; center on the core predicate and merge its dependent components (e.g., "Photosynthesis (core) → requires light and chloroplasts" into one paragraph) to avoid long paragraphs that lead to semantic confusion and provide clear input for subsequent keyword extraction.

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

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

[0095] ②Lay the foundation for the new democratic revolution.

[0096] Specifically, BiLSTM-CRF is a sequence labeling model that combines a bidirectional long short-term memory network (to capture context) with a conditional random field (to constrain label transfer). Traditional TF-IDF only counts word frequency and cannot identify specialized terms (such as "separation of powers") or long-tail entities (such as "Monroe Doctrine").

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

[0098] Example: Input paragraph ①, and the model labels "May Fourth Movement (B-event), Marxism (B-theory)".

[0099] In detail, semantic disambiguation is done by matching knowledge graphs to determine the specific meaning of polysemous words in context (e.g., does "apple" refer to fruit or a company); candidates may misuse terms (e.g., "cell wall" is written as "cell membrane"), or a word may have multiple meanings (e.g., "evolution" has different meanings in different disciplines).

[0100] Specifically, keywords are aligned with graph entities (e.g., “evolution” matches “species evolution” in biology but not “cultural evolution” in sociology); based on the disambiguated entities, the relationship between paragraphs is inferred (e.g., paragraph ① “May Fourth Movement” and paragraph ② “New Democratic Revolution” constitute a causal chain).

[0101] For example, if the answer appears to be “Cellular respiration requires mitochondria”, after disambiguation, it is confirmed that “mitochondria” is a biological entity, not the “mitochondrial landform” in geography.

[0102] Specifically, the disambiguated paragraphs are integrated into a tree or graph structure according to hierarchical relationships (general-specific, cause-and-effect, and progressive). A flat paragraph collection cannot reflect the logical depth of the answer (such as whether the argument is fully supported). Predefined tags (such as "Support" and "Contrast") are used to mark the relationship between paragraphs. Structured data in JSON format is generated, including paragraph content, keyword lists, and relationship tags.

[0103] Specifically, this solution achieves multi-dimensional analysis of answers to subjective questions through a four-step process: dependency syntax segmentation, BiLSTM-CRF annotation, knowledge graph disambiguation, and hierarchical relationship integration. For example, in a history exam, if a candidate's answer merely lists event names (e.g., "May Fourth Movement," "New Culture Movement") without causal explanation, the system will identify the answer as weak based on the completeness of the logical chain within the semantic paragraph set (e.g., the lack of a "→ promote" relationship) and, using dynamic weighting, reduce the candidate's final score. This design precisely quantifies the logical quality of answers, completely resolving the pain point of grading answers characterized by keyword-heavy answers but confusing logic.

[0104] S3. Extract entities and logical connectives corresponding to the subject knowledge graph from the semantic paragraph set, and construct a directed logic graph of the answers to the subjective questions based on the semantic associations between the entities and the logical connectives.

[0105] In an embodiment of the present invention, extracting entities and logical connectives 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 sequence label the semantic paragraphs and identify subject-related entities (such as historical events, 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 well-annotated subject corpora (such as textbooks and test question analysis) to ensure comprehensive coverage of entity types (events, people, theories); combined with rule templates (such as "therefore" and "however") and dependency syntax analysis, it can identify logical conjunctions that represent causal, transitional, and progressive relationships.

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

[0108] In detail, the grammatical rationality of the conjunctions (e.g., whether “therefore” connects two independent clauses) is verified through the dependency syntax tree.

[0109] In an embodiment of the present invention, constructing a directed logical graph of the answers to the subjective questions based on the semantic associations between the entities and the logical connectives includes:

[0110] Constructing an initial logical graph by using the entities as nodes and the logical connectives as edges;

[0111] Verifying the grammatical association between the entities based on the dependency syntax;

[0112] The edges that fail to pass the verification in the initial logic graph are deleted through a graph pruning algorithm to obtain a directed logic graph with syntax verification.

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

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

[0115] Specifically, a list of disambiguated entities (e.g., “May Fourth Movement”, “Marxism”) is obtained from the semantic paragraph set in the previous step.

[0116] Specifically, logical connectives (such as "so" for causality and "however" for transition) 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 the edges "promote" and "therefore".

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

[0119] Specifically, a relationship type is predefined for each logical connective (such as "therefore → cause and effect", "but → turning point"); repeated entity nodes are merged (for example, only one node is retained for "New Culture Movement" in different paragraphs) to avoid redundant nodes interfering with the logical chain analysis and ensure the simplicity and accuracy of the graph.

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

[0121] Specifically, a dependency tree is generated for sentences containing logical connections (e.g., “May Fourth Movement (subject) → promote (predicate) → spread (object)”); and the grammatical roles corresponding to the edges are checked to see if they are reasonable (e.g., “promote” needs to connect the subject and object, not the adverbial component).

[0122] Example: If the answer is "Marxism was spread because of the May Fourth Movement", dependency analysis finds that "because" modifies the adverbial clause rather than the causal subject, so the edge is invalid.

[0123] Specifically, higher weights are assigned to edges that pass grammatical verification (e.g., grammatically compliant edges are weighted +20%), whereas those that do not are weighted lower or deleted. The grammatical confidence scoring formula is designed as: , if S is lower than a threshold (such as 0.6), mark the edge as "to be pruned".

[0124] In detail, graph pruning is to remove edges in the initial graph that do not conform to grammatical or semantic rules through rules or machine learning models, while retaining valid logical chains.

[0125] Furthermore, the semantic rationality probability of the edge is predicted by a graph neural network (GNN), the input of the graph neural network (GNN) includes but is not limited to: node features, connective type and historical pruning data, and the output of the graph neural network (GNN) is the edge validity probability.

[0126] In detail, the initial graph may contain invalid associations randomly piled up by the candidates (such as "Newton→Xinhai Revolution"), and the logical structure needs to be purified through pruning.

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

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

[0129] In detail, rule pruning refers to directly deleting edges that fail grammatical verification (such as confidence S < 0.6); probability pruning refers to using graph neural networks (GNN) to predict the rationality of edges and delete low-probability edges (such as probability < 0.5).

[0130] Example: In the initial graph, “Marxism → (therefore) → Reform and Opening Up” was pruned due to a time sequence contradiction (Reform and Opening Up came later than the spread of Marxism).

[0131] Specifically, a directed logic graph with syntax checking is an optimized graph that retains only syntactically and semantically valid edges after pruning; the optimization includes denoising, weight adjustment, and logic chain closure checking; merging parallel edges (e.g., if multiple "therefore"s connect the same node pair, only the one with the highest weight is retained); and checking whether the logic chain is closed (e.g., "A→B→C" must form a complete causal chain to avoid chain breaks).

[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 included three entities: "May Fourth Movement, New Culture Movement, and Reform and Opening Up", and forcibly connected them with "therefore".

[0134] Specifically, the initial graph generated three nodes and two edges (May Fourth Movement → Therefore → New Culture Movement; New Culture Movement → Therefore → Reform and Opening Up); grammatical verification found that "therefore" had no reasonable grammatical dependency in the second place (there was no direct causal relationship between the New Culture Movement and Reform and Opening Up); the pruning algorithm deleted the second edge, and the final logic graph retained only the valid part (May Fourth Movement → New Culture Movement); the logical coherence score dropped from the initial 0.6 (two edges) to 0.3 (one edge), accurately reflecting the quality of the answer.

[0135] In summary, through a three-step process involving constructing an initial logical graph, verifying dependency syntax, and optimizing graph pruning, this solution achieves precise modeling of the logical structure of answers to subjective questions. For example, in the history section, if a candidate piles up "May Fourth Movement" and "Reform and Opening Up" but misuses the "therefore" connection, the system will identify invalid edges through syntax validation, prune only valid logical chains, and calculate a score based on the weight and integrity of the remaining edges. This design not only solves the problem of "keyword stacking with confusing logic" but also improves the scientific and adaptable nature of scoring by combining dynamic weighting with grammatical rules.

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

[0137] In an embodiment of the present invention, generating the relationship strength between the logic graph nodes of the directed logic graph by the graph attention network includes:

[0138] Input the nodes and edges in the directed logical graph into the graph attention network as initial data;

[0139] Calculating an initial attention score between each pair of entity nodes using a 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] Performing weight distribution aggregation on the normalized attention scores to obtain aggregated node features between logic graph nodes in the directed logic graph;

[0142] Numerical mapping is performed on the aggregate node features to obtain the relationship strength between the logic graph nodes.

[0143] In an embodiment of the present invention, generating a logical coherence score of the answer to the subjective question based on the relationship strength includes:

[0144] Aggregating node features and edge weights in the directed logical graph based on the relationship strength to obtain a global logical view;

[0145] generating a completeness score of a logic chain in the directed logic graph according to the global logic view;

[0146] The completeness score is normalized to obtain a logical coherence score for the answer to the subjective question.

[0147] In detail, the entity nodes (such as "May Fourth Movement" and "Marxism") and logical connective edges (such as "therefore" and "promote") in the directed logic graph are converted into input features of the graph attention network (GAT); traditional graph neural networks (GNNs) find it difficult to distinguish the importance of different logical relationships, while GAT can dynamically learn the strength of associations between nodes through the attention mechanism.

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

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

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

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

[0152]

[0153] in, and Node and The eigenvector of is a learnable attention parameter vector, is the transpose of the learnable attention parameter vector, is a learnable weight matrix, Is the splicing symbol, splicing along the feature dimension and , is the leaky rectified linear unit activation function,

[0154] Example: In the edge “May Fourth Movement → Promotion → Marxism”, multi-head attention calculates 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) converts the score into a probability distribution; directly using the original attention score may cause gradient disappearance or numerical instability.

[0156] In detail, the attention score is applied (negative interval slope = 0.2), retain the weak negative signal; for the node All neighbors calculate ,For example, the initial attention score is 1.2,−0.5, becomes 1.2,−0.1, and after normalization it becomes 0.85,0.15.

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

[0158] In detail, after the “Marxism” node aggregates the “May Fourth Movement” features, its representation is updated from 0.7, 0.3 to 0.8, 0.4, 0.6, ..., which contains causal and temporal information.

[0159] Specifically, the aggregated features are mapped into a relationship strength scalar (e.g., 0-1) through a fully connected layer. The relationship strength needs to be quantified to support subsequent scoring (e.g., strength > 0.5 is considered valid logic).

[0160] In detail, the mapping formula is: ,in, is the logic graph node and the strength of the relationship between is the activation function, is a learnable weight vector, is a node The eigenvector of is a learnable bias term.

[0161] In detail, the formula is obtained by The eigenvector of Perform weighted summation (dot product operation) and add bias term, and then process it through Sigmoid activation function to get a score representing the strength or probability of a certain association .

[0162] Specifically, 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 relationship strength of “May Fourth Movement → Marxism” is 0.82, while the strength of “Marxism → Reform and Opening Up” is only 0.35 due to the time sequence 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] Specifically, use max pooling or average pooling to compress full-graph features (e.g., historical answer graph → pooled into a 128-dimensional vector).

[0165] Specifically, the integrity of the logic chain refers to whether the key nodes form a closed path (for example, "A→B→C→A" forms a loop, indicating a logical contradiction); use depth-first search (DFS) to traverse the logic graph and count the length of the valid path (for example, the longest causal chain contains 3 nodes).

[0166] Specifically, the completeness score is equal to the effective path length divided by the theoretical maximum length, multiplied by the mean relationship strength. For example, in the history answer, "May Fourth Movement → Marxism → New Democratic Revolution" is a complete causal chain (length 3), with a score of 0.8; if the intermediate node is missing, the score is 0.3.

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

[0168] Specifically, the threshold is dynamically adjusted based on the overall performance of the examinee (e.g., if the average chain length in this exam is shorter, the threshold is lowered by 10%); the original completeness score of 0.75 is standardized to 0.68, reflecting that the logic of the answer is better than the average but not excellent.

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

[0170] In detail, if the candidate's answer is "The May Fourth Movement therefore promoted reform and opening up", GAT will identify that the relationship strength of "therefore" is low (time sequence error), the completeness score will be zero, and the logical score = 0; while the traditional method may make a misjudgment due to the high coverage of the keywords "May Fourth Movement" and "Reform and Opening Up".

[0171] In summary, this solution dynamically quantifies the strength of logical relationships through the multi-head attention mechanism, nonlinear activation, feature aggregation, and numerical mapping of the graph attention network. It also generates a logical coherence score by combining a global view with path analysis. For example, if a candidate constructs a complete causal chain from the May Fourth Movement to Marxism to the 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 a maximum of 1.0). However, a stacked answer, due to its low relationship strength and broken logical chain, will receive a score below 0.3.

[0172] S5. Combine the entity coverage score of the subject knowledge graph and the logical coherence score, and calculate the final score of the answer to the subjective question based on the dynamic weight coefficient.

[0173] In an embodiment of the present invention, the combination of the entity coverage score of the subject knowledge graph and the logical coherence score, and the calculation of the final score of the answer to the subjective question 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, wherein the logic weight of the essay question is high (such as logic: coverage = 7:3), and the coverage weight of the short answer question is high (such as logic: coverage = 3:7), and the final score of the answer to the subjective question is calculated according to the determined weight ratio.

[0174] In detail, the dynamic weight coefficients in the dynamic weight coefficients assigned to the entities in the subject knowledge graph are determined according to the question type and subject characteristics.

[0175] In an embodiment of the present invention, the steps for calculating the entity coverage score are as follows:

[0176] Comparing the entity corresponding to the answer to the subjective question with the standard answer entity;

[0177] The Jaccard similarity between the entity corresponding to the answer to the subjective question and the standard answer entity is calculated based on the comparison result, 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, chemical concepts) from the 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" mistakenly written as "cell wall"), and standardization is required 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" is mapped to a standard biological entity); standard entity sets are obtained from pre-constructed subject knowledge graphs (such as the standard answers to history questions include "May Fourth Movement", "Spread of Marxism", and "New Democratic Revolution").

[0180] Example: The candidate's answer mentioned that "the New Culture Movement promoted ideological emancipation", which was mapped into "New Culture Movement (event)" and "ideological emancipation (concept)" after standardization.

[0181] In detail, entity comparison is to calculate the overlap (intersection) and difference (union-intersection) between the candidate's answer entity and the standard answer entity; only counting the number of entities cannot reflect the coverage of core test points (for example, stacking 10 secondary entities is not as good as covering 3 core entities).

[0182] Specifically, define the candidate entity set A and the standard entity set B, calculate the intersection A∩B and the union A∪B; record uncovered standard entities (such as the candidate omitting "the spread of Marxism") and redundant entities (such as the candidate mistakenly adding "the Westernization Movement").

[0183] In detail, if the standard answer needs 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 the coverage accuracy (avoiding redundant entity interference) and recall (encouraging coverage of more standard entities), where , among which, based on the statistics of past exam questions and expert annotations, each entity in the knowledge graph is given a weight .

[0185] Specifically, if the weight of the May Fourth Movement is 0.9, the weight of Marxism is 0.8, the weight of the New Democratic Revolution is 0.7, and the weight of the New Culture Movement is 0.5, then , which is significantly higher than the traditional Jaccard (0.25).

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

[0187] Specifically, a graph embedding model such as TransE is used to calculate entity similarity. If the similarity is less than a threshold (such as 0.6), it is considered an error and its weight is reset to zero. The weight of the incorrect entity is not included in the numerator (intersection), but is included in the denominator (union).

[0188] Example: The candidate mistakenly wrote "Marxism" as "Marx Theory", and the semantic similarity = 0.4 (below the threshold), then the intersection weight = 0, and the union weight = 0.8 (standard entity weight) + 0.3 (default weight of the incorrect term) = 1.1.

[0189] like Figure 2 , which is a functional module diagram of an online examination integrated intelligent scoring system provided by one embodiment of the present invention.

[0190] The integrated online exam intelligent grading system 100 described in the present invention can be installed in an electronic device. Depending on the functionality implemented, the integrated online exam intelligent grading system 100 can include a weight assignment module 101, a multi-granularity analysis module 102, a logic diagram construction module 103, a coherence evaluation module 104, and a final score generation module 105. A module, also referred to as a unit, is a series of computer program segments that can be executed by an electronic device processor and perform a fixed function, and is stored in the electronic device's memory.

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

[0192] The weight assignment 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 analysis module 102 is used to perform multi-granularity analysis on the answers to the subjective questions based on semantic units to obtain a set of semantic paragraphs and a set of keywords of the answers to the subjective questions;

[0194] The logic graph construction module 103 is used to extract entities and logical connectives corresponding to the subject knowledge graph from the semantic paragraph set, and construct a directed logic graph of the answers to the subjective questions based on the semantic associations between the entities and the logical connectives;

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

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

[0197] In the several embodiments provided by the present 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 example, the module division is merely a logical function 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, and 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 elements. Some or all of the modules may be selected to achieve the purpose of the solution of this embodiment according to actual needs.

[0199] In addition, the functional modules in various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or 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 the present application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to achieve 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 limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An integrated intelligent scoring method for online examinations, characterized in that: The method comprises: 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 the subjective questions based on semantic units to obtain a set of semantic paragraphs and a set of keywords for the answers to the subjective questions; S3. Extracting entities and logical connectives corresponding to the subject knowledge graph from the semantic paragraph set, and constructing a directed logical graph of the answers to the subjective questions based on the semantic associations between the entities and the logical connectives; S4. Generating the relationship strength between the logic graph nodes of the directed logic graph through the graph attention network, and generating a logical coherence score for the answer to the subjective question based on the relationship strength; S5. Combine the entity coverage score of the subject knowledge graph and the logical coherence score, and calculate the final score of the answer to the subjective question based on the dynamic weight coefficient.

2. The online examination integrated intelligent scoring method according to claim 1, characterized in that: The construction of the subject knowledge graph includes: Analyze the subject knowledge system based on textbooks and past examination papers to obtain an entity set of knowledge points; Establishing logical relationships between the knowledge points through expert annotation; A subject knowledge graph is constructed based on the entity set and the logical relationship.

3. The online examination integrated intelligent scoring method according to claim 1, characterized in that: The multi-granularity analysis of the answers to the subjective questions based on semantic units to obtain the semantic paragraph set and keyword set of the answers to the subjective questions includes: Divide the answers to subjective questions into semantic paragraphs based on dependency syntax; Using a BiLSTM-CRF model to identify each of the semantic paragraphs to obtain a keyword set for the answer to the subjective question; Performing semantic disambiguation on the keywords in the keyword set using the matching strategy of the subject knowledge graph, and determining the hierarchical relationship between the semantic paragraphs based on the semantically disambiguated keywords; A semantic paragraph set of the answer to the subjective question is determined based on the semantic paragraphs and the hierarchical relationship.

4. The online examination integrated intelligent scoring method according to claim 3, characterized in that: The step of constructing a directed logic graph of the answers to the subjective questions based on the semantic associations between the entities and the logical connectives includes: Constructing an initial logical graph by using the entities as nodes and the logical connectives as edges; Verifying the grammatical association between the entities based on the dependency syntax; The edges that fail to pass the verification in the initial logic graph are deleted through a graph pruning algorithm to obtain a directed logic graph with syntax verification.

5. The online examination integrated intelligent scoring method according to claim 4, characterized in that: The generating of the relationship strength between the logic graph nodes of the directed logic graph by the graph attention network includes: Input the nodes and edges in the directed logical graph into the graph attention network as initial data; Calculating an initial attention score between each pair of entity nodes using a 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; Performing weight distribution aggregation on the normalized attention scores to obtain aggregated node features between logic graph nodes in the directed logic graph; Numerical mapping is performed on the aggregate node features to obtain the relationship strength between the logic graph nodes.

6. The online examination integrated intelligent scoring method according to claim 1, characterized in that: Generating a logical coherence score of the answer to the subjective question based on the relationship strength includes: Aggregating node features and edge weights in the directed logical graph based on the relationship strength to obtain a global logical view; generating a completeness score of a logic chain in the directed logic graph according to the global logic view; The completeness score is normalized to obtain a logical coherence score for the answer to the subjective question.

7. The online examination integrated intelligent scoring method according to claim 1, characterized in that: The calculation steps of the entity coverage score are as follows: Comparing the entity corresponding to the answer to the subjective question with the standard answer entity; The Jaccard similarity between the entity corresponding to the answer to the subjective question and the standard answer entity is calculated based on the comparison result, and the Jaccard similarity is used as the entity coverage score.

8. The online examination integrated intelligent scoring method according to claim 4, characterized in that: The graph pruning algorithm is a pruning algorithm based on grammatical confidence scoring and semantic rationality verification.

9. The online examination integrated intelligent scoring method according to any one of claims 1 to 8, characterized in that: The distribution of the dynamic weight coefficient is dynamically adjusted according to the question type and subject characteristics.

10. An integrated intelligent scoring system for online examinations, characterized in that: The system comprises: A weight assignment module is used to construct a subject knowledge graph and assign dynamic weight coefficients to entities in the subject knowledge graph; A multi-granularity analysis module is used to perform multi-granularity analysis on the answers to the subjective questions based on semantic units to obtain a set of semantic paragraphs and a set of keywords for the answers to the subjective questions; A logic graph construction module is used to extract entities and logical connectives corresponding to the subject knowledge graph from the semantic paragraph set, and construct a directed logic graph of the answers to the subjective questions based on the semantic associations between the entities and the logical connectives; a coherence evaluation module, configured to generate the relationship strength between the logic graph nodes of the directed logic graph through a graph attention network, and generate a logical coherence score for the answer to the subjective question based on the relationship strength; The final score generation module is used to combine the entity coverage score of the subject knowledge graph and the logical coherence score to calculate the final score of the answer to the subjective question based on the dynamic weight coefficient.

Images