Education knowledge graph quality evaluation method based on confidence evaluation

Abstract

The invention discloses an education knowledge graph quality evaluation method based on confidence evaluation, and relates to the technical field of natural language processing, and the method comprises the steps: S1, obtaining multi-source education data, and achieving the data standardization through a subject term unified dictionary and an education field exclusive relationship type mapping table; s2, constructing a dimension weight adaptive model, and dynamically adjusting weights of accuracy, integrity and consistency in combination with subject types, education stages and application scenes; s3, based on the standardized data and the real-time weight, respectively evaluating the confidence of each dimension through double mechanisms, three sub-dimensions and double detection; s4, fusing to obtain a comprehensive confidence coefficient, and if the comprehensive confidence coefficient does not reach the standard, adjusting parameters according to priorities and returning to re-evaluation; and S5, updating a model coefficient according to teacher feedback and student learning effect data to realize iteration. According to the method, the evaluation suitability and accuracy are improved, and reliable quality guarantee is provided for educational knowledge graph supporting teaching and personalized learning.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a method for evaluating the quality of educational knowledge graphs based on confidence assessment. Background Technology

[0002] With the popularization of online education and personalized learning, educational knowledge graphs need to integrate fragmented resources and build a systematic knowledge network. Educational data comes from diverse sources and varies in quality. Traditional assessment methods, such as manual annotation, are insufficient to handle large-scale dynamic data, necessitating automated confidence assessment technologies.

[0003] Educational knowledge and student needs are constantly evolving, requiring real-time updates to the knowledge graph and adjustments to its confidence level. For example, dynamic knowledge graphs in financial risk control scenarios support millisecond-level incremental updates, a technology that can be transferred to the education field.

[0004] Educational knowledge contains a great deal of implicit logic, such as subject prerequisite relationships and cognitive development patterns. Traditional quality assessment methods, such as coverage and accuracy, cannot fully reflect its complexity. For example, the RKQG model divides subject knowledge into five levels, emphasizing the integration of explicit and implicit knowledge.

[0005] Therefore, there is an urgent need for a quality evaluation method for educational knowledge graphs based on confidence assessment. In the face of the problem of insufficient synergy in multi-dimensional assessment, including the mutual constraints among quality dimensions such as accuracy, completeness, and consistency, existing methods lack systematic integration. Summary of the Invention

[0006] This invention aims to provide a quality evaluation method for educational knowledge graphs based on confidence assessment. By standardizing multi-source educational data, adaptively adjusting dimensional weights, conducting multi-dimensional exclusive confidence assessments, implementing closed-loop optimization and feedback iteration, it addresses the problems of insufficient multi-dimensional collaboration, fixed weights detached from educational scenarios, and disconnect between assessment and teaching practice in existing educational knowledge graph quality evaluation methods.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] Methods for evaluating the quality of educational knowledge graphs based on confidence assessment include:

[0009] S1. Acquire structured, semi-structured, and unstructured educational data; normalize knowledge point entities by constructing a unified dictionary of subject terms; standardize relationships by establishing a mapping table of relationship types specific to the education field; and obtain standardized graph data.

[0010] S2. Construct a dimension weight adaptive model, determine the core dimensions of quality evaluation as accuracy, completeness and consistency, extract subject type, education stage and application scenario as domain feature factors, construct a weight calculation function based on the factors, and dynamically adjust the real-time weights of each dimension.

[0011] S3. Based on standardized graph data and real-time weights, multi-dimensional confidence assessment is conducted. Among them, accuracy is assessed by a dual mechanism of semantic verification of the BERT model in the education field and feedback correction of student answer records. Integrity is assessed by weighted evaluation from three sub-dimensions: entity coverage, relationship coverage, and Bloom's cognitive hierarchy coverage. Consistency is assessed by dual detection of subject logical conflict detection and data internal conflict detection.

[0012] S4. Combine the confidence scores of multiple dimensions according to the real-time weights to obtain the comprehensive confidence score. If it is lower than the preset threshold, determine the optimization priority according to the product of the dimension deviation and the weight, adjust the corresponding evaluation parameters, and return to step S3.

[0013] S5. Collect teacher feedback scores and student learning effect data as correction factors, update the feature factor influence coefficients of the dimension weight adaptive model, and achieve iterative optimization.

[0014] Preferably, S1 specifically includes:

[0015] The method for constructing the unified dictionary of subject terminology is as follows: collect knowledge point terms from the national curriculum standards, authoritative textbooks and core academic literature of the corresponding subject, use the edit distance algorithm to calculate the character similarity of synonyms, combine the pre-trained language model in the education field to calculate the semantic similarity of terms, weight and fuse the character similarity and semantic similarity, merge terms with similarity ≥ 0.85, establish a one-to-one mapping relationship between non-standard terms and standard terms, and form a unified dictionary of subject terminology.

[0016] Preferably, S1 further includes:

[0017] For the preprocessing of unstructured educational data, for student answer record texts, a word segmentation tool specifically for the education field is used to divide knowledge point related segments, and a named entity recognition model is used to extract knowledge point entities from the answer records; for teacher teaching feedback texts, sentiment polarity analysis is used to screen effective feedback, and the knowledge point relationship correction information involved in the feedback is extracted, and then the relationship standardization is completed by combining it with an education-specific relationship type mapping table.

[0018] Preferably, S2 specifically includes:

[0019] The quantification method for the domain characteristic factors is as follows: in the subject type, natural sciences are assigned a value of 1.0 and humanities and social sciences are assigned a value of 0.8; in the education stage, basic education is assigned a value of 0.9 and higher education is assigned a value of 1.1; in the application scenario, knowledge point retrieval scenario is assigned a value of 0.9 and personalized learning recommendation scenario is assigned a value of 1.2; the quantified values ​​are calibrated using the Delphi method of domain experts, and the calibration cycle is once per semester.

[0020] Preferably, S3 further includes:

[0021] In the dual-mechanism evaluation of accuracy, the training corpus of the BERT model in the education field includes the curriculum standards, textbook chapter content, and analysis of high-difficulty test questions for the corresponding subject. During fine-tuning, a triple (s,p,o) semantic matching task is used, with the input being the concatenated text "s+p+o". The output layer uses the sigmoid function to generate an initial accuracy confidence score in the 0-1 interval. The feedback correction of student answer records only includes answer data with a recent sample size ≥50. When the accuracy is lower than 60% of the preset threshold, the correction coefficient is forcibly set to 0.6, and when it is higher than 140% of the threshold, it is forcibly set to 1.4.

[0022] Preferably, S3 further includes:

[0023] In the weighted evaluation of integrity across its three sub-dimensions, redundant entities with duplicate annotations are removed when calculating entity coverage. Relationship coverage prioritizes the coverage ratio of core relationships, with core relationships accounting for no less than 60% of the total relationship types. The weights of the three sub-dimensions of Bloom's cognitive hierarchy coverage are dynamically split according to the allocation ratio of the integrity dimension in the real-time weights, with entity coverage accounting for 40%, relationship coverage accounting for 35%, and cognitive hierarchy coverage accounting for 25%.

[0024] Preferably, S3 further includes:

[0025] In the dual-detection evaluation of consistency, the subject-specific logical conflict detection constructs a subject-specific knowledge logical rule base and uses a graph traversal algorithm to scan the paths between entities. For each conflicting path found, the logical conflict coefficient increases by 0.1 (maximum 1). In the internal data conflict detection, the attribute conflict rate is equal to the number of entities with attribute value differences divided by the total number of entities. When the difference value is within the allowable error range, it is not counted as a conflict.

[0026] Preferably, S4 specifically includes:

[0027] When fusing multi-dimensional confidence scores with real-time weights to obtain a comprehensive confidence score, the confidence score evaluation results of each dimension are first uniformly converted to the same evaluation scale.

[0028] When the overall confidence level is lower than the preset threshold, the difference between the confidence level of each dimension and the preset quality threshold of the corresponding dimension is first calculated as the dimension deviation. Then, the optimization priority of each dimension is determined by multiplying the dimension deviation with the real-time weight of that dimension. The larger the product, the higher the optimization priority of the dimension.

[0029] For dimensions with higher priority, adjust their corresponding evaluation parameters. Specifically, adjust the semantic verification parameters for the accuracy dimension, adjust the sub-dimension weight allocation parameters for the integrity dimension, and adjust the conflict detection parameters for the consistency dimension. After the parameter adjustment is completed, re-execute the multi-dimensional confidence evaluation in step S3.

[0030] Preferably, S5 specifically includes:

[0031] When collecting teacher feedback scores, the feedback content is limited to specific evaluations by teachers on the accuracy of normalization of knowledge point entities in the knowledge graph, the rationality of standardization of relationships between entities, and the relevance of multi-dimensional confidence assessment results to actual teaching.

[0032] The collected teacher feedback ratings are validated for validity. Subjective evaluations that are not directly related to the quality of the knowledge graph are removed to form a valid feedback dataset. This dataset is used as the core component of the correction factor to adjust the influence coefficients of domain feature factors that are directly related to the teaching scenario in the dimensional weight adaptive model.

[0033] Preferably, S5 further includes:

[0034] The collected student learning outcome data includes data on changes in the mastery of knowledge points after learning activities based on the current knowledge graph, feedback data on the adaptability of learning paths, and data on the improvement of knowledge application ability.

[0035] The student learning outcome data was filtered for correlation, retaining only the data directly related to the knowledge graph quality evaluation dimension. The filtered data was then integrated with teacher feedback scores to form a correction factor.

[0036] When updating the feature factor influence coefficients of the adaptive model based on the correction factor, a phased update mechanism is adopted. First, local coefficients are adjusted to address the prominent issues reflected in the student learning performance data, and then the overall coefficients are calibrated in conjunction with teacher feedback scores.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] S1 constructs a unified dictionary of subject-specific terminology and a mapping table of relationship types specific to the education field to normalize the knowledge point entities and standardize the relationships of multi-source heterogeneous educational data (structured, semi-structured, and unstructured). This avoids evaluation bias caused by messy data formats and inconsistent terminology, and ensures the accuracy of the data source for subsequent multi-dimensional confidence assessments.

[0039] The S2-built dimensional weight adaptive model takes accuracy, completeness, and consistency as its core dimensions and dynamically adjusts the weights by combining key characteristic factors in the education field such as subject type, education stage, and application scenario, rather than using general fixed weights. This allows the assessment focus to meet the needs of different education scenarios, such as basic education focusing on knowledge completeness and higher education focusing on logical consistency, thus solving the pain point of insufficient coordination in multi-dimensional assessment.

[0040] S3 designs an education-specific assessment mechanism for each dimension: Accuracy combines the BERT model in the education field (ensuring semantic rationality) with student answer feedback (linking to actual learning outcomes), avoiding pure model assessment from being divorced from teaching practice; Completeness covers entities, relationships, and Bloom's hierarchy of knowledge (matching educational cognitive patterns), rather than just assessing the amount of data; Consistency takes into account subject logic (such as prerequisite relationships) and internal data conflicts, ensuring that the knowledge graph conforms to the educational knowledge logic system.

[0041] S4 uses a comprehensive confidence threshold judgment and priority optimization mechanism to achieve precise parameter adjustment when the assessment fails to meet the standards, avoiding blind optimization; S5 updates the model coefficients by combining teacher teaching feedback and student learning effect data, so that the assessment model can continuously adapt to changes in the educational scenario, ensuring the long-term effectiveness of the assessment, and ultimately enhancing the supporting value of educational knowledge graphs for personalized learning and teaching assistance. Attached Figure Description

[0042] Figure 1 This is a flowchart of a confidence-based evaluation method for the quality assessment of educational knowledge graphs. Detailed Implementation

[0043] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0044] Reference Figure 1 As shown, the quality assessment method for educational knowledge graphs based on confidence evaluation includes:

[0045] S1. Acquire structured, semi-structured, and unstructured educational data; normalize knowledge point entities by constructing a unified dictionary of subject terms; standardize relationships by establishing a mapping table of relationship types specific to the education field; and obtain standardized graph data.

[0046] S2. Construct a dimension weight adaptive model, determine the core dimensions of quality evaluation as accuracy, completeness and consistency, extract subject type, education stage and application scenario as domain feature factors, construct a weight calculation function based on the factors, and dynamically adjust the real-time weights of each dimension.

[0047] S3. Based on standardized graph data and real-time weights, multi-dimensional confidence assessment is conducted. Among them, accuracy is assessed by a dual mechanism of semantic verification of the BERT model in the education field and feedback correction of student answer records. Integrity is assessed by weighted evaluation from three sub-dimensions: entity coverage, relationship coverage, and Bloom's cognitive hierarchy coverage. Consistency is assessed by dual detection of subject logical conflict detection and data internal conflict detection.

[0048] S4. Combine the confidence scores of multiple dimensions according to the real-time weights to obtain the comprehensive confidence score. If it is lower than the preset threshold, determine the optimization priority according to the product of the dimension deviation and the weight, adjust the corresponding evaluation parameters, and return to step S3.

[0049] S5. Collect teacher feedback scores and student learning effect data as correction factors, update the feature factor influence coefficients of the dimension weight adaptive model, and achieve iterative optimization.

[0050] It should be noted that the structured data consists of curriculum standards / test question banks, which achieve semantic alignment through ontology mapping; the semi-structured data consists of teaching plans / learning materials, which use pattern-aware parsing technology to extract cognitive relationships; and the unstructured data consists of classroom videos / discussion texts, which utilize multimodal BERT embedding spatial projection.

[0051] A subject-specific terminology dictionary and a relation type mapping table form a dual-track normalization engine to solve the "semantic fragmentation" problem caused by the heterogeneity of educational data.

[0052] S1 specifically includes:

[0053] The method for constructing the unified dictionary of subject terminology is as follows: collect knowledge point terms from the national curriculum standards, authoritative textbooks and core academic literature of the corresponding subject, use the edit distance algorithm to calculate the character similarity of synonyms, combine the pre-trained language model in the education field to calculate the semantic similarity of terms, weight and fuse the character similarity and semantic similarity, merge terms with similarity ≥ 0.85, establish a one-to-one mapping relationship between non-standard terms and standard terms, and form a unified dictionary of subject terminology.

[0054] For the preprocessing of unstructured educational data, student answer records are segmented into knowledge point-related segments using a specialized education-specific word segmentation tool, and knowledge point entities are extracted from the answer records using a named entity recognition model. For teacher feedback texts, effective feedback is screened using sentiment polarity analysis, and information on correcting knowledge point relationships within the feedback is extracted. Finally, relationship standardization is achieved using an education-specific relationship type mapping table. (Supplementary explanation: No flowcharts or tables are required.)

[0055] It should be noted that the construction of a unified dictionary for subject-specific terminology is as follows:

[0056] National standards anchor the main framework, using the Ministry of Education's curriculum standards as the core terminology to ensure policy compliance;

[0057] The textbook and literature are dynamically expanded, integrating three types of textbooks (People's Education Press, Beijing Normal University Press, and Ministry-compiled edition) and five core journal terminology databases, covering 98% of teaching scenarios;

[0058] The cross-version alignment engine resolves the mapping of cross-textbook terms such as "redox reaction" (Chemistry Compulsory 1) and "electron transfer reaction" (Elective 4) through semantic vector space projection;

[0059] Mixed similarity calculation:

[0060]

[0061] In the formula, Sim fusion λ represents the fusion similarity, which is the result calculated by combining character similarity and semantic similarity; λ represents the dynamic weight, which is used to balance the contributions of character similarity and semantic similarity. In basic disciplines, λ is set to 0.4; in terminology-sensitive fields, λ is set to 0.6.

[0062] EditDist represents the edit distance, which is the minimum number of editing operations (including insertion, deletion, and replacement) required to transform one string into another; MaxLen represents the length of the longer string; CosEmbed represents the cosine embedding, which is the cosine similarity between two vectors; BERT edu (t1) and BERT edu (t2) represents the vector representation obtained by encoding two texts t1 and t2 using the educational pre-trained model BERT_edu.

[0063] Character similarity was optimized, the edit distance algorithm was improved, and subject-specific morpheme weights were introduced, such as "function" having a 30% higher weight than "number".

[0064] Semantic similarity enhancement employs the educational pre-trained model BERT. edu Accurately capture differences in teaching context

[0065] A dynamic weighting strategy is used, setting λ = 0.4 (basic disciplines) or 0.6 (terminology-sensitive areas) to balance glyphic and semantic contributions.

[0066] Cognitive hierarchy terminology is isolated by constructing a two-dimensional matrix of "terminology-cognitive level" to avoid cross-level merging, such as not merging "force" in junior high school with "Newton's Third Law" in senior high school; Bloom's Taxonomy is integrated into terminology normalization to solve the essential teaching problem of "knowing ≠ understanding";

[0067] A dialect terminology conversion channel will be established, creating a regional textbook terminology conversion table, such as "chlorine" (national) → "green gas" (Hong Kong and Macao), supporting special scenarios such as the Greater Bay Area.

[0068] Unstructured educational data preprocessing:

[0069] Student answer records, knowledge point related segmentation, education-specific word segmentation, word segmentation model based on attention mechanism; for example, segmenting "prove that the two base angles of an isosceles triangle are equal" into [isosceles triangle|property proof|base angles are equal]; cognitive behavior labeling, automatically labeling problem-solving strategy types, such as "reverse deduction" and "analogy transfer";

[0070] Dual-channel verification for entity extraction:

[0071]

[0072] Teacher feedback on text refinement, emotion-content separation framework:

[0073] Original feedback → Sentiment analysis → Negative, extracting equipment / management issues; Neutral / Positive, knowledge point relationship correction extraction → mapping relationship type library;

[0074] The noise filtering mechanism removes emotional expressions irrelevant to the teaching content, such as "the projector is too dark"; the information positioning is corrected by using a pointer network to accurately extract descriptive segments, such as "Newton's laws should be explained first".

[0075] S2 specifically includes:

[0076] The quantification method for the domain characteristic factors is as follows: in the subject type, natural sciences are assigned a value of 1.0 and humanities and social sciences are assigned a value of 0.8; in the education stage, basic education is assigned a value of 0.9 and higher education is assigned a value of 1.1; in the application scenario, knowledge point retrieval scenario is assigned a value of 0.9 and personalized learning recommendation scenario is assigned a value of 1.2; the quantified values ​​are calibrated using the Delphi method of domain experts, and the calibration cycle is once per semester.

[0077] It should be noted that for the Natural Sciences category (1.0): the knowledge structure is characterized by a highly structured conceptual system, such as the force-motion-energy logic chain in physics; the assessment sensitivity is such that a single conceptual error can lead to systemic cognitive biases, such as a balancing error in a chemical equation; the assignment principle is to strengthen the accuracy weight (βacc↑15%) to ensure scientific rigor.

[0078] Humanities and Social Sciences (0.8): Knowledge flexibility allows for multiple interpretations, such as multi-perspective analysis of historical events; teaching inclusiveness prioritizes the integrity of relationships over absolute precision, such as the influence network of literary schools; the assignment principle is to reduce consistency constraints (αcon↓20%) and increase the priority of integrity.

[0079] The constructivist logic of assigning value to educational stages:

[0080]

[0081] The essence of assigning value differences is that higher education needs to support critical thinking (which requires more rigorous consistency verification) and innovative research (which relies on a more complete knowledge boundary);

[0082] Application scenario assignment: Knowledge point retrieval (0.9), the core requirement is fast and accurate positioning (accuracy weight ≥70%), partial incompleteness is allowed, such as the temporary lack of less popular knowledge points;

[0083] Personalized learning recommendations (1.2) have the core requirement of comprehensively covering learners' cognitive status (completeness weight ↑35%) and updating knowledge connections in real time.

[0084] Delphi calibration:

[0085] The expert group is designed to create an ecological framework, consisting of subject matter experts, provincial-level curriculum researchers, front-line teachers, representatives of top-level teachers, educational technology experts, AI engineers, cognitive scientists, and learning analysts. The Delphi expert group comprises subject matter experts, provincial-level curriculum researchers, front-line teachers, representatives of top-level teachers, educational technology experts, AI engineers, and cognitive scientists.

[0086] The system is based on a four-dimensional authoritative balance, covering the essence of the discipline (40%), teaching practice (30%), technological implementation (20%), and cognitive laws (10%); a dynamic elimination mechanism is in place, where experts whose prediction deviation is >15% for two consecutive rounds are suspended from participation.

[0087] Calibration process: In the pre-calibration stage, machine learning is used to pre-tune parameters and generate initial suggested values ​​based on 100,000+ teaching behavior logs; outlier filtering removes proposals that deviate from the historical mean by more than 2 standard deviations.

[0088] Three rounds of iterative optimization:

[0089] Round Core task Education data support First round Propose benchmark value National curriculum quality standards Second round In-depth debate on controversial points Regional academic quality monitoring report Third round Converge to consensus Cross-school atlas application effect evaluation

[0090] The consensus quantification index uses Kendall's concordance coefficient (W≥0.85) to determine the convergence of opinions: W=12Σd 2 / (m 2 (n 3 -n)), where m = number of experts, n = parameter terms.

[0091] Semester calibration cycle:

[0092] The educational cycle is aligned with the schedule, with at least one major teaching phase transition (midterm / final) per semester;

[0093] Typical adjustment case: During the college entrance examination preparation period → the weight of natural sciences was increased to 1.05;

[0094] New semester curriculum standards update → Consistency threshold for humanities and social sciences lowered by 0.03;

[0095] An emergency response mechanism will be activated to trigger temporary calibration in the following situations: the release of national curriculum reform policies (activated within 72 hours); a sharp drop in the pass rate of a certain subject area exceeding 10% (response within 48 hours);

[0096] Assignment Iteration Formula:

[0097] In the formula, V new V represents the new assignment or the updated value; old This represents the old assignment, the initial value; η represents the educational gain coefficient, a weighting factor used to adjust the impact of the difference between the actual teaching effectiveness indicators and the expected effects recommended by the graph on the new assignment, with a default value of 0.15; Eff app Indicators representing actual teaching effectiveness, such as the increase in knowledge point mastery rate; Eff pred This indicates the expected effect of the map recommendation, that is, the teaching effect predicted based on a certain model or algorithm.

[0098] S3 further includes:

[0099] In the dual-mechanism evaluation of accuracy, the training corpus of the BERT model in the education field includes the curriculum standards, textbook chapter content, and analysis of high-difficulty test questions for the corresponding subject. During fine-tuning, a triple (s,p,o) semantic matching task is used, with the input being the concatenated text "s+p+o". The output layer uses the sigmoid function to generate an initial accuracy confidence score in the 0-1 interval. The feedback correction of student answer records only includes answer data with a recent sample size ≥50. When the accuracy is lower than 60% of the preset threshold, the correction coefficient is forcibly set to 0.6, and when it is higher than 140% of the threshold, it is forcibly set to 1.4.

[0100] In the weighted evaluation of integrity across its three sub-dimensions, redundant entities with duplicate annotations are removed when calculating entity coverage. Relationship coverage prioritizes the coverage ratio of core relationships, with core relationships accounting for no less than 60% of the total relationship types. The weights of the three sub-dimensions of Bloom's cognitive hierarchy coverage are dynamically split according to the allocation ratio of the integrity dimension in the real-time weights, with entity coverage accounting for 40%, relationship coverage accounting for 35%, and cognitive hierarchy coverage accounting for 25%.

[0101] In the dual-detection evaluation of consistency, the subject-specific logical conflict detection constructs a subject-specific knowledge logical rule base and uses a graph traversal algorithm to scan the paths between entities. For each conflicting path found, the logical conflict coefficient increases by 0.1 (maximum 1). In the internal data conflict detection, the attribute conflict rate is equal to the number of entities with attribute value differences divided by the total number of entities. When the difference value is within the allowable error range, it is not counted as a conflict.

[0102] It should be noted that the accuracy is assessed using a dual-mechanism approach:

[0103] BERT model in the education field: Corpus selection is based on: curriculum standards to ensure that knowledge expression conforms to national education goals (covering 100% of core competency items); textbook chapters to extract progressive expressions that conform to students' cognitive ladder (such as physics concepts from "phenomenon description" to "mathematical modeling"); and high-difficulty test question analysis to internalize expert problem-solving thinking patterns (such as strategic knowledge encoding of challenging math problems).

[0104] By incorporating Bloom's cognitive categorization (memory → creation) into text encoding, the model can understand the semantic difference between "describing Newton's laws" and "applying Newton's laws".

[0105] Fine-tuning task: Triple semantic matching design, input text construction example:

[0106] text = "Trigonometric functions [s] have the property [p] periodicity [o]";

[0107] label = 0.93, confidence level (expert assessment);

[0108] Sigmoid output layer: Set a threshold of 0.85 as the passing grade for teaching; if the value is below this, manual review is triggered. Introduce a cognitive decay factor to automatically reduce the weight of triples that "exceed the cognitive level of the grade level" by 20%.

[0109] Answer feedback and correction, dynamic sample selection strategy:

[0110]

[0111] In the linear correction formula, C is the corrected weighting coefficient; P act The actual accuracy rate, i.e., the percentage of questions students actually answered correctly; P th Threshold accuracy is a set standard accuracy rate used to determine whether a student needs extra attention or resources.

[0112] Three-dimensional cognitive structure mapping for integrity assessment:

[0113] Coverage calculation: Entity deduplication mechanism, redundancy judgment criteria:

[0114] Synonyms (such as "Pythagorean theorem" and "Pythagorean theorem") are merged; parent-child relationships are preserved for cross-granularity entities (such as "function" and "quadratic function"); and the transitive reduction algorithm based on the educational ontology hierarchy tree reduces redundancy by 72%.

[0115] The principle of prioritizing core relationships: relationships mandated by the curriculum standards, such as "force → motion" in physics; relationships that support ≥3 higher-order cognitive objectives, such as "event A → impact → event B" in history;

[0116] Dynamic monitoring automatically triggers the knowledge extraction enhancement module when the core relationship ratio is less than 60%.

[0117] Bloom's hierarchy of quantification and weight allocation cognitive model:

[0118]

[0119] In the formula, W entity The weight for entity coverage refers to the degree of coverage of knowledge points or concepts during the educational process; W relation The weight of relational coverage refers to the degree of understanding of the relationships between knowledge points during the educational process; W bloom The weight of the cognitive hierarchy refers to the degree of coverage of Bloom's cognitive hierarchy (such as memory, comprehension, application, analysis, evaluation, and creation) in the educational process; W completeness (EduFeature) is the integrity weight of the educational feature. It is a function used to calculate the integrity of the educational feature, which may include a comprehensive consideration of the above-mentioned entity coverage, relation coverage and cognitive level.

[0120] Entity coverage (40%) ensures the basic knowledge framework (corresponding to memory / comprehension levels);

[0121] Relationship coverage (35%), constructing a cognitive network (corresponding to application / analysis levels);

[0122] Cognitive level (25%): Supports innovative thinking (corresponding to evaluation / creativity level);

[0123] Hierarchical coverage algorithm: Constructed based on the gold standard, it decomposes cognitive objectives based on the curriculum standards, such as the mathematical "derivative" which needs to cover 6 levels of objectives;

[0124] Gap detection: If the current level coverage is less than the standard requirement, the gap value = (standard value - current value) × cognitive weight factor;

[0125] Consistency assessment:

[0126] A rule engine for detecting subject-specific logical conflicts, and the educational characteristics of its logical rule base:

[0127] Rule type Example Conflict coefficient Disciplinary adaptation Cause and effect inversion Must learn equations before inequalities +0.15 Mathematical cognitive progression Contradictory conclusions Biological "evolution" vs. creationism +0.30 Education of the nature of science Interdisciplinary conflict Geographical plate movement → historical civilization migration +0.10 STEM integrated curriculum

[0128] The graph traversal algorithm was optimized, and the teaching path was scanned. An improved DFS algorithm was adopted (depth limit = cognitive depth of the learning stage): conflict path weight = base value 0.1 × path teaching importance coefficient;

[0129] Fault-tolerant design for data conflict detection, and tolerance mechanism for attribute differences;

[0130] Error tolerance range table:

[0131] Attribute type Error range Education scene example Numerical value ±5％ Difference in measured value of physical constant Time sequence ± 3 years Controversy over the time of historical events Qualitative description Synonymous expression "Cell division" vs. "Cell proliferation"

[0132] Educational calibration for conflict rate calculation:

[0133]

[0134] In the formula, Conflict attr Attribute conflict rate, representing the proportion of instances where a conflict occurs on a particular attribute; N error N represents the number of erroneous samples, i.e., the number of samples that are incorrect on a certain attribute; total Total sample size, i.e., all samples considered; Tolerance, i.e., the maximum allowable deviation; Actual diff This refers to the actual difference, i.e., the difference value actually observed.

[0135] When the actual difference is within the tolerance range (Actual diff ≤Tolerance), Conflict attr =0, when the difference exceeds the limit, the conflict value is amplified according to the degree of deviation index.

[0136] S4 specifically includes:

[0137] When fusing multi-dimensional confidence scores with real-time weights to obtain a comprehensive confidence score, the confidence score evaluation results of each dimension are first uniformly converted to the same evaluation scale.

[0138] When the overall confidence level is lower than the preset threshold, the difference between the confidence level of each dimension and the preset quality threshold of the corresponding dimension is first calculated as the dimension deviation. Then, the optimization priority of each dimension is determined by multiplying the dimension deviation with the real-time weight of that dimension. The larger the product, the higher the optimization priority of the dimension.

[0139] For dimensions with higher priority, adjust their corresponding evaluation parameters. Specifically, adjust the semantic verification parameters for the accuracy dimension, adjust the sub-dimension weight allocation parameters for the integrity dimension, and adjust the conflict detection parameters for the consistency dimension. After the parameter adjustment is completed, re-execute the multi-dimensional confidence evaluation in step S3.

[0140] It should be noted that the educational scale for confidence level fusion is consistent:

[0141] Multidimensional dimensional normalization, educational evaluation space mapping, and the use of piecewise linear transformation functions to transform confidence scores of different dimensions to the [0,1] interval:

[0142]

[0143] In the formula, C norm The value is the normalized value, i.e., the value after processing by the piecewise linear transformation function, and its range is within the interval [0,1]; C raw The original value, i.e., the value before normalization; Q eduThis serves as a benchmark for passing the subject teaching test. Different subjects have different benchmark values; for example, it is 0.75 for science subjects and 0.68 for humanities subjects. min The lower bound for dimensional tolerance is dynamically calculated as the historical mean minus twice the standard deviation (2σ); U max This is the upper limit of dimensional tolerance, which is usually set to 1 because the range of normalized values ​​is [0,1].

[0144] By segmenting and mapping teaching baselines, the integrity dimension of humanities subjects is prevented from being "numerically suppressed".

[0145] Cognitive characteristic compensation design:

[0146] Dimension Compensation factor Compensation principle Accuracy ×1.05 High hazard of basic knowledge errors Integrity ×0.95 Allow gradual completion Consistency ×1.10 Systematic cognitive confusion caused by logical conflicts

[0147] A teaching value-driven model that optimizes priorities:

[0148] Dimensional deviation calculation, dynamic threshold setting, subject difference benchmark:

[0149] Natural Sciences, threshold _acc = 0.85, strictly controlling the accuracy of knowledge;

[0150] Threshold _comp = 0.75, allowing for gradual improvement;

[0151] Humanities and social sciences, threshold _acc = 0.78, accepting diversity of expression;

[0152] Threshold _comp = 0.82, emphasizing the overall context;

[0153] Scenario-sensitive adjustments have been made, raising the consistency threshold to 0.88 in exam preparation scenarios to prevent misleading associations;

[0154] Decoding the Educational Significance of Deviation Measurement:

[0155] Δ dim = |C dim -T dim ×Edu significance

[0156] In the formula, Δ dim This represents the dimensionality deviation, specifically the difference between the actual performance of a knowledge point in teaching and its target performance; C dim T represents the actual dimension value, that is, the student's actual performance on a certain knowledge point; dim This represents the target dimension value, that is, the target performance that students should achieve in a certain knowledge point; Edu significance The criticality coefficient for instruction is determined by the centrality of the knowledge point in the course graph. It is used to adjust the weight of the deviation and reflects the importance of the knowledge point.

[0157] Prioritization decision-making and the pedagogical rationale of the product model:

[0158] In the formula, Priority represents the priority of tasks or problems, that is, the execution order determined according to a certain criterion under limited resources; Δ dim This represents the dimensionality deviation, specifically the difference between the actual performance of a knowledge point in teaching and its target performance; W dim This represents the dimensional weights, i.e., the importance of different dimensions or knowledge points in the overall evaluation; Optimize Cost This refers to the optimization cost, which is the resources or effort required to improve the performance of a certain dimension or knowledge point.

[0159] Optimize cost factors:

[0160] Dimension Cost coefficient Influencing factors Accuracy 1.2 Need expert verification, time-consuming 3-5 days High efficiency of automated completion (hour level) 0.8 Consistency Involves interdisciplinary coordination 1.5 Problem type

[0161] Emergency intervention threshold: priority calculation will be skipped and processing will proceed directly when the following situations occur: accuracy deviation > 0.3 and affects core test points; consistency conflict causes collective cognitive confusion in ≥ 2 classes;

[0162] Parameter adjustment for educational scenarios:

[0163] Semantic validation tuning for accuracy, with targeted parameter adjustments:

[0164] Adjustment parameter Teaching significance Fuzzy concept expression BERT fine-tuning learning rate ↑ 30% Strengthen semantic differentiation ability Advanced knowledge missing Test analysis corpus weight × 1.5 Supplement high-order cognitive materials Discipline

[0165] Optimization of cognitive structure in terms of integrity:

[0166] Incompleteness → Gap type → Entity missing → Increase the intensity of textbook terminology extraction;

[0167] Gap type → Sparse relationship → Enhance the depth of cognitive relationship mining;

[0168] Gap type → Hierarchical fracture → Inject Bloom classification template;

[0169] Conflict resolution in the consistency dimension:

[0170] The rule base is dynamically enhanced, and a conflict path transformation mechanism is implemented. Rules are automatically generated for each valid conflict path detected.

[0171] Tolerance threshold elasticity:

[0172] Attribute error threshold Dynamic adjustment trigger condition Experimental science Involve safe operation (such as chemical reagent proportioning) ±3％→±1.5％ Historical research Time error ± 10 years Key turning point events (such as the Industrial Revolution) Invalid feedback type

[0173] S5 specifically includes:

[0174] When collecting teacher feedback scores, the feedback content is limited to specific evaluations by teachers on the accuracy of normalization of knowledge point entities in the knowledge graph, the rationality of standardization of relationships between entities, and the relevance of multi-dimensional confidence assessment results to actual teaching.

[0175] The collected teacher feedback ratings are validated for validity. Subjective evaluations that are not directly related to the quality of the knowledge graph are removed to form a valid feedback dataset. This dataset is used as the core component of the correction factor to adjust the influence coefficients of domain feature factors that are directly related to the teaching scenario in the dimensional weight adaptive model.

[0176] The S5 also includes:

[0177] The collected student learning outcome data includes data on changes in the mastery of knowledge points after learning activities based on the current knowledge graph, feedback data on the adaptability of learning paths, and data on the improvement of knowledge application ability.

[0178] The student learning outcome data was filtered for correlation, retaining only the data directly related to the knowledge graph quality evaluation dimension. The filtered data was then integrated with teacher feedback scores to form a correction factor.

[0179] When updating the feature factor influence coefficients of the adaptive model based on the correction factor, a phased update mechanism is adopted. First, local coefficients are adjusted to address the prominent issues reflected in the student learning performance data, and then the overall coefficients are calibrated in conjunction with teacher feedback scores.

[0180] It should be noted that there is a specific evaluation based on teacher feedback:

[0181] Evaluation dimensions, entity normalization accuracy verification:

[0182] Verification mechanism: If a teacher's terminology mapping deviation, such as "redox reaction" being incorrectly classified as "electrolysis reaction", is detected, and the ontology tree is reconstructed when ≥3 teachers report similar issues, the ontology tree reconstruction is triggered; the teacher's terminology sensitivity is converted into a graph semantic calibration signal;

[0183] Diagnosis of the rationality of relationship standardization, and review of cognitive logic:

[0184] Detect "forced associations" (such as hard links like "Du Fu → Calculus");

[0185] Identify "missing pivots" (e.g., "trigonometric functions" are not associated with "Fourier analysis");

[0186] Validity verification, subjective noise removal rules:

[0187] Filtering strategy Education principle Teaching equipment complaints Keyword shielding (such as "projector") Focus on the quality of knowledge ontology General emotional evaluation Ensure data objectivity Automatically discarded if sentiment value <-0.7 Interdisciplinary suggestions Save to cross-domain optimization pool Prevent weight model overload Learning effect data

[0188] Learning outcome data:

[0189] Data collection, tracking of changes in mastery levels, and dynamic cognitive modeling:

[0190] Pre-test knowledge level → Graphical learning intervention;

[0191] Post-test mastery → Δ mastery rate = Post-test - Pre-test;

[0192] Significance threshold: Δ > 15% is considered a valid improvement (p < 0.05);

[0193] Relevance filtering, knowledge graph attribution chain:

[0194] Retention condition Attribution strength Knowledge point mastery degree improvement Directly matched with atlas recommended content Problem solving ability transition 4 Involve atlas new relationship chain Interdisciplinary application innovation 3 Trigger atlas cross-domain association module Parameter type 2

[0195] Data fusion weighting strategy, teacher-student data weight ratio:

[0196] Basic subjects: 60% (teachers): 40% (students);

[0197] Practical courses: 45% (teachers): 55% (students);

[0198] Theoretical courses rely on teachers' experience, while practical courses focus on behavioral feedback.

[0199] Phased update mechanism:

[0200] Local emergency response mechanism, highlighting problem identification algorithm:

[0201]

[0202] In the formula, Urgency represents the urgency of the problem and is used to determine whether immediate action is needed; ΔScore represents the change in score, indicating the magnitude of the change in student performance, which can be positive or negative and is used to measure the fluctuation in performance; Student_Count represents the number of students, i.e., the total number of students affected; Days_Since_Update represents the number of days since the last update, indicating the length of time since the problem occurred.

[0203] Triggering condition: Urgency > 0.7 (e.g., a weekly drop in the grade for the electricity chapter in physics at a certain school > 25%);

[0204] Coefficient adjustment constraint rules:

[0205] Single maximum adjustment amplitude Anti-shock mechanism Disciplinary type factor Forbidden twice adjustment within 72 hours ±0.10 Application scene weight Association academic warning system linkage ±0.15 ​

[0206] Overall coefficient calibration:

[0207] The calibration cycle is synchronized with the teaching cycle. During the monthly exam period, the focus is on the mastery of knowledge points (student weight × 1.2); during the semester summary, the comprehensive evaluation of teachers is strengthened (teacher weight × 1.3).

[0208] Cross-factor coupling optimization:

[0209]

[0210] In the formula, α, β, and γ are the original weight coefficients used to assess the importance of different factors in the educational process; α′, β′, and γ′ are the optimized weight coefficients, adjusted after cross-factor coupling optimization; η is the adjustment coefficient used to fine-tune the weight coefficients to reflect the latest educational feedback or data; Δ T Δ S Δ C These represent the changes in time, space, and content dimensions, respectively, and are used to adjust the weighting coefficients to adapt to changes in the educational environment.

[0211] A matrix diagonal length greater than 0.7 ensures the stability of the principal factors; off-diagonal elements enable cross-dimensional collaboration, such as teacher feedback Δ. T Improve consistency (γ).

[0212] Example:

[0213] This embodiment uses junior high school mathematics, basic education, and personalized learning recommendation as examples to explain in detail the specific implementation process of this method:

[0214] Step 1: Obtain multi-source junior high school mathematics education data. The structured data is a database of knowledge points from junior high school mathematics textbooks (including entities such as "linear equations in one variable" and "inequalities" and explicit relationships such as "solution dependencies"); the semi-structured data is the hierarchical knowledge modules in the Ministry of Education's junior high school mathematics curriculum outline and the People's Education Press's lesson plans; the unstructured data is the monthly exam answer records of 1,000 students in the second year of junior high school in a certain middle school and the teaching feedback texts of 30 mathematics teachers.

[0215] Step 2: Construct a unified dictionary of junior high school mathematics terminology, collect terms from curriculum standards, textbooks and authoritative teaching aids, merge synonyms such as "linear equation in one variable" and "formula for linear equation in one variable", and establish mapping relationships; establish a mapping table of relationship types specific to the education field, clarify the standard definitions of 8 core relationships such as "prerequisite-subsequent", "subordinate-included", and "solution dependency", complete the normalization of knowledge point entities and the standardization of relationships, and output standardized graph data of junior high school mathematics.

[0216] Step 3: Determine the core dimensions of quality evaluation as accuracy, completeness, and consistency, and extract domain characteristic factors: subject type is natural science, education stage is basic education, and application scenario is personalized learning recommendation scenario;

[0217] Step 4: Construct a weight calculation function based on the above feature factors, and set the initial weights as follows: accuracy 0.4, completeness 0.3, consistency 0.3; obtain the influence coefficients of each feature factor through training with historical evaluation data, and dynamically adjust them in combination with the quantified values ​​of the feature factors to finally obtain the real-time weights as follows: accuracy 0.42, completeness 0.33, consistency 0.25;

[0218] Step 5: Using the BERT model in the education field (fine-tuned based on junior high school math textbooks and curriculum standards), the input triple "linear equation in one variable - solution dependency - equality properties" is used, and the semantic matching score of 0.85 is output as the initial accuracy confidence score; student answer records (800 in total) associated with this triple are extracted, and the answer accuracy rate is 75%. After correction with a preset threshold, the final accuracy confidence score is 0.83.

[0219] Step 6: Entity coverage is calculated as the ratio of the 280 junior high school math knowledge points actually included in the graph to the 300 core knowledge points stipulated in the curriculum standard (0.93); Relationship coverage is the ratio of the 420 relationships actually included in the graph to the total of 450 subject-specific logical relationships (0.93); In Bloom's hierarchy of knowledge coverage, the "memory" and "understanding" levels account for 60% of the weight, and the corresponding knowledge points account for 0.95. The three sub-dimensions are weighted and summed according to real-time weights to obtain an integrity confidence score of 0.94.

[0220] Step 7: Subject-specific logical conflict detection. By traversing the graph, no conflicting paths such as "irreversible prior relationships" were found, and the logical conflict coefficient was 0. Internal data conflict detection. By comparing the "definition of the solution of a linear equation in one variable" attribute in the multi-source data, there were no differences exceeding the allowable error, the attribute conflict rate was 0, and the calculated consistency confidence level was 0.98.

[0221] Step 8: Fuse multi-dimensional confidence scores according to real-time weights, and calculate the overall confidence score = (0.42×0.83+0.33×0.94+0.25×0.98) = 0.89. Set a preset threshold of 0.9. If the overall confidence score is lower than the threshold, calculate the deviations of each dimension: accuracy deviation 0.07, completeness deviation 0.06, and consistency deviation 0.02. Combine this with the real-time weights to obtain the priority: accuracy > completeness > consistency. Adjust the BERT model semantic matching threshold for the accuracy dimension and return to S3 for re-evaluation.

[0222] Step 9: Collect feedback scores from 30 teachers on the rationality of the association between knowledge points in the knowledge graph and the relevance of the assessment results to actual teaching, as well as data on the improvement in knowledge point mastery and learning efficiency of 1000 students after using the knowledge graph for personalized learning. Filter out effective information to form correction factors. Based on these correction factors, update the feature factor influence coefficients of the dimensional weight adaptive model, complete the model iteration, and use it for the next round of quality evaluation of junior high school mathematics knowledge graphs.

[0223] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for evaluating the quality of educational knowledge graphs based on confidence assessment, characterized in that, include: S1. Acquire structured, semi-structured, and unstructured educational data; normalize knowledge point entities by constructing a unified dictionary of subject terms; standardize relationships by establishing a mapping table of relationship types specific to the education field; and obtain standardized graph data. S2. Construct a dimension weight adaptive model, determine the core dimensions of quality evaluation as accuracy, completeness and consistency, extract subject type, education stage and application scenario as domain feature factors, construct a weight calculation function based on the factors, and dynamically adjust the real-time weights of each dimension. S3. Based on standardized graph data and real-time weights, multi-dimensional confidence assessment is conducted. Among them, accuracy is assessed by a dual mechanism of semantic verification of the BERT model in the education field and feedback correction of student answer records. Integrity is assessed by weighted evaluation from three sub-dimensions: entity coverage, relationship coverage, and Bloom's cognitive hierarchy coverage. Consistency is assessed by dual detection of subject logical conflict detection and data internal conflict detection. S4. Combine the confidence scores of multiple dimensions according to the real-time weights to obtain the comprehensive confidence score. If it is lower than the preset threshold, determine the optimization priority according to the product of the dimension deviation and the weight, adjust the corresponding evaluation parameters, and return to step S3. S5. Collect teacher feedback scores and student learning effect data as correction factors, update the feature factor influence coefficients of the dimension weight adaptive model, and achieve iterative optimization.

2. The method for evaluating the quality of educational knowledge graphs based on confidence assessment according to claim 1, characterized in that, S1 specifically includes: The method for constructing the unified dictionary of subject terminology is as follows: collect knowledge point terms from the national curriculum standards, authoritative textbooks and core academic literature of the corresponding subject, use the edit distance algorithm to calculate the character similarity of synonyms, combine the pre-trained language model in the education field to calculate the semantic similarity of terms, weight and fuse the character similarity and semantic similarity, merge terms with similarity ≥ 0.85, establish a one-to-one mapping relationship between non-standard terms and standard terms, and form a unified dictionary of subject terminology.

3. The method for evaluating the quality of educational knowledge graphs based on confidence assessment according to claim 1, characterized in that, S1 further includes: For the preprocessing of unstructured educational data, for student answer record texts, a word segmentation tool specifically for the education field is used to divide knowledge point related segments, and a named entity recognition model is used to extract knowledge point entities from the answer records; for teacher teaching feedback texts, sentiment polarity analysis is used to screen effective feedback, and the knowledge point relationship correction information involved in the feedback is extracted, and then the relationship standardization is completed by combining it with an education-specific relationship type mapping table.

4. The method for evaluating the quality of educational knowledge graphs based on confidence assessment according to claim 1, characterized in that, S2 specifically includes: The quantification method for the domain characteristic factors is as follows: in the subject type, natural sciences are assigned a value of 1.0 and humanities and social sciences are assigned a value of 0.8; in the education stage, basic education is assigned a value of 0.9 and higher education is assigned a value of 1.1; in the application scenario, knowledge point retrieval scenario is assigned a value of 0.9 and personalized learning recommendation scenario is assigned a value of 1.2; the quantified values ​​are calibrated using the Delphi method of domain experts, and the calibration cycle is once per semester.

5. The method for evaluating the quality of educational knowledge graphs based on confidence assessment according to claim 1, characterized in that, S3 further includes: In the dual-mechanism evaluation of accuracy, the training corpus of the BERT model in the education field includes the curriculum standards, textbook chapter content, and analysis of high-difficulty test questions for the corresponding subject. During fine-tuning, a triple (s,p,o) semantic matching task is used, with the input being the concatenated text "s+p+o". The output layer uses the sigmoid function to generate an initial accuracy confidence score in the 0-1 interval. The feedback correction of student answer records only includes answer data with a recent sample size ≥50. When the accuracy is lower than 60% of the preset threshold, the correction coefficient is forcibly set to 0.6, and when it is higher than 140% of the threshold, it is forcibly set to 1.

4.

6. The method for evaluating the quality of educational knowledge graphs based on confidence assessment according to claim 1, characterized in that, S3 further includes: In the weighted evaluation of integrity across its three sub-dimensions, redundant entities with duplicate annotations are removed when calculating entity coverage. Relationship coverage prioritizes the coverage ratio of core relationships, with core relationships accounting for no less than 60% of the total relationship types. The weights of the three sub-dimensions of Bloom's cognitive hierarchy coverage are dynamically split according to the allocation ratio of the integrity dimension in the real-time weights, with entity coverage accounting for 40%, relationship coverage accounting for 35%, and cognitive hierarchy coverage accounting for 25%.

7. The method for evaluating the quality of educational knowledge graphs based on confidence assessment according to claim 1, characterized in that, S3 further includes: In the dual-detection evaluation of consistency, the subject-specific logical conflict detection constructs a subject-specific knowledge logical rule base and uses a graph traversal algorithm to scan the paths between entities. For each conflicting path found, the logical conflict coefficient increases by 0.1 (maximum 1). In the internal data conflict detection, the attribute conflict rate is equal to the number of entities with attribute value differences divided by the total number of entities. When the difference value is within the allowable error range, it is not counted as a conflict.

8. The method for evaluating the quality of educational knowledge graphs based on confidence assessment according to claim 1, characterized in that, S4 specifically includes: When fusing multi-dimensional confidence scores with real-time weights to obtain a comprehensive confidence score, the confidence score evaluation results of each dimension are first uniformly converted to the same evaluation scale. When the overall confidence level is lower than the preset threshold, the difference between the confidence level of each dimension and the preset quality threshold of the corresponding dimension is first calculated as the dimension deviation. Then, the optimization priority of each dimension is determined by multiplying the dimension deviation with the real-time weight of that dimension. The larger the product, the higher the optimization priority of the dimension. For dimensions with higher priority, adjust their corresponding evaluation parameters. Specifically, adjust the semantic verification parameters for the accuracy dimension, adjust the sub-dimension weight allocation parameters for the integrity dimension, and adjust the conflict detection parameters for the consistency dimension. After the parameter adjustment is completed, re-execute the multi-dimensional confidence evaluation in step S3.

9. The method for evaluating the quality of educational knowledge graphs based on confidence assessment according to claim 1, characterized in that, S5 specifically includes: When collecting teacher feedback scores, the feedback content is limited to specific evaluations by teachers on the accuracy of normalization of knowledge point entities in the knowledge graph, the rationality of standardization of relationships between entities, and the relevance of multi-dimensional confidence assessment results to actual teaching. The collected teacher feedback ratings are validated for validity. Subjective evaluations that are not directly related to the quality of the knowledge graph are removed to form a valid feedback dataset. This dataset is used as the core component of the correction factor to adjust the influence coefficients of domain feature factors that are directly related to the teaching scenario in the dimensional weight adaptive model.

10. The method for evaluating the quality of educational knowledge graphs based on confidence assessment according to claim 1, characterized in that, The S5 also includes: The collected student learning outcome data includes data on changes in the mastery of knowledge points after learning activities based on the current knowledge graph, feedback data on the adaptability of learning paths, and data on the improvement of knowledge application ability. The student learning outcome data was filtered for correlation, retaining only the data directly related to the knowledge graph quality evaluation dimension. The filtered data was then integrated with teacher feedback scores to form a correction factor. When updating the feature factor influence coefficients of the adaptive model based on the correction factor, a phased update mechanism is adopted. First, local coefficients are adjusted to address the prominent issues reflected in the student learning performance data, and then the overall coefficients are calibrated in conjunction with teacher feedback scores.

Images