Academic information conflict resolution method based on evidence intensity index grading

Abstract

The invention discloses an academic information conflict resolution method based on evidence intensity index grading. According to the method, all conflict viewpoints and traceability chains thereof are reserved by adopting a'full-amount retention, non-merging and non-deleting 'strategy; calculating a multi-dimensional weighted evidence intensity index of a comprehensive journal influence factor, an introduced frequency and an author academic influence for each literature; judging the conflict severity according to the difference value of the evidence intensity indexes of the two conflict parties, wherein the smaller the difference value is, the higher the severity is; modeling a conflict cluster into an auxiliary node layer in a graph database, and superposing the auxiliary node layer on a core graph without interfering with an ontology framework; conflict severity ranking results drive calibration reduction of downstream weight calculations. The method can be used for information conflict management in academic knowledge graph construction, and interpretable conflict severity grading reference is provided.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, specifically to a method for resolving academic information conflicts based on a graded evidence strength index. Background Technology

[0002] With the deepening development of knowledge graph technology in academic research and industrial applications, automatically extracting structured knowledge from multi-source heterogeneous literature has become a core step in knowledge graph construction. However, in the actual construction process, different literatures frequently contradict each other in their descriptions of the same fact. This is especially true in empirical sciences such as geology, medicine, and ecology, where academic consensus is dynamically evolving. Different research teams, based on different samples, methods, and theoretical frameworks, often give drastically different judgments on the genesis, properties, or classification of the same geological bodies.

[0003] Existing methods for handling knowledge graph conflicts mainly fall into two categories: voting mechanisms and confidence ranking. Voting mechanisms employ a "majority wins" strategy, using viewpoints supported by multiple sources as the final choice; confidence ranking, on the other hand, takes the viewpoint from the source with the highest confidence level. However, both methods are flawed in academia. Voting mechanisms implicitly assume "majority is correct," but the history of science repeatedly demonstrates that minority viewpoints ultimately replace majority consensus—a common occurrence in academic evolution, as evidenced by the long-standing rejection of plate tectonics by the majority proponents of geosyncline theory. Confidence ranking, on the other hand, ignores the fact that highly cited literature may have accumulated more citations due to its earlier publication date, while newer research with lower citation counts may contain more accurate insights.

[0004] Furthermore, knowledge graph embedding-based methods (such as link prediction models like TransE) implicitly judge the rationality of triples in the vector space. While this can assess the consistency of knowledge to some extent, it cannot preserve the complete viewpoints of both conflicting parties, nor can it provide an interpretable judgment on the severity of the conflict. Constraint verification tools such as SHACL aim to "correct errors," treating inconsistencies as defects that need to be corrected. However, in academic contexts, conflict is not an error but an expression of cognitive uncertainty. Forcibly "correcting errors" may delete important minority viewpoints.

[0005] Therefore, existing technologies lack a systematic approach that can both preserve all conflicting viewpoints and provide a differentiated grading reference for the strength of evidence, in order to support the reliable application of knowledge graphs in academic information management.

[0006] To address the aforementioned issues, this invention provides a method for resolving academic information conflicts based on evidence strength index grading. This method aims to solve the technical problem in existing technologies where multi-source document conflicts cannot simultaneously achieve full retention and differentiated grading management during knowledge graph construction. Summary of the Invention

[0007] The purpose of this invention is to provide a method for resolving academic information conflicts based on a strength of evidence index grading system. The method specifically includes the following steps: S1. Describe the contradictions of the same fact supported by multiple sources in the knowledge graph. In the knowledge graph triple extraction stage, identify the contradictory descriptions of the same fact supported by multiple sources in the knowledge graph. Use a full retention strategy to save all conflicting viewpoints. Prohibit the automatic merging or deletion of any conflicting viewpoints in the extraction stage. Each conflicting viewpoint is accompanied by a complete literature tracing chain, including literature DOI, fingerprint deduplication, original text paragraph location, extraction timestamp, and extraction model version. S2. Deduplicatize the documents. When a document has a DOI, use the DOI as the unique primary key. Different language versions of the same document with the same DOI are considered the same document. When a document does not have a DOI, use a combination of the document title, publication year, and first author to generate a document fingerprint as the deduplication key. S3. To support conflicting viewpoints, calculate the multidimensional weighted Evidence Strength Index (ESI) for each document. The Evidence Strength Index is a weighted sum of three dimensions: the normalized value of the impact factor of the journal in which the document is located, the normalized value of the total number of citations of the document, and the normalized value of the academic influence index of the first author or corresponding author of the document. S4. For conflicting viewpoints on the same fact, calculate the difference between the evidence strength indices of the corresponding documents, and determine the severity level of the conflict based on the difference. The smaller the difference, the higher the severity. S5. Model each group of conflicting viewpoints as conflict auxiliary nodes in the graph database. Conflicting viewpoints are connected to conflict auxiliary nodes through the establishment of disputed view relationships. Conflicting viewpoints are linked in two directions through contradictory relationships. Conflict auxiliary nodes are superimposed on the core graph as a label layer and do not participate in the definition of the core graph's ontology TBox framework. S6. Drive the calibration reduction of downstream weight calculations by using the conflict severity classification results. Furthermore, the complete formula for calculating the Evidence Strength Index (ESI) in S3 is as follows: ESI = α×IF_norm + β×Citation_norm + γ×H_norm (1) Wherein, IF_norm is the normalized value of the impact factor of the journal in which the article is published, Citation_norm is the normalized value of the total number of citations of the article, H_norm is the normalized value of the H index of the first author or corresponding author of the article, and α, β, γ are weight coefficients and α+β+γ=1. When the non-zero fill rate of the above H-index data is lower than a preset threshold, it automatically switches to the simplified calculation formula: ESI_simple = α'×IF_norm + β'×Citation_norm (2) Among them, α' and β' are obtained by renormalizing α and β in the full version according to the original ratio, that is, α' = α / (α + β), β' = β / (α + β).

[0008] Furthermore, the determination of the conflict severity level in S4 adopts counterintuitive logic, and the specific determination rules are as follows: S41. High level (high): ESI difference ≤ 0.30, the authority of the literature evidence on both sides is close, and the disagreement is difficult to be automatically adjudicated by the evidence strength, and manual experts need to be involved in the adjudication; S42. Medium level (medium): 0.30 < ESI difference ≤ 0.50, there is a certain gap between the evidence on both sides, but it still constitutes a substantial disagreement, and manual review is recommended; S43. Low level (low): ESI difference > 0.50, the evidence of one side is significantly more authoritative, and the conflict can be naturally alleviated through evidence grading, and the view of the side with high ESI can be used as the priority reference.

[0009] Furthermore, the graph modeling of the conflict auxiliary nodes in S5 includes creating a dispute cluster auxiliary node for each conflict cluster, and creating a dispute cluster auxiliary node in the graph database with the conflict fact description, conflict severity level, ESI difference and conflict status as node attributes.

[0010] Furthermore, the establishment method of the dispute view relationship in S5 is as follows: in the graph database, taking the conflict view triple node as the starting node and the dispute cluster auxiliary node corresponding to the conflict cluster to which the conflict view belongs as the target node, perform a directed relationship creation operation, specify the relationship type as the dispute view, and write the evidence strength index score, evidence strength level and literature traceability chain of the literature corresponding to the view into the attribute field of this relationship.

[0011] Furthermore, the establishment method of the contradiction relationship in S5 is as follows: for each pair of triple nodes with contradictory views within the same dispute cluster, perform two symmetric relationship creation operations in the graph database. The first time, create a contradictory opposition relationship with the first triple node as the starting node and the second triple node as the target node. The second time, create a contradictory opposition relationship with the second triple node as the starting node and the first triple node as the target node. The attributes of both relationships include the contradiction type.

[0012] Furthermore, the calculation formula for the downstream impact of the calibration reduction in S6 is: IWI_final = FW × ConflictCalibration ···(3) Wherein, FW is the basic fusion weight of the entity node, ConflictCalibration is the conflict calibration factor, which is 0.85 at high levels and 1.0 at medium and low levels, and IWI_final is the final comprehensive weight index after calibration.

[0013] The present invention has the following beneficial effects: By adopting the strategy of "full retention, no merging and no deletion", the present invention avoids the problem of losing minority academic viewpoints due to automatic merging or voting mechanisms during the knowledge extraction stage, so that the knowledge graph can fully record the diversity and uncertainty of academic cognition, and retain a complete information foundation for subsequent manual review and academic analysis.

[0014] This invention calculates an evidence strength index for each document across three dimensions: journal level, article level, and author level. This enables a quantitative assessment of the authority of documentary evidence, providing an interpretable and auditable quantitative basis for conflict management and overcoming the irrationality of existing voting mechanisms and confidence ranking methods in the academic field.

[0015] This invention employs a counterintuitive judgment logic that "the smaller the difference, the higher the severity," thereby accurately identifying high-risk conflicts where the authority of the evidence from both parties is similar and the academic dispute is most substantial. This allows conflicts requiring human intervention to be prioritized and processed, improving the efficiency and accuracy of conflict management.

[0016] By modeling conflict clusters as an auxiliary node layer superimposed on the core graph, the structured expression and management of conflict information are achieved without destroying the core graph's ontological structure, ensuring that the topological integrity and semantic consistency of the core graph are unaffected by conflict annotations. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the Evidence Strength Index (ESI) calculation process of the present invention; Figure 3 This is a schematic diagram of the conflict auxiliary node graph modeling structure of the DisputeCluster of the present invention; Figure 4 This is a schematic diagram of the conflict severity classification and determination logic of the present invention; Figure 5 This is a schematic diagram of the sensitivity analysis results of the ESI weight coefficients in this invention; Figure 6 This is a schematic diagram of a module of an academic information conflict resolution system based on evidence strength index grading according to the present invention. Detailed Implementation

[0018] This invention provides a method for resolving academic information conflicts based on a graded evidence strength index. Taking the multi-source literature conflict in the metallogenic mechanism of porphyry copper deposits during the construction of a geological knowledge graph as an application scenario, the implementation process of this invention is described in detail. In this scenario, when constructing a knowledge graph of the porphyry copper deposit metallogenic system, structured triplets were extracted from 376 academic documents in Chinese, English, and Russian. Numerous factual descriptions concerning magma water content, alteration zoning sequences, mineralization temperature ranges, and mineral assemblages contained numerous inter-documentary contradictions. The specific steps are as follows: S1. In the knowledge graph triple extraction stage, contradictory descriptions of the same fact supported by multiple source documents in the knowledge graph are identified. When the subject entity and predicate relationship of two or more triples are the same, but the object entity or attribute value is contradictory, it is identified as a set of conflicts. S11. Adopting the rigid principle of "full retention, no merging and no deletion", any conflicting viewpoints are prohibited from being automatically merged or deleted during the extraction stage. All conflicting viewpoints are fully retained, and each viewpoint is accompanied by a complete literature tracing chain, including the literature DOI or deduplication fingerprint, the location of the original paragraph in the literature, the extraction timestamp, and the extraction model version.

[0019] S12. The academic rationale for the "full data retention" principle can be demonstrated from three dimensions: (1) From the perspective of the history of science, a large number of cases show that it is normal for minority opinions to eventually replace the consensus of the majority in academic evolution. For example, in geology, the plate tectonics theory replaced the geosyncline theory, and the rise and controversy of the mantle plume theory all show that the minority at that time later became the new consensus.

[0020] (2) From the perspective of information theory, conflict information itself carries a quantitative signal of the uncertainty of the fact. If a fact is not conflicted, its credibility can be directly adopted; if there is a conflict, the ESI difference of the conflict itself is a quantitative measure of uncertainty.

[0021] (3) From an engineering perspective, once a conflict is merged during the extraction phase, the deleted minority viewpoints cannot be recovered later, while the retained ones can be filtered at any time. Irreversible operations should be performed at the end of the pipeline rather than at the beginning.

[0022] Taking the magma water content of porphyry copper deposits as an example, Reference A (published in Economic Geology, 2018) states that "the initial magma water content of porphyry copper deposits is usually in the range of 4-6 wt%"; Reference B (published in Nature Geoscience, 2022) proposes that "the latest fluid inclusion evidence shows that the initial water content can reach 8-12 wt%". The two descriptions of the fact of "initial magma water content of porphyry copper deposits" are contradictory, and both viewpoints are fully preserved.

[0023] S2. Perform document deduplication to prevent the same document from being counted repeatedly due to different sources. The deduplication key generation specification is as follows: S21. DOI-Priority Fingerprint: When a document has a DOI identifier, the DOI is used as the unique primary key of the document. Different language versions of the same document (e.g., Chinese and English versions) are considered the same document if they have the same DOI, and the strength of evidence index is taken as the highest value among the versions. S22. Combined fingerprint: When a document does not have a DOI identifier (such as some early Russian documents and gray documents), a fingerprint is generated by combining the document title, publication year, and first author as the deduplication key. When generating the fingerprint, the title is lowercase and punctuation is removed, and the author's name is taken in the standardized form of the surname. For example, 23 duplicate documents were identified from 376 documents (17 of which were different language versions with the same DOI, and 6 were combined fingerprints without DOI). After deduplication, 353 unique documents were obtained.

[0024] S3, such as Figure 3 As shown, a multi-dimensional weighted Evidence Strength Index (ESI) is calculated for each article supporting conflicting viewpoints. The ESI integrates three dimensions: First Dimension: Journal Impact Factor Normalized Value (IF_norm): This value is the latest impact factor of the journal containing the document. Min-Max normalization is performed on the current document set, mapping it to the [0, 1] interval. The normalization formula is: IF_norm = (IF_i - IF_min) / (IF_max - IF_min) ··· Where IF_i is the impact factor of the journal in which the i-th article is located, and IF_min and IF_max are the minimum and maximum impact factors in the current document set, respectively; The second dimension: Normalized total citation count (Citation_norm): This value represents the total number of citations a document has received since publication. It is also normalized using the Min-Max formula: Citation_norm = (Cite_i - Cite_min) / (Cite_max - Cite_min) ··· Where Cite_i is the total number of citations for the i-th document, and Cite_min and Cite_max are the minimum and maximum number of citations in the current document set, respectively; The third dimension: Normalized value of author's academic influence index H_norm: Obtain the H-index of the first author or corresponding author of the paper, and perform Min-Max normalization. The normalization formula is as follows: H_norm = (H_i - H_min) / (H_max - H_min) ··· Where H_i is the H-index of the first / corresponding author of the i-th document, and H_min and H_max are the minimum and maximum H-index values ​​in the current document set, respectively; S31. The complete formula for calculating the strength of evidence index is as follows: ESI = α×IF_norm + β×Citation_norm + γ×H_norm··· Where α, β, and γ are weighting coefficients, α=0.40, β=0.35, γ=0.25, α+β+γ=1, and the determination of α, β, and γ is based on the following: From a bibliometric perspective, journal level (impact factor) has the highest discriminative power for document quality, and is therefore assigned the largest weight of 0.40; article level (citation count) reflects the actual influence of the document in the academic community, and is assigned the second largest weight of 0.35; author level (H-index) is most affected by differences in field and discipline size, and is assigned the lowest weight of 0.25.

[0025] From a sensitivity analysis perspective, grid enumeration was performed on the weight coefficients within the range of 0.20 to 0.60 (summing up to 1) across all dimensions. Using the "Kendall τ correlation coefficient between ESI ranking and manual ranking by 8 domain experts" as the evaluation metric, combinations that maximize τ were selected, as shown in Table 1.

[0026] Analysis of Table 1 shows that the combination of α=0.40, β=0.35, and γ=0.25 results in Kendall τ reaching its highest value of 0.86. From a robustness verification perspective, the optimal weight combination was perturbed within ±0.05, and the reversal rate of the conflict severity determination results (i.e., the proportion of a certain conflict severity level that changes before and after the perturbation) was calculated. In the 87th group of conflicts, the reversal rate was 3.4% (3 groups), which is lower than the acceptable threshold of 5%, proving that the coefficient selection has good stability.

[0027] When the non-zero fill rate of the H-index data is lower than a preset threshold (70%), it automatically switches to the simplified calculation formula of the evidence strength index: ESI_simple = α'×IF_norm + β'×Citation_norm··· Among them, α' = α / (α + β) = 0.40 / (0.40 + 0.35) ≈ 0.53, β' = β / (α + β) = 0.35 / (0.40 + 0.35) ≈ 0.47. The simplified version coefficient is obtained by renormalizing the full version according to the original ratio, keeping the relative weight relationship between the two dimensions unchanged.

[0028] Among the 353 documents, 312 (88.4%) have H-index data available, and the filling rate is 88.4%, which is higher than the 70% threshold. Therefore, the full version calculation formula of the evidence strength index is automatically enabled.

[0029] Taking the conflict of magma water content as an example, the ESI calculation processes of two documents are as follows: Document A (Economic Geology, 2018): Journal impact factor 4.23, IF_norm = 0.62; total citations 187, Citation_norm = 0.71; first author H-index 28, H_norm = 0.55. ESI_A = 0.40×0.62 + 0.35×0.71 + 0.25×0.55 = 0.248 + 0.249 + 0.138 = 0.635.

[0030] Document B (Nature Geoscience, 2022): Journal impact factor 18.32, IF_norm = 0.95; total citations 42, Citation_norm = 0.23; first author H-index 19, H_norm = 0.38. ESI_B = 0.40×0.95 + 0.35×0.23 + 0.25×0.38 = 0.380 + 0.081 + 0.095 = 0.556.

[0031] The ESI difference between the two documents is |0.635 - 0.556| = 0.079.

[0032] S4. As Figure 4 shown, the conflict severity is determined based on the ESI difference between the two conflicting documents. The present invention adopts a counterintuitive determination logic: the smaller the difference, the higher the severity. The specific determination rules are as follows: (1) High level (high): ESI difference ≤ 0.30. The evidence authority of the two documents is close, and the disagreement is difficult to be automatically adjudicated by the evidence strength, and manual expert intervention is required for adjudication.

[0033] (2) Medium level (medium): 0.30 < ESI difference ≤ 0.50. There is a certain gap between the two pieces of evidence, but it still constitutes a substantial disagreement, and manual review is recommended.

[0034] (3) Low level: ESI difference > 0.50, one side’s evidence is obviously more authoritative, and the conflict can be naturally alleviated through evidence classification. The view of the side with high ESI can be given priority reference.

[0035] The basis for this counterintuitive design is that when the ESI scores of two articles are very close, it means that they are evenly matched in terms of journal quality, academic influence, and author authority. It is impossible to determine which is better based on quantitative indicators alone. It is necessary for experts in the field to intervene and analyze the specific argumentation logic and experimental data. Conversely, when the ESI scores are significantly different, the viewpoint of the article with the higher ESI score can be given higher initial credibility, and the actual impact of the conflict can be controlled.

[0036] In S3, the ESI difference between literature A and literature B is 0.079, which is much smaller than the first threshold of 0.30. Therefore, the conflict is judged to be of a high level. This judgment is consistent with the actual situation in the geological community: Economic Geology and Nature Geoscience are both high-impact journals, and the disagreement about magma water content is indeed a current hot topic of academic debate.

[0037] In the 87th group of conflicts, after ESI calculation and difference determination: 31 groups of high-level conflicts (35.6%), 24 groups of medium-level conflicts (27.6%), and 32 groups of low-level conflicts (36.8%).

[0038] S5, such as Figure 3 As shown, each set of conflicting viewpoints is modeled as a conflict auxiliary node structure in a graph database (such as Neo4j); S51. Each conflicting viewpoint (i.e., a triplet node containing a contradictory description) is connected to the DisputeCluster auxiliary node through the DISPUTED_VIEW relationship. The DISPUTED_VIEW relationship carries the following attributes: the ESI score (esiScore) of the corresponding document, the strength level of evidence (with values ​​of HIGH / MEDIUM / LOW), and the source chain (including document DOI / fingerprint, original paragraph location, etc.). S52. Triple nodes holding contradictory viewpoints establish opposing links through the CONTRADICTS bidirectional relationship. The attributes of the CONTRADICTS relationship include the contradiction type (contradictionType, with values ​​such as VALUE_CONFLICT / CLASSIFICATION_CONFLICT / CAUSATION_CONFLICT, etc.). S53. Key design constraints: The DisputeCluster auxiliary nodes are only superimposed on the core knowledge graph as an annotation layer. All DisputeCluster nodes and the DISPUTED_VIEW and CONTRADICTS relationships do not participate in the definition of the ontology TBox framework of the core graph, and do not affect the Schema hierarchical structure and inference logic of the core graph. When the downstream application does not require conflict information, the DisputeCluster layer can be directly excluded through label filtering to obtain a "clean" view of the core graph. Taking the conflict of magma water content as an example, its graph modeling structure is as follows: Create a DisputeCluster node DC-001 with attributes {factDescription: "Initial magma water content range of porphyry copper deposit", severityLevel: "high", esiDelta: 0.079, status: "OPEN"}.

[0039] The triple node T-A ("view of water content 4-6 wt%", source literature A) is connected to DC-001 through the DISPUTED_VIEW relationship (esiScore: 0.635, strengthLevel: "HIGH").

[0040] The triple node T-B ("view of water content 8-12 wt%", source literature B) is connected to DC-001 through the DISPUTED_VIEW relationship (esiScore: 0.556, strengthLevel: "MEDIUM").

[0041] A CONTRADICTS two-way relationship (contradictionType: "VALUE_CONFLICT") is established between A and T-B.

[0042] S6. The calibration reduction of downstream weight calculation is driven by the conflict severity classification result. The specific reduction rules are as follows: For conflicts with a high severity level (ESI difference ≤ 0.30), multiply the weight of the entity node associated with this conflict by the reduction factor 0.85. The intention of this reduction is that a high severity level conflict indicates a large uncertainty in the reliability of this knowledge point, and the contribution of its downstream weight calculation should be appropriately reduced to prompt the decision maker to treat this information with caution.

[0043] For conflicts with a medium severity level (0.30 < ESI difference ≤ 0.50) and a low severity level (ESI difference > 0.50), the weights of the associated entity nodes remain unchanged, and the calibration reduction factor is 1.0.

[0044] The formula for the downstream effects of calibration reduction is: IWI_final = FW × ConflictCalibration··· Wherein, FW is the basic fusion weight of the entity node, ConflictCalibration is the conflict calibration factor, which is 0.85 at high level and 1.0 at medium and low level, and IWI_final is the final comprehensive weight index after calibration. In this study, the weights of the 47 entity nodes involved in the 31 high-level conflicts were all multiplied by a reduction factor of 0.85. The selection of the reduction factor of 0.85 was based on the need to strike a balance between "signal preservation" and "risk labeling." Excessive reduction (such as 0.50) would lead to over-suppression of highly controversial but potentially correct knowledge; while excessive reduction (such as 0.95) would fail to effectively label risks. 0.85 corresponds to a 15% reduction in weight. In subsequent experimental verification, this reduction magnitude ensured that highly controversial conflict areas were appropriately downweighted in the downstream prediction model without completely obscuring their information contribution. The specific calculation details of how the reduction factor is calibrated and incorporated into the downstream weight synthesis formula are not within the scope of this method. This method only provides the output interface for conflict severity classification and reduction factor.

[0045] In this embodiment, in the porphyry copper deposit knowledge graph construction scenario, 12,847 triples were extracted from 376 documents (353 after deduplication), 87 conflicting triples were identified, involving 284 conflicting triples. The statistical results after ESI calculation and conflict severity classification are shown in Table 2: Typical controversial topics involved in high-level conflicts include: magma water content range (3 groups), mineralization alteration zonation sequence (5 groups), proportion of ore-forming fluid sources (4 groups), emplacement depth of porphyry bodies (2 groups), and ore-forming temperature range (6 groups), all of which are active research frontiers in the current porphyry copper deposit metallogenic theory.

[0046] The determination results of 87 conflict groups were manually verified: three geologists with more than 10 years of experience in porphyry copper deposit research independently assessed the actual severity of each conflict group. The majority consensus was used as the benchmark and compared with the ESI classification results of this invention. The consistency rate was 82.8% (72 / 87), with the highest consistency rate for high-level conflicts (90.3%, 28 / 31). This shows that the determination logic of "the smaller the difference, the more serious" is highly consistent with the intuitive judgment of experts in the field.

[0047] This invention takes approximately 2.3 seconds to perform ESI calculations on 353 documents on a standard computing device (Intel i7-12700, 32GB RAM), approximately 0.1 seconds to determine the severity of 87 conflict groups, and approximately 1.8 seconds to perform batch writing of the DisputeCluster graph model in the Neo4j database. The entire process can be completed within 5 seconds.

[0048] This invention can be used in an academic information conflict resolution system based on evidence strength index grading, including a conflict identification and retention module, an evidence strength index calculation module, a conflict severity determination module, a conflict graph modeling module, and a calibration and reduction module. The conflict identification and retention module is used to identify contradictory descriptions of the same fact supported by multiple sources in the knowledge graph, and adopts a full retention strategy to retain all conflicting viewpoints, prohibiting the automatic merging or deletion of any conflicting viewpoints. The evidence strength index calculation module is used to calculate a multi-dimensional weighted evidence strength index for each document supporting the conflicting viewpoint, and performs a weighted summation of three dimensions: the normalized value of the journal impact factor, the normalized value of the total number of citations of the document, and the normalized value of the author's academic influence index. The conflict severity determination module is used to calculate the difference in evidence strength indices between the conflicting documents, and determine the severity level of the conflict based on the difference. The smaller the difference, the higher the severity. The conflict graph modeling module is used to model each group of conflicting viewpoints as conflict auxiliary nodes in the graph database, and construct a conflict annotation layer through disputed view relationships and contradictory relationships. The calibration and reduction module is used to perform differentiated reduction on the downstream weight calculation based on the conflict severity grading results.

[0049] This invention can be used in an electronic device, including a processor and a memory, wherein the processor is configured to run a computer program to perform an academic information conflict resolution method based on an evidence strength index grading.

[0050] This invention can be used in a computer-readable storage medium storing a computer program that, when executed by a processor, implements steps for resolving academic information conflicts based on a strength of evidence index grading.

[0051] Although specific embodiments of the present invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of the present invention. Various modifications and variations that can be made by those skilled in the art without inventive effort within the scope described in the claims are still within the scope of protection of the present invention.

Claims

1. A method for resolving academic information conflicts based on a strength of evidence index grading system, characterized in that: The method specifically includes the following steps: S1. Describe the contradictions of the same fact supported by multiple sources in the knowledge graph. In the knowledge graph triple extraction stage, identify the contradictory descriptions of the same fact supported by multiple sources in the knowledge graph. Use a full retention strategy to save all conflicting viewpoints. Prohibit the automatic merging or deletion of any conflicting viewpoints in the extraction stage. Each conflicting viewpoint is accompanied by a complete literature tracing chain. S2. Deduplicatize the documents. When a document has a DOI, use the DOI as the unique primary key. Different language versions of the same document with the same DOI are considered the same document. When a document does not have a DOI, use a combination of the document title, publication year, and first author to generate a document fingerprint as the deduplication key. S3. To support conflicting viewpoints, calculate the multidimensional weighted Evidence Strength Index (ESI) for each document. The Evidence Strength Index is a weighted sum of three dimensions: the normalized impact factor of the journal in which the document is located, the normalized total citation count of the document, and the normalized academic influence index of the first author and corresponding author of the document. S4. For conflicting viewpoints on the same fact, calculate the difference between the evidence strength indices of the corresponding documents, and determine the severity level of the conflict based on the difference. The smaller the difference, the higher the severity. S5. Model each group of conflicting viewpoints as conflict auxiliary nodes in the graph database. Conflicting viewpoints are connected to the conflict auxiliary nodes through the establishment of disputed view relationships. Conflicting viewpoints are linked in two directions through contradictory relationships. The conflict auxiliary nodes are superimposed on the core graph as a labeling layer and do not participate in the definition of the core graph's ontology TBox framework. S6. The calibration reduction of downstream weight calculation is driven by the conflict severity classification results.

2. The academic information conflict resolution method based on evidence strength index grading as described in claim 1, characterized in that: The complete formula for calculating the Evidence Strength Index (ESI) in S3 is as follows: ESI = α×IF_norm + β×Citation_norm + γ×H_norm (1) Wherein, IF_norm is the normalized value of the impact factor of the journal in which the article is published, Citation_norm is the normalized value of the total number of citations of the article, H_norm is the normalized value of the H index of the first author or corresponding author of the article, and α, β, γ are weight coefficients and α+β+γ=1. When the non-zero fill rate of the above H-index data is lower than a preset threshold, it automatically switches to the simplified calculation formula: ESI_simple = α'×IF_norm + β'×Citation_norm (2) Among them, α' and β' are obtained by normalizing α and β in the complete version according to the original proportion, that is, α'=α / (α+β) and β'=β / (α+β).

3. The academic information conflict resolution method based on evidence strength index grading as described in claim 1, characterized in that: The difference between the evidence strength indices of the corresponding documents in S4 is calculated as follows: obtain the evidence strength index of each document that supports the viewpoints of both sides of the conflict, denote the evidence strength indices of the documents of both sides of the conflict as ESI_A and ESI_B, and calculate the absolute value of the difference between the two as the ESI difference, that is, ΔESI=|ESI_A−ESI_B|.

4. The academic information conflict resolution method based on evidence strength index grading as described in claim 1, characterized in that: The determination of the conflict severity level in S4 uses counterintuitive logic, and the specific determination rules are as follows: S41, High Level: ESI difference ≤ 0.30, the authority of the evidence from both sides is similar, the disagreement is difficult to be automatically decided by the strength of evidence, and human expert intervention is required for adjudication; S42, Medium: 0.30 < ESI difference ≤ 0.50, there is a certain gap between the evidence from both sides, but it still constitutes a substantial disagreement, and manual review is recommended; S43, Low level: ESI difference > 0.50, one side's evidence is significantly more authoritative, and the conflict can be naturally alleviated through evidence classification. The viewpoint of the side with higher ESI can be given priority.

5. The academic information conflict resolution method based on evidence strength index grading as described in claim 1, characterized in that: The graph modeling of the conflict auxiliary nodes in S5 includes creating a dispute cluster auxiliary node for each conflict cluster, and creating dispute cluster auxiliary nodes in the graph database with conflict fact description, conflict severity level, ESI difference and conflict status as node attributes.

6. The academic information conflict resolution method based on evidence strength index grading as described in claim 1, characterized in that: The method for establishing the disputed view relationship in S5 is as follows: In the graph database, with the conflict view triple node as the starting node and the dispute cluster auxiliary node corresponding to the conflict cluster to which the conflict view belongs as the target node, a directed relationship creation operation is performed, the relationship type is specified as disputed view, and the evidence strength index score, evidence strength level and literature tracing chain of the literature corresponding to the view are written into the attribute fields of the relationship.

7. The academic information conflict resolution method based on evidence strength index grading as described in claim 1, characterized in that: The contradictory relationship in S5 is established as follows: for each pair of triple nodes holding contradictory views within the same dispute cluster, two symmetrical relationship creation operations are performed in the graph database. The first operation creates a contradictory relationship with the first triple node as the starting node and the second triple node as the target node. The second operation creates a contradictory relationship with the second triple node as the starting node and the first triple node as the target node. Both relationships include the contradictory type as their attributes.

8. The academic information conflict resolution method based on evidence strength index grading as described in claim 1, characterized in that: The formula for calculating the downstream effects of the calibration reduction in S6 is as follows: IWI_final = FW × ConflictCalibration···(3) Wherein, FW is the basic fusion weight of the entity node, ConflictCalibration is the conflict calibration factor, which is 0.85 at high levels and 1.0 at medium and low levels, and IWI_final is the final comprehensive weight index after calibration.

Images