Method and system for disambiguating author name

Abstract

The invention discloses an author name disambiguation method and system, and belongs to the technical field of natural language processing. The method comprises the following steps: acquiring and preprocessing a paper data set; blocking processing according to name fingerprints; performing first-stage merging based on same mechanism and text similarity; second-stage cross-mechanism judgment is carried out based on bidirectional partner coverage, time cross constraint and mechanism transfer prior; executing constrained graph clustering; according to the method, the disambiguation accuracy and recall rate of the authors with the same name can be improved under the condition that only abstracts and metadata exist, the situation that the authors replace work units can be further accurately recognized, and the method is suitable for scenes such as academic literature retrieval and researchers evaluation.

Description

Technical Field

[0001] This application belongs to the field of natural language processing technology and discloses a method and system for author name disambiguation. Background Technology

[0002] In the fields of natural language processing, information retrieval, and data mining, accurately identifying paper authors is a fundamental and crucial task in applications such as academic literature retrieval, researcher evaluation, and academic collaboration network analysis. With the rapid development of academic research, the number of research papers is growing exponentially, posing numerous challenges to author identification: different researchers may share the same name, especially prevalent in Chinese; researchers may change employers during their careers, resulting in different affiliations for papers published by the same author at different times; paper metadata may contain quality issues such as spelling errors and inconsistent formatting; and modern research involves increasing interdisciplinary collaboration, meaning the same author may publish papers in different academic fields.

[0003] Existing methods for disambiguating author names have significant shortcomings: some methods employ multi-feature weighted similarity and threshold merging, lacking consideration for the directionality and time decay of collaborator relationships, and failing to specifically design a mechanism for identifying changes in work unit; some methods use a two-stage clustering strategy, with the first stage based on collaborator relationship graph similarity and AP clustering, and the second stage based on inter-cluster semantic similarity of document vectors, and adopt a conservative strategy of "not merging different schools," affecting the recall rate of authors who have changed work units; some methods prioritize email consistency, combining similarity of institutions, disciplines, co-authorship networks, and citation relationships, but lack a hard constraint mechanism in the time dimension, resulting in poor performance in sparse data scenarios; and some methods use Chinese-to-English name conversion rules combined with TF-IDF features and classifiers, failing to fully utilize the structural information of collaborator networks and insufficiently considering scenarios involving changes in work unit.

[0004] Therefore, the urgent technical problem to be solved is to improve the accuracy and recall of disambiguation of authors with the same name, especially to accurately identify cases where authors have changed their work units, given only abstracts and metadata. Summary of the Invention

[0005] This application provides a method and system for author name disambiguation, which solves the problem of improving the accuracy and recall of disambiguation of authors with the same name when only abstracts and metadata are available, especially in accurately identifying cases where authors have changed their work units. The content is as follows.

[0006] On the one hand, an author name disambiguation method, the method comprising: Obtain the paper dataset and preprocess it; Authors are blocked after preprocessing based on their name and fingerprint. The first phase of merging is conducted among authors after blocking, based on similarity between the same institution and text. Based on bidirectional collaborator coverage, time overlap constraints, and institutional transfer priors, a second-stage cross-institutional determination is made for the authors after the first-stage merger. Constrained graph clustering was performed on the authors after the second-stage determination; Output the author entity cluster and related evidence.

[0007] Optionally, the preprocessing includes name normalization, organization normalization, text vector extraction, and collaborator profile construction.

[0008] Optionally, the name normalization includes converting traditional Chinese names to simplified Chinese names, converting full-width characters to half-width characters, standardizing the pinyin representation of Chinese names, and processing abbreviations of names.

[0009] Optionally, the institution normalization includes unifying the way institution names are expressed, mapping institutions to standardized hierarchical IDs based on a thesaurus, fuzzy matching, and coding trees, and handling equivalent sets of cases such as institution name changes, campus changes, and school mergers.

[0010] Optionally, the calculation of the bidirectional collaborator coverage includes calculating the coverage of the first author set to the second author set, calculating the coverage of the second author set to the first author set, and calculating the harmonic mean of the two-way coverage as the bidirectional collaborator coverage.

[0011] Optionally, the determination of the time crossover constraint includes calculating the time interval overlap of the two author clusters. When the overlap exceeds a preset threshold and there is no evidence of multiple affiliations, the merging is rejected.

[0012] Optionally, the calculation of the institutional transfer prior is based on the migration probability of subject area, geographical location, and institutional level.

[0013] Optionally, the second-stage cross-institutional determination also includes a comprehensive score, the comprehensive score function being Score=α·semantic similarity+β·bidirectional collaborator coverage+γ·institutional transfer prior-δ·temporal overlap, where α, β, γ, and δ are weight parameters.

[0014] Optionally, the constrained graph clustering includes constructing a weighted graph, where nodes in the weighted graph represent author clusters, edge weights are the combined similarity scores between two clusters, and constrained hierarchical clustering is performed based on the weighted graph and iteratively updated until convergence.

[0015] On the other hand, an author name disambiguation system is provided, including a data acquisition module, a preprocessing module, a blocking module, a first-stage merging module, a second-stage judgment module, a clustering module, and a result output module; The data acquisition module is used to acquire the thesis dataset; The preprocessing module is used to preprocess the paper dataset; The blocking module is used to block the preprocessed authors based on their name fingerprints; The first-stage merging module is used to perform a first-stage merging of authors after blocking based on the same institution and text similarity; The second-stage determination module is used to perform a second-stage cross-institutional determination on the authors merged in the first stage based on bidirectional collaborator coverage, time crossover constraints, and institutional transfer priors. The clustering module is used to perform constrained graph clustering on the authors after the second-stage determination; The result output module is used to output the author entity cluster and related evidence. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the author name disambiguation method of the present invention; Figure 2 This describes the calculation process for Bidirectional Collaborator Coverage (BCC) in this embodiment of the invention. Figure 3 This is the determination process for the Time Crossover Constraint (CTOC) in this embodiment of the invention; Figure 4 This is an architecture diagram of the system in an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0018] In this article, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0019] The author name disambiguation method and system described in this application will be explained in detail below through specific embodiments.

[0020] Example 1 like Figures 1-3 The flowchart shown illustrates an author name disambiguation method. An author name disambiguation method, comprising: Step 101: Obtain the paper dataset and perform preprocessing.

[0021] This embodiment uses "disambiguation of the names of academic paper authors in the field of computer science" as the application scenario, and the data source is 100,000 papers included in IEEE Xplore and ACM Digital Library from 2010 to 2024.

[0022] Obtain the complete metadata and abstracts of the papers from the above databases, including paper titles, abstract texts, author lists (with "corresponding author" and "first author" markings), original institution strings (such as "Tsinghua Univ, Dept Comp Sci & Technol", "Department of Computer Science and Technology, Tsinghua University"), publication time (accurate to year, month, and day, such as 2020-05-18), keywords, email addresses of some authors (such as zhangsan@tsinghua.edu.cn), and ORCID (such as 0000-0002-1234-5678).

[0023] Perform preprocessing on the obtained data, including name normalization, institution normalization, text vector extraction, and co-author portrait construction.

[0024] Among them, name normalization includes converting traditional Chinese names to simplified Chinese, converting full-width characters to half-width characters, unifying the pinyin representation of Chinese names, and handling abbreviated forms of names.

[0025] In one example, there are the following examples of name normalization.

[0026] Input the traditional Chinese names "陳曉明" and "黃麗娜", and convert them to simplified Chinese "陈晓明" and "黄丽娜" through the "OpenCC" traditional and simplified conversion tool; input "ZhangWei (traditional Chinese name pinyin: ChangWei)", and uniformly convert it to "张伟 (ZhangWei)".

[0027] Input the full-width character names "ZhangSan" and "LiWei", and convert them to half-width "ZhangSan" and "LiWei" through character encoding conversion (full-width ASCII code - 65248 = half-width ASCII code); input the mixed full-width and half-width "张Wei", and convert it to "张伟 (ZhangWei)".

[0028] Input the pinyin "ZhangWei", "ZhangW.", and "WeiZhang", and uniformly standardize them in the format of "surname pinyin full spelling + name pinyin full spelling (first letter capitalized)" as "ZhangWei"; input "ZhangWei (pinyin tone: ZhāngWěi)", and standardize it to "ZhangWei" after removing the tone.

[0029] Enter the abbreviations “Z.San”, “L.Wei”, and “ChenX.M.”, and complete them to “ZhangSan”, “LiWei”, and “ChenXiaoming” by combining them with the “Co-author Names” and “Institution Information” in the paper’s author list. If only the abbreviation “X.Zhang” exists and there is no other supporting information, it will be completed as “XiaoZhang” by default (common name completion, which will be further verified by collaborator information later).

[0030] Institutional normalization includes standardizing the way institution names are expressed, mapping institutions to standardized hierarchical IDs based on vocabularies, fuzzy matching, and coding trees, and handling equivalent sets such as institution name changes, campus changes, and school mergers.

[0031] In one example, organization normalization is shown below.

[0032] The original string "TsinghuaUniversity,DeptofComputerScience" or "Tsinghua University Computer Science and Technology Institute" is standardized to "Tsinghua University Computer Science and Technology Department"; the string "PekingUnivHealthSciCtr" is standardized to "Peking University Health Science and Technology Institute".

[0033] Based on the "National University and Research Institution Hierarchical Coding Tree" (Level 1: School / Research Institute, Level 2: College / Department, Level 3: Laboratory / Research Group), the standardized institution "Intelligent Systems Laboratory, Department of Computer Science and Technology, Tsinghua University" is hierarchically mapped to generate the ID "TSINGHUA-COMP-INTEL-001", where "TSINGHUA" is the school code, "COMP" is the college code (Department of Computer Science and Technology), and "INTEL" is the laboratory code (Intelligent Systems Laboratory).

[0034] Inputting "Beijing Institute of Posts and Telecommunications (before 2004)" and "Beijing University of Posts and Telecommunications (after 2004)" will group them into the same equivalent set, with a unified institution ID of "BUPT-000". For campus changes: inputting "Tsinghua University (Haidian Campus)" and "Tsinghua University (Shenzhen International Graduate School)" will map them as "TSINGHUA-HAIDIAN-001" and "TSINGHUA-SHENZHEN-001" respectively (not equivalent, distinguishing between campuses). For mergers: inputting "Former Beijing Medical University" and "Peking University School of Medicine (after merger in 2000)" will group them into the same equivalent set, with a unified ID of "PEKING-MED-000".

[0035] Step 102: Block the preprocessed authors by name and fingerprint; A "name fingerprint" is generated for the preprocessed authors. Based on the normalized name (e.g., "Zhang Wei" and "ZhangWei" are unified as "Zhang Wei_ZhangWei"), a unique fingerprint is generated by combining the first letter of the name's pinyin ("ZW") and the fixed code of the surname ("Zhang" corresponds to "ZHANG"). Authors with the same fingerprint are assigned to the same blocking group. For example, the "Zhang Wei_ZW_ZHANG" group contains all authors whose normalized name is "Zhang Wei". A total of 120 blocking groups are generated, with the number of authors in each group ranging from 50 to 2000, which significantly reduces the computational complexity of the subsequent process.

[0036] Step 103: Perform the first-stage merging of blocked authors based on the similarity of the same institution and text; For authors within each blocked group, the "same institution" criteria are first determined. If two authors have the same normalized institution ID (e.g., "Department of Computer Science and Technology, Tsinghua University" is mapped to "TSINGHUA-COMP-001"), they are listed as candidate pairs. Next, text similarity is calculated: a BERT pre-trained model is used to extract semantic vectors from the abstracts of the two authors' papers, and cosine similarity (Sim_sem) is calculated. A threshold T1=0.8 is set. If Sim_sem≥0.8, they are merged into the same author cluster. For example, within the blocked group, "Zhang Wei (Institution ID: TSINGHUA-COMP-001, Paper Title: Machine Learning)" and "Zhang Wei (Institution ID: TSINGHUA-COMP-001, Paper Title: Deep Learning)" have a Sim_sem=0.85, satisfying the condition, and are merged into cluster A.

[0037] Step 104: Based on bidirectional collaborator coverage, time cross-constraints, and institutional transfer priors, conduct the second-stage cross-institutional determination of the authors after the first-stage merger; Two-way collaborator coverage is one of the core innovations of this invention. Traditional collaborator similarity calculations typically use the Jaccard coefficient or cosine similarity, which only considers the number of co-collaborators and ignores the directionality of collaborator relationships, time decay effects, and the importance of author position.

[0038] Specific implementation:

[0039]

[0040] in: Coverage(A→B) represents the degree to which the set of collaborators of author A covers the set of collaborators of author B; a(c) represents the collaborator weight, which takes into account frequency, time decay and author position. match(c,CB) represents the collaborator matching function, which allows fuzzy name matching combined with auxiliary information such as organization / email / ORCID; BCC is the harmonic mean of the two-way coverage, with a value range of [0,1].

[0041] In one example, for “cross-agency candidate pairs” that were not merged in the first phase (such as cluster A: agency IDTSINGHUA-COMP-001, cluster B: agency IDPEKING-COMP-002), the following operations are performed.

[0042] Bidirectional collaborator coverage (BCC) was calculated, yielding a BCC of 0.676 for clusters A and B. Cluster A's timeframe was 2018-2022, and cluster B's was 2020-2024, with an overlap of 0.5. There was no evidence of multiple affiliations, initially classifying it as a "potential conflict." Further investigation revealed that an author of cluster A had a record of being a "visiting scholar at Peking University" from 2020-2022 (evidence of multiple affiliations), thus resolving the conflict. Both clusters A and B belong to the field of "Computer Science," are located in Beijing, and are affiliated with "985 universities." The time interval Δt = 2 years, and ITP = 0.8. The score was 0.5928. An adaptive threshold of 0.5 was set; if the score ≥ the threshold, clusters A and B were merged.

[0043] The calculation of the bidirectional collaborator coverage includes calculating the coverage of the first author set to the second author set, calculating the coverage of the second author set to the first author set, and calculating the harmonic mean of the two-way coverage as the bidirectional collaborator coverage.

[0044] In one example, we calculate the bidirectional collaborator coverage (BCC) using “Author Cluster A (Zhang Wei, Computer Science)” and “Author Cluster B (Zhang Wei, Computer Science)” as examples.

[0045] The set of collaborators CA of cluster A is: {C1 (weight 0.8, collaborator in 2023, corresponding author), C2 (weight 0.6, collaborator in 2022, first author), C3 (weight 0.4, collaborator in 2021, other authors)}; The collaborator set CB of cluster B is: {C2 (weight 0.7, collaborator in 2023, corresponding author), C3 (weight 0.5, collaborator in 2022, first author), C4 (weight 0.3, collaborator in 2021, other authors)}.

[0046] Match collaborators in CA and CB: C2 (match), C3 (match), C1 (not match), C4 (not match); Numerator: Σ (weight of matching collaborator) = 0.6 + 0.4 = 1.0; Denominator: Σ (weights of all collaborators in CA) = 0.8 + 0.6 + 0.4 = 1.8; Coverage(A→B)=1.0 / 1.8≈0.588.

[0047] Match CB with collaborators in CA: C2 (match), C3 (match), C4 (not match), C1 (not match); Numerator: Σ (weight of matching collaborator) = 0.7 + 0.5 = 1.2; Denominator: Σ (weights of all collaborators in CB) = 0.7 + 0.5 + 0.3 = 1.5; Coverage(B→A)=1.2 / 1.5=0.8.

[0048] BCC=2×Coverage(A→B)×Coverage(B→A) / (Coverage(A→B)+Coverage(B→A))=2×0.588×0.8 / (0.588+0.8)=0.9408 / 1.388≈0.676.

[0049] The determination of the time crossover constraint includes calculating the time interval overlap of the two author clusters. When the overlap exceeds a preset threshold and there is no evidence of multiple affiliations, merging is rejected.

[0050] The temporal overlap hard constraint is a mechanism specifically designed to identify changes in author affiliation. This mechanism is based on a reasonable assumption: if two authors with the same name are affiliated with different institutions during the same time period, and there is no clear evidence of multiple affiliations, then they are likely different people.

[0051] Specifically Overlap(A,B)=|[tA_start,tA_end]∩[tB_start,tB_end]| / min(tA_end-tA_start,tB_end-tB_start) ifOverlap(A,B)>τandnothas_multi_aff_evidence(A,B): Merge refused (CTOC conflict) in, Overlap(A,B) represents the degree of overlap in the time intervals of two author clusters; τ is the time crossover threshold, which is usually set to 0.25-0.35; The `has_multi_aff_evidence(A,B)` function is used to detect the existence of evidence of multiple aff relationships.

[0052] In one example, taking "Author Cluster A (Zhang Wei, Institution: Tsinghua University, Publication Date: 2018-03-01 to 2022-12-31)" and "Author Cluster B (Zhang Wei, Institution: Peking University, Publication Date: 2020-05-01 to 2024-06-30)" as examples, the Time Cross Constraint (CTOC) decision is performed.

[0053] The time interval of cluster A [tA_start,tA_end] = [2018-03-01,2022-12-31], and the interval length is 4 years and 10 months (approximately 4.83 years). The time interval of cluster B [tB_start,tB_end] = [2020-05-01,2024-06-30], and the interval length is 4 years and 2 months (approximately 4.17 years).

[0054] The overlap interval [t_overlap_start, t_overlap_end] = [2020-05-01, 2022-12-31], and the overlap length is 2 years and 8 months (approximately 2.67 years). Overlap(A,B) = Overlap length / min(interval length of cluster A, interval length of cluster B) = 2.67 / 4.17 ≈ 0.64.

[0055] The preset time crossover threshold τ = 0.3, 0.64 > 0.3, proceed to multi-membership evidence check; Examining the metadata of papers in clusters A and B reveals that in the three papers published by cluster A between 2020 and 2022, the authors' affiliations are listed as "visiting scholars from Tsinghua University and Peking University," while in the two papers published by cluster B during the same period, the authors' affiliations are listed as "part-time staff from Peking University and Tsinghua University." This demonstrates clear evidence of multiple affiliations.

[0056] Because of the existence of multiple membership evidence, even if Overlap(A,B)>τ, cluster A and cluster B are still allowed to merge; if no multiple membership evidence is found (e.g., cluster A's unit is only "Tsinghua University", and cluster B's unit is only "Peking University"), then merging is rejected (CTOC conflict).

[0057] The calculation of the institutional transfer prior is based on the migration probability of subject area, geographical location and institutional level.

[0058] Institutional transfer priors, based on statistical patterns of academic personnel mobility, provide prior knowledge for cross-institutional author identification. Significant differences exist in the probability of academic personnel mobility between institutions of different disciplines, regions, and levels.

[0059] Specific implementation:

[0060] in: "Field" refers to a subject area; "region" indicates a geographical location. tier indicates the level of an institution (such as 985 / 211 universities, regular undergraduate institutions, research institutes, etc.); Δt represents the time interval.

[0061] In one example, the Institutional Transfer Prior (ITP) is calculated based on four dimensions: “subject field”, “region”, “tier”, and “time interval”.

[0062] Subject areas: Computer Science (F1), Biomedicine (F2); Geographic location (region): First-tier cities (R1: Beijing, Shanghai, Guangzhou), New first-tier cities (R2: Hangzhou, Chengdu), Third-tier and below cities (R3); Institutional Tier: 985 / 211 Universities (T1), Ordinary Undergraduate Institutions (T2), Research Institutes (T3); Time interval (Δt): less than 1 year (Δt1), 1-3 years (Δt2), more than 3 years (Δt3); Based on statistics on the mobility of academic personnel, the migration probability (ITP base score) for each dimension combination is preset: for example, F1+R1+T1→F1+R1+T1+Δt2, migration probability 0.8 (ITP=0.8); F2+R3+T2→F2+R1+T3+Δt1, migration probability 0.3 (ITP=0.3).

[0063] Example 1: Cluster A (field=F1, region=R1, tier=T1, 2018-2022) → Cluster B (field=F1, region=R1, tier=T1, 2020-2024), Δt=2 years (Δt2), matches “F1+R1+T1→F1+R1+T1+Δt2”, ITP=0.8; Example 2: Cluster C (field=F2, region=R3, tier=T2, 2021-2023) → Cluster D (field=F2, region=R1, tier=T3, 2022-2024), Δt=1 year (Δt1), matches “F2+R3+T2→F2+R1+T3+Δt1”, ITP=0.3; Example 3: Cluster E (field=F1, region=R2, tier=T2, 2019-2021) → Cluster F (field=F2, region=R1, tier=T1, 2022-2024), Δt=3 years (Δt3), cross-domain (F1→F2) + cross-level (T2→T1) + Δt3, migration probability 0.15, ITP=0.15.

[0064] Step 105: Perform constrained graph clustering on the authors after the second-stage determination; The second stage of cross-institutional determination also includes a comprehensive score, the comprehensive score function being Score = α·semantic similarity + β·bidirectional collaborator coverage + γ·institutional transfer prior - δ·temporal overlap, where α, β, γ, and δ are weight parameters.

[0065] The comprehensive scoring function organically combines features from multiple dimensions to form a unified judgment standard. It balances the importance of different features through a weighted summation method.

[0066] Specific implementation:

[0067] in: Sim_sem represents the semantic similarity of texts; BCC indicates bidirectional collaborator coverage; ITP stands for Institutional Transfer Prior; Overlap indicates the degree of time overlap; α, β, γ, δ are weighting parameters that can be adjusted according to specific application scenarios.

[0068] Therefore, using the author clusters determined in the second stage as nodes, a weighted graph is constructed with "comprehensive score" as the edge weight (e.g., nodes: cluster A+B, cluster C, cluster D; edge weights: cluster A+B-cluster C=0.4, cluster A+B-cluster D=0.62, cluster C-cluster D=0.35). A constrained hierarchical clustering algorithm is adopted, and the hard constraint of "no CTOC conflict" is strictly followed during the clustering process: first, cluster A+B with the highest edge weight is merged with cluster D (weight 0.62, no conflict) to form a new cluster A+B+D; ​​then the edge weight between the new cluster and cluster C is calculated to be 0.45, which is still no conflict but has not reached the adaptive threshold (0.5), so merging is stopped; the characteristics of the clusters are iteratively updated (e.g., the time interval of the merged cluster is 2018-2024, and the collaborator set is the collaborator of clusters A, B, and D) until the number of clusters is stable (convergence), and finally 32 author entity clusters are obtained.

[0069] In one example, taking "cross-organizational determination of cluster A and cluster B" as an example, a comprehensive score is performed based on the calculation results of claims 6-8.

[0070] Text semantic similarity (Sim_sem): Papers from both cluster A and cluster B focus on "computer vision", and the cosine similarity calculated based on BERT vectors is 0.7; Bidirectional collaborator coverage (BCC): 0.676 calculated according to claim 6; Institutional Transfer Prior (ITP): Calculated as in Example 1 of claim 8, the result is 0.8; Temporal overlap: 0.64 calculated according to claim 7; Preset weight parameters: α=0.4 (Sim_sem weight), β=0.3 (BCC weight), γ=0.2 (ITP weight), δ=0.1 (Overlap penalty weight).

[0071] Score=α·Sim_sem+β·BCC+γ·ITP-δ·Overlap =0.4×0.7+0.3×0.676+0.2×0.8-0.1×0.64 =0.28 + 0.2028 + 0.16 - 0.064 =0.5788.

[0072] Based on the number of authors in the blocking group (120 people) and data sparsity (clusters A and B each have 15 papers, and the data is sufficient), the adaptive threshold is calculated to be 0.5. Since Score = 0.5788 ≥ 0.5, cluster A and cluster B are determined to meet the merging condition.

[0073] The constrained graph clustering includes constructing a weighted graph, where nodes in the weighted graph represent author clusters and edge weights are the combined similarity scores between two clusters. Constrained hierarchical clustering is performed based on the weighted graph and iteratively updated until convergence. Those skilled in the art can implement this equilibrium control method and system based on the above description.

[0074] Specifically, constraint graph clustering optimizes and integrates the results of the previous steps. By constructing a weighted graph and performing hierarchical clustering, the final author entity cluster is obtained.

[0075] Graph construction: Nodes: Each node represents an author cluster. Edge weight: The weight of an edge is the combined similarity score between the two clusters. Constraints: Hard constraints such as time-interval constraints should be considered. Clustering algorithms: A constrained hierarchical clustering algorithm is adopted to respect hard constraints during the clustering process and ensure the rationality of the clustering results.

[0076] Iterative optimization: After each round of clustering, the feature representation of the cluster is updated, and the process is repeated iteratively until convergence.

[0077] defconstrained_graph_clustering(clusters): # Constructing a weighted graph graph=build_weighted_graph(clusters) #Perform constrained hierarchical clustering final_clusters=constrained_hierarchical_clustering(graph) # Iterate and update until convergence whilenotconverged(final_clusters): update_cluster_features(final_clusters) final_clusters=constrained_hierarchical_clustering(graph) return final_clusters In one example, constrained graph clustering is performed on the “Computer Science Domain Blocking Group (containing clusters A, B, C, and D)”.

[0078] Nodes: Cluster A, Cluster B, Cluster C, Cluster D (all are author clusters after the first phase of merging); Edge weights (overall score): AB=0.5788, AC=0.42, AD=0.38, BC=0.61, BD=0.45, CD=0.32; Constraints: No CTOC conflict (all cluster pairs were pre-checked, only CD had Overlap=0.4>τ=0.3 and no evidence of multiple memberships, marked as "No Merging Constraint").

[0079] Step 1: Sort the edges by weight from highest to lowest: BC (0.61), AB (0.5788), BD (0.45), AC (0.42), AD (0.38), CD (0.32, forbidden constraint). Step 2: Prioritize merging B and C (highest weight, no constraints) to form a new cluster B+C, and update the cluster characteristics (time interval: 2020-2024 for B + 2019-2023 → 2019-2024 for C; collaborator set: collaborators of B + collaborators of C). Step 3: Calculate the edge weights of the new cluster B+C with other clusters: B+CA=0.52 (recalculated based on the merged semantic vector and collaborator set), B+CD=0.41; Step 4: Merge B+C with A (weight 0.52 ≥ threshold 0.5, no constraints) to form a new cluster A+B+C; Step 5: Calculate the edge weights of A+B+C and D. The weights are 0.39 < 0.5, and CD have forbidden constraints. Stop merging.

[0080] Update the cluster features of A+B+C (time interval 2018-2024, collaborator set merge, semantic vector recalculation), and check all cluster pairs again (only A+B+C and D). The edge weight is still 0.39 < 0.5, there is no possibility of new merging, the clustering converges, and finally two author entity clusters are obtained: A+B+C and D.

[0081] Step 106: Output the author entity cluster and related evidence.

[0082] Furthermore, the output includes information for each entity cluster, including all papers within the cluster (title, publication date, institution), author names and aliases (e.g., "Zhang Wei", "Zhang W.", "Wei Zhang"), institution change timeline (2018-2020: Tsinghua University, 2020-2022: Tsinghua University + Peking University visiting scholars, 2022-2024: Peking University), key evidence (BCC=0.676, ITP=0.8, multiple affiliation records), and confidence level (0.92). Simultaneously, the collaboration network (nodes represent collaborators, edges represent collaboration frequency) and the evolution of research topics (keyword cloud map) are displayed in graphical form.

[0083] Example 2 Reference Figure 4 As shown in the figure, this application provides an author name disambiguation system, including a data acquisition module, a preprocessing module, a blocking module, a first-stage merging module, a second-stage determination module, a clustering module, and a result output module.

[0084] The data acquisition module is used to acquire paper datasets. It uses the open API interfaces of IEEEXplore and PubMed to acquire paper metadata and abstracts in batches, storing the data in a PostgreSQL database. It supports filtering data by "domain (e.g., computer science, biomedicine)" and "time range (e.g., 2010-2024)," with a maximum single acquisition volume of 100,000 records per hour.

[0085] The preprocessing module is used to preprocess the paper dataset. It calls the name normalization subroutine (performing simplified / traditional Chinese conversion, full / half-width character conversion, etc.), the institution normalization subroutine (performing institution name standardization, hierarchical mapping, etc.), the text vector extraction subroutine (generating semantic vectors based on the BERT model), and the collaborator profile construction subroutine (statistically calculating collaborator weights). The preprocessed data is then output to the blocking module, with a processing speed of ≥1000 papers / hour.

[0086] The blocking module is used to block preprocessed authors based on their name fingerprints. It receives normalized name data from the preprocessing module, generates "name fingerprints" and groups them, and sends the blocked group data to the first-stage merging module, supporting parallel processing of 10 blocking groups.

[0087] The collaborator profile construction includes collecting collaborator information for each author, calculating collaborator weights, and considering collaboration frequency, time decay, and author ranking.

[0088] defpreprocess_and_normalize(papers): #Name Normalization normalized_names=normalize_names(papers.authors) #Institutional normalization normalized_affiliations=normalize_affiliations(papers.affiliations) #Text Vector Extraction text_vectors=extract_text_embeddings(papers.abstracts) #Collabor Profile Building coauthor_profiles=build_coauthor_profiles(papers) return{ 'names':normalized_names, 'affiliations':normalized_affiliations, 'text_vectors':text_vectors, 'coauthor_profiles':coauthor_profiles } Furthermore, the same-name blocking step is a preprocessing step designed to reduce computational complexity. By grouping authors with the same or similar names, the complexity of subsequent calculations can be significantly reduced.

[0089] The blocking strategy is as follows: Name fingerprint generation: Generate a name fingerprint for each author, considering different representations of the name; Group processing: Assign authors with the same name fingerprint to the same blocking group; Candidate pair generation: Generate candidate author pairs within each blocking group.

[0090] defblock_by_name_fingerprint(processed_data): blocks = defaultdict(list) forpaperinprocessed_data: # Generate name fingerprints, considering different name representation formats name_fingerprint=generate_name_fingerprint(paper.name) blocks[name_fingerprint].append(paper) returnblocks The first-stage merging module is used to perform a first-stage merging of blocked authors based on the same institution and text similarity; it reads the blocked group data and the "institution ID" and "text vector" output by the preprocessing module, determines candidate pairs of the same institution and calculates the text similarity, merges clusters according to the threshold T1=0.8, and sends the merged cluster data to the second-stage determination module.

[0091] The second-stage determination module is used to perform a second-stage cross-institutional determination on the authors merged in the first stage based on bidirectional collaborator coverage, time crossover constraints, and institutional transfer priors. It integrates the BCC calculation subroutine, CTOC determination subroutine, ITP calculation subroutine, and comprehensive scoring subroutine, reads the first-stage cluster data, calculates BCC, ITP, and time overlap, performs CTOC constraint checks and comprehensive scoring, and sends the cluster data determined to be "mergeable" to the clustering module.

[0092] The clustering module is used to perform constrained graph clustering on the authors after the second stage determination; it receives the second stage cluster data, constructs a weighted graph and performs constrained hierarchical clustering, iteratively updates the cluster features until convergence, and sends the final cluster data to the result output module.

[0093] The output module is used to output author entity clusters and related evidence. It reads the final cluster data from the clustering module and generates an "Author Entity Cluster Report" (including an institutional change timeline and key evidence), which supports PDF export and visualization (collaboration network, topic evolution diagram), with a response time of ≤3 seconds per query.

[0094] The above description is merely an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for disambiguating author names, characterized in that, The method includes: Obtain the paper dataset and preprocess it; Authors are blocked after preprocessing based on their name and fingerprint. The first phase of merging is conducted among authors after blocking, based on similarity between the same institution and text. Based on bidirectional collaborator coverage, time overlap constraints, and institutional transfer priors, a second-stage cross-institutional determination is made for the authors after the first-stage merger. Constrained graph clustering was performed on the authors after the second-stage determination; Output the author entity cluster and related evidence.

2. The method according to claim 1, characterized in that, The preprocessing includes name normalization, organization normalization, text vector extraction, and collaborator profile construction.

3. The method according to claim 2, characterized in that, The name normalization includes converting traditional Chinese names to simplified Chinese names, converting full-width characters to half-width characters, standardizing the pinyin representation of Chinese names, and processing abbreviations of names.

4. The method according to claim 2, characterized in that, The institution normalization includes unifying the way institution names are expressed, mapping institutions to standardized hierarchical IDs based on vocabularies, fuzzy matching, and coding trees, and handling equivalent sets such as institution name changes, campus changes, and school mergers.

5. The method according to claim 1, characterized in that, The calculation of the bidirectional collaborator coverage includes calculating the coverage of the first author set to the second author set, calculating the coverage of the second author set to the first author set, and calculating the harmonic mean of the two-way coverage as the bidirectional collaborator coverage.

6. The method according to claim 1, characterized in that, The determination of the time crossover constraint includes calculating the time interval overlap of the two author clusters. When the overlap exceeds a preset threshold and there is no evidence of multiple affiliations, merging is rejected.

7. The method according to claim 1, characterized in that, The calculation of the institutional transfer prior is based on the migration probability of subject area, geographical location, and institutional level.

8. The method according to claim 1, characterized in that, The second stage of cross-institutional determination also includes a comprehensive score, the comprehensive score function being Score = α·semantic similarity + β·bidirectional collaborator coverage + γ·institutional transfer prior - δ·temporal overlap, where α, β, γ, and δ are weight parameters.

9. The method according to claim 1, characterized in that, The constrained graph clustering includes constructing a weighted graph, where nodes in the weighted graph represent author clusters and edge weights are the combined similarity scores between two clusters. Constrained hierarchical clustering is performed based on the weighted graph and iteratively updated until convergence.

10. An author name disambiguation system implementing the method of claim 1, characterized in that, It includes a data acquisition module, a preprocessing module, a blocking module, a first-stage merging module, a second-stage decision module, a clustering module, and a result output module; The data acquisition module is used to acquire the thesis dataset; The preprocessing module is used to preprocess the paper dataset; The blocking module is used to block the preprocessed authors based on their name fingerprints; The first-stage merging module is used to perform a first-stage merging of authors after blocking based on the same institution and text similarity; The second-stage determination module is used to perform a second-stage cross-institutional determination on the authors merged in the first stage based on bidirectional collaborator coverage, time crossover constraints, and institutional transfer priors. The clustering module is used to perform constrained graph clustering on the authors after the second-stage determination; The result output module is used to output the author entity cluster and related evidence.

Images