From raw affiliations to persistent identifiers

Frow raw affiliations to persistent identifiers Myrto Kallipoliti1, Serafeim Chatzopoulos2, Miriam Baglioni3, Eleni Adamidi2, Paris Koloveas2,4, Thanasis Vergoulis2 1 OpenAIRE AMKE, Athens, Greece 2ATHENA RC, Athens, Greece 3CNR-ISTI, Pisa, Italy 4University of the Peloponnese, Tripolis, Greece TPDL 2025, Tampere, Finland | 23-26 September 2025 Introduction •The growth rate of research products is constantly increasing •Why? •The number of researchers increases worldwide •Pressure to publish more – “publish or perish” Open Science Initiatives •Open Science Initiatives (e.g., EOSC, I4OC) •Large amounts of scholarly data are openly available •Scholarly data often represented as Scholarly Knowledge Graphs •E.g, OpenAIRE Graph, OpenAlex •Rich and relatively clean sources of information about research products Scholarly Knowledge Graphs (SKGs) •SKGs are inherently heterogeneous graphs •Research products and related entities are represented as nodes (with metadata) Scholarly Knowledge Graphs (SKGs) •SKGs are inherently heterogeneous graphs •Research products and related entities are represented as nodes (with metadata) •Connections between them are represented as edges (with semantics) CITES Scholarly Knowledge Graphs (SKGs) •SKGs are inherently heterogeneous graphs •Research products and related entities are represented as nodes (with metadata) •Connections between them are represented as edges (with semantics) •AFFILIATED_WITH relations are incomplete •Use available metadata to extract and enrich these relations CITES Motivation •Why affiliation relations matter? •Track institutional and national contributions to global challenges and innovation •Data interoperability across research databases •Large-scale data integration and bibliometric analyses Motivation •Why affiliation relations matter? •Track institutional and national contributions to global challenges and innovation •Data interoperability across research databases •Large-scale data integration and bibliometric analyses •Goal: Infer affiliation relations from the metadata •Map affiliation strings to organization persistent identifiers (like ROR ids) The Challenge •Raw affiliation stings are often unstructured •They can be inconsistent in formatting •Frequently reference multiple organizations in one string Our approach Our approach Preprocessing phase Matching phaseDisambiguation phase Preprocessing Phase •Cleaning and stemming •Lowercase stings, remove stopwords, multi-digit numbers etc •Stemming: reduce words to their root or base form to enable easier comparison and matching Preprocessing Phase •Cleaning and stemming •Lowercase stings, remove stopwords, multi-digit numbers etc •Stemming: reduce words to their root or base form to enable easier comparison and matching •Keyword labeling and partitioning •Splits affiliation strings into partitions (or segments) •Identifies common keywords (such as ‘hospital’, ‘university’ etc) and country names inside partitions Preprocessing Phase •Partition pruning •Partitions that do not contain any keywords are not considered for the upcoming similarity computations •Country names are set aside and considered later in the ‘Candidate identification’ step Preprocessing Phase •Partition pruning •Partitions that do not contain any keywords are not considered for the upcoming similarity computations •Country names are set aside and considered later in the ‘Candidate identification’ step •Shortening of the remaining partitions •Preserves terms near the keywords (using a window parameter) Matching & Disambiguation Phases •Candidate identification •Performs string similarity among remaining partitions and the names in the organization database •Restrict matching to organizations located in the countries mentioned in the affiliation string •Apply different similarity thresholds: for universities (sim_u) and other organizations (sim_o) Matching & Disambiguation Phases •Candidate identification •Performs string similarity among remaining partitions and the names in the organization database •Restrict matching to organizations located in the countries mentioned in the affiliation string •Apply different similarity thresholds: for universities (sim_u) and other organizations (sim_o) •Results refinement and disambiguation •If multiple candidate organization names are identified for a single partition, they are compared to the original (clean and stemmed) affiliation string to determine the best match •If multiple organizations with the same name exist, city information is considered to determine the best match Our approach AffRoDB: Expert-Curated Dataset •Affiliation strings to ROR ids mappings •Each entry independently annotated by at least two experts with one of the three categories: exact match, ancestor match, vague •Stats: 1,500 records from Crossref / 1,374 unique affiliation strings / 1,475 affiliation relations (25% exact / 75% ancestor) •All expert annotations are openly available on Zenodo* Regular updates: adding matches for new affiliation strings and revising existing entries h release will include a curated batch (~70% new strings, ~30% previously processed) to validate identifiers and catch missing matches *https://zenodo.org/records/15322098 AffRo parameter analysis •3 parameters: window, sim_u, sim_o •F1 score improves when increasing window –peaks at window = 3 •Precision increases with higher values of similarity thresholds •Higher thresholds (sim_o > 0.7, sim_u > 0.5) lead to lower recall Conclusions •Introduced AffRo algorithm to tackle the problem of affiliation matching •Released a fully expert-curated dataset to facilitate relevant studies Conclusions •Introduced AffRo algorithm to tackle the problem of affiliation matching •Released a fully expert-curated dataset to facilitate relevant studies Future work: •Use the structure of the SKGs to infer affiliation relations •Compare against recent studies: AffilGood* * N Duran-Silva, P Accuosto, P Przybyła, H Saggion: AffilGood: Building reliable institution name disambiguation tools to improve scientific literature analysis Thank you! [email protected] Code: https://github.com/mkallipo/affiliation-matching API: https://affro-api.imsi.athenarc.gr/docs Dataset: https://zenodo.org/records/15322098