docs/columns.md

Columns

Part of MBDef. The columns of the five tables, one list per table. What a row is: What one row is.

terms

• concept_id, term: the concept and its name (its most frequent spelling).
• spellings: every spelling, with how many papers write it.
• papers: how many papers define the concept. strict: true when at least one of its rows passes the stricter cut.
• definitions, description, synonyms, related_terms, subtypes, supertypes, examples: the merged items, most often stated first (how items merge). Each item has:
  • text: its most frequent form;
  • papers: how many papers state it;
  • links_to: for names, the concept the item names, if it is one of the dataset's concepts;
  • variants: the forms that were merged, with their papers (empty when there was one form);
  • sources: every paper and row that states it (paper_id, mention_id, def_id for definitions, as_written when the paper's wording differs from text, and citations).

concepts

• concept_id: a stable id ("g" plus 12 hex digits).
• label: the most frequent spelling. labels: all spellings, most frequent first.
• match_key: the normalised word set that the spellings share.
• n_instances, n_papers, n_definitions: the counts. n_strict: how many of the concept's rows pass the stricter cut.
• merge_flags: for concepts that join different spellings, the kinds of join someone may want to check: same words in a different order, spacing, a dropped bracket or acronym, a plural with two possible singulars, an irregular or "-ics" plural, British spelling. 1,052 concepts have at least one flag.
• joined_by: every key that two or more of the concept's spellings share, with those spellings (tracing a merge).
• mention_ids: the concept's rows in instances.

instances

• mention_id: the row id. concept_id: the concept the row belongs to.
• term: the term as the paper writes it (its original label).
• word_set_key, spacing_key: the two normalised keys the merging compares.
• normalisation: each step that changed the term before matching, in plain words.
• paper_id: the source paper.
• corpus: seeds (found by the literature search) or closure (cited by a seed paper).
• n_definitions: how many definitions the row has.
• ensemble_score: the relevance score that decided the row, between 0 and 1. strict: true when the row also passes the stricter cut (0.5850).
• classifier_score, basal_score, judge_score: the three member scores.
• classifier_ref_rank, basal_ref_rank, judge_ref_rank: each member score placed among that member's scores on the labelled training records, 0 to 1 (how scores are combined).
• description, synonyms, related_terms, subtypes, supertypes, examples: everything else the extraction lists for the term in this paper, each a list of {text, citations} (empty when the paper gives none; what they are).

definitions

• def_id: the row id.
• mention_id, concept_id, paper_id, term: the row's term-in-paper, concept, paper and term.
• text: the definition as quoted.
• citations: the references the paper gives for it, as written. 40,927 definitions have at least one.

papers

• paper_id: the paper's id in this dataset. corpus: seeds or closure.
• doi: known for 5,287 papers. title.
• year: known for 5,289 papers (5,086 from OpenAlex).
• openalex_id. oa_status: OpenAlex's open-access status.
• open_licence: the most open licence that OpenAlex records for any version of the paper. licence_url: where that version is.
• n_instances, n_definitions: the counts.