Classifiers

bibr combines alias rules, local classifiers, and LLM fallbacks for section types, paper types, OECD domains, and citation linking. Multi-study heading scopes keep repeated Methods and Results sections attached to the right study.

Section classifier

Module: bibr/structure/section_classifier.py

The section classifier maps paper section headers to canonical IMRaD+ categories. It uses a three-tier cascade:

Tier 1: Lookup-based matching

First, headers are matched against a dictionary of known aliases (CANONICAL_SECTION_ALIASES in paper_contents.py). This handles common headers like "Introduction", "Methods", "Results", etc.

• Exact match -- normalized header text matches an alias directly (confidence: 1.0), bypassing both later tiers
• Substring match -- a trusted alias appears as a word-boundary substring of the header (confidence: 0.95), also bypassing both later tiers
• A generic alias hit that isn't trusted is kept as a low-confidence prior, used only if both later tiers return unknown

Tier 2: Trained classifier

For lookup misses, headers route through a trained MiniLM two-head model (ML_SECTION_CLASSIFIER_MODEL_ID, default scienceverse/bibr-section-classifier). It predicts the canonical type and whether the heading is top-level, using relative position and neighboring headings. Predictions below ML_SECTION_CLASSIFIER_MIN_CONFIDENCE (default 0.5) collapse to unknown. If the model is disabled or unavailable, the pipeline uses the LLM path. Classifier load and inference failures are recorded as degraded operation; ML_CLASSIFIERS_REQUIRED=true makes configured classifier availability a requirement instead.

Tier 3: LLM-based classification

Headers the trained model leaves as unknown (when ML_SECTION_CLASSIFIER_LLM_ESCALATION is on, the default) or that skip Tier 2 entirely are sent to the configured LLM provider in a batch. The LLM returns one of the canonical section values for each header (confidence: 0.85). Results are validated against the known section types. Ambiguous headings can include body-text snippets and study-scope context.

Exported section rows include classification_source alongside classification_score, so consumers can distinguish model, lookup, LLM, and later hierarchy decisions. With no_llm=True, heading classification uses the lookup-only path.

Canonical section types

Type	Value	Examples
Title	title	Paper title (root/level-0 heading)
Abstract	abstract	Abstract, Summary, Executive Summary, Statement of Relevance
Introduction	intro	Introduction, Background, Related Work, Literature Review
Methods	method	Methods, Materials and Methods, Study Design, Participants
Results	results	Results, Findings, Experimental Results
Discussion	discussion	Discussion, General Discussion, Limitations, Conclusion
References	references	References, Bibliography, Works Cited
Acknowledgment	acknowledgment	Acknowledgments, Declarations, Corresponding Author
Author Contributions	author_contributions	Author Contributions, CRediT Authorship Contribution Statement
Conflict of Interest	coi	Conflict of Interest, Competing Interests, Declaration of Interest
Ethics	ethics	Ethics Statement, IRB Approval, Informed Consent
Funding	funding	Funding, Financial Support
Keywords	keywords	Keywords
Endnote	endnote	Supplementary Materials, Future Work, Outlook
Appendix	appendix	Appendix, Supplementary Material, Supporting Information
Data Availability	open_data	Data Availability, Code Availability, Reproducibility Statement
Footnote	footnote	Footnotes, Notes
Table	table	Table (caption/label region)
Figure	figure	Figure (caption/label region)
Unknown	unknown	(fallback for unrecognized headers)

IMRaD enforcement

After classification, enforce_imrad_order() deduplicates sections that should only appear once. Most section types are repeatable -- real papers often have multiple subsections with the same classification.

Only two types are unique:

• Abstract -- a paper has one abstract
• References -- a paper has one bibliography

For duplicate Abstract or References sections, the highest-trust classification_source wins; document order breaks ties. Other candidates are reset to unknown with classification_source="imrad_dedup". A later clean alias heading can therefore supersede an earlier weak classification. Deduplication is global, including across study scopes. Every other canonical type is allowed to repeat, and section order is not forced into IMRaD order.

Layout hints

When the PP-DocLayoutV3 layout model tags a region with a semantic label (abstract, reference, footnote), PDFParser._handle_section_hint uses that label to create or reuse an implicit section for it directly — independent of, and prior to, header-based classification. This doesn't override an existing classification score; it's how sections without a text heading (e.g. a layout-detected reference block) get a section in the first place.

Front-role classifier

Module: bibr/extract/front_role.py (bundle loading, scoring), bibr/extract/front_role_features.py (the feature contract shared with the trainer), consumed by bibr/extract/front_matter.py and bibr/extract/ref_locator.py

An optional gradient-boosted model scores every OCR text region with a distribution over eleven roles: title, byline, affiliation, abstract, keywords, doi_line, masthead, heading, ref_header, body, other. Its features are the region's geometry in a page-relative frame, its font relative to the largest type on the page, and script-independent text shape (initial density, separator counts, affiliation and correspondence cues). It is trained from publisher JATS projected onto cached OCR regions, so its labels are verbatim ground truth rather than an LLM's opinion.

The scores are evidence, not decisions. Front-matter ownership (resolve_front_matter) keeps its lexical heuristics and adds the model's roles on top:

• a region the model calls a byline (probability at or above ML_FRONT_ROLE_MIN_CONFIDENCE, default 0.5) is admitted as a byline even when the English byline shape or the 45-word cap rejects it (consortium bylines, separator-rich house styles, non-Latin scripts), including page-1 rows the section classifier mistyped;
• a model title seeds a record for scripts the uppercase test cannot read;
• affiliation and abstract roles are added the same way;
• a confident masthead (ML_FRONT_ROLE_MASTHEAD_CONFIDENCE, default 0.8) cannot root a record unless layout labelled the region doc_title;
• a row the heuristics seeded as a title but the model confidently types as something else (ML_FRONT_ROLE_RECORD_ROOT_CONFIDENCE, default 0.9) keeps the title role and loses only the right to root a second record. Correspondence, A R T I C L E I N F O, CITATION and Key Features all score heading at 1.00 and all sit above an abstract, which is anatomy enough to develop a record — so admitting a correct byline above them used to cut the real title away from its own abstract. Layout's doc_title and a match against the reference parser's detected title both outrank the veto, and 1.0 disables it.

RefLocator accepts a heading the model scores as ref_header as the reference-section header, which covers the non-English headings (Literaturverzeichnis, Bibliografía, Список литературы, 参考文献) the English regex misses.

ML_FRONT_ROLE_MODEL_ID defaults to scienceverse/bibr-front-role-v1, pinned to a commit by ML_FRONT_ROLE_REVISION so a hub push cannot change front-matter output. Set the id to null to fall back to the lexical heuristics alone, or ML_FRONT_ROLE_ENABLED=false to keep a configured bundle unloaded. The bundle is a joblib pickle, read through the gadget-restricted loader — only point the setting at a checkpoint you control. scikit-learn and joblib are core dependencies, so the model runs on a torch-free install. Loading is soft: an unavailable bundle logs one warning and the pipeline runs on the heuristics alone. Resolutions influenced by the model carry the front_role_model reason flag, and every candidate records the roles the model contributed (model_roles) with its top scores (model_scores). Native DOCX/JATS/HTML inputs have no OCR regions and are never scored.

Paper type classifier

Module: bibr/structure/paper_classifier.py (taxonomy constants), bibr/extract/core_metadata.py (LLM classification, via CoreMetadataExtractor; bibr/extract/extractor.py delegates to it)

The default scienceverse/bibr-paper-classifier model is a SPECTER2-based multitask classifier. It reads the resolved title and abstract and predicts paper type together with OECD L1/L2. ML_PAPER_CLASSIFIER_MODEL_ID selects the model, and ML_PAPER_CLASSIFIER_REVISION pins its revision.

When the paper-type confidence falls below ML_PAPER_CLASSIFIER_MIN_CONFIDENCE (default 0.5), the pipeline can ask the LLM paper-type labeler to resolve it. This escalation is controlled by ML_PAPER_CLASSIFIER_LLM_ESCALATION (default true). If no model result is available, core metadata extraction falls back to LLM classification of paper type and OECD domains. The exported paper_type_confidence records the model or escalation confidence when available.

The output supports these types; notice labels are also handled by title guards during core metadata extraction:

Type	Description
empirical	Original research with data collection/analysis
review	Systematic review, literature review, scoping review
meta-analysis	Quantitative synthesis of multiple studies
case-study	Case study, case report, case series
commentary	Commentary, editorial, opinion piece
corrigendum	Notice amending a previously published article
erratum	Notice amending a previously published article
retraction	Notice withdrawing a previously published article

Unrecognized or uncertain values are coerced to null when the type cannot be determined, rather than an unknown sentinel.

OECD domain classifier

Module: bibr/structure/paper_classifier.py (taxonomy constants), bibr/extract/core_metadata.py (LLM classification, via CoreMetadataExtractor; bibr/extract/extractor.py delegates to it)

The same multitask classifier predicts the broad domain and subdomain in bibr's OECD Frascati Manual taxonomy. On the trained-model path, OECD labels are not sent to an LLM for confidence escalation. A subdomain below ML_PAPER_CLASSIFIER_L2_MIN_CONFIDENCE (default 0.5) is exported as null; setting the threshold to 0.0 disables that gate. oecd_confidence is the L1 confidence.

Level 1: Broad domain (6 categories)

Domain
Natural Sciences
Engineering and Technology
Medical and Health Sciences
Agricultural and Veterinary Sciences
Social Sciences
Humanities and the Arts

Level 2: Subdomain

Each L1 domain has 4-10 subcategories. For example, "Social Sciences" includes: Psychology and Cognitive Sciences, Economics and Business, Education, Sociology, Law, Political Science, Social and Economic Geography, Media and Communications.

When the trained classifier is disabled or unavailable, the LLM fallback returns both L1 and L2. Its labels are canonicalized against the taxonomy; a recognized L2 also determines its parent L1. Native formats that already provide structured metadata can bypass the core extraction/classification path entirely.

Multi-study hierarchy

Module: bibr/structure/section_tree.py

Headings such as Study 1, Study II, and Experiment A establish scopes before classification. A combined heading such as Study 2: Methods can supply both a study marker and a section label. Hierarchy reconstruction uses numbering, classifier top-level predictions, and scoped IMRaD anchors to keep one study's sections from becoming children of another study.

Citation linking

Module: bibr/structure/citation_linker.py

The citation linker detects inline citations in text and resolves them to bibliography entries. It uses a 3-tier hybrid approach:

Tier 1: Numeric citations

Bracket citations [1], [3-5], [1, 2, 3] and superscript citations ^{3}, ^{15,16} are detected and checked against bibliography IDs. Parenthesized numeric and flattened superscript styles have additional document-style and local-context guards to avoid treating equations, measurements, and ordinary digits as citations.

Tier 2: Author-year citations

Parenthetical (Smith, 2020) and narrative Smith (2020) citation patterns are detected via regex and fuzzy-matched against reference author/year fields.

Tier 3: LLM fallback

Remaining ambiguous or unresolved citation candidates are sent to the LLM for resolution against the reference list.

Resolved citations are stored as PaperXref objects with xref_type="bib". The exported extraction.diagnostics.citation_linking receipt records candidate spans, evidence, accepted/rejected decisions, and coverage; this makes unresolved citations visible alongside the successful xref rows. Rejected numeric candidates do not automatically become LLM requests.

refs="off" skips bibliographic citation linking along with reference extraction. no_llm=True skips citation linking as part of its reduced post-processing path.