ANTHEM-202605 — an Enriched ACL Anthology for Scholarly Comparison

117,386 ACL Anthology papers linked to ROR-normalised author affiliations and Computer Science Ontology (CSO) topic labels, with an evidence span retained for every extraction decision.

Code: https://github.com/MoritzStaudinger/anthem

Configs

Config	Rows	What
papers	117,386	one row per paper; nested fields kept as JSON strings
affiliations	516,008	flat: paper x author x affiliation, with ROR id and evidence
topics	1,447,513	flat: paper x CSO label, source = syntactic | semantic
fulltext	117,379	parsed text per paper: GROBID TEI, PyMuPDF/LLM Markdown, OCR
affiliation_gold	781	hand-annotated affiliation gold set (781 papers)
benchmark_questions	99	comparative-QA benchmark: question, persona, the ROR/topic/year gold specification, and the resolved gold_paper_ids
cso_validation	104	hand-judged CSO topic labels
linking_adjudication	805	hand-verified ROR linking decisions, four strategies side by side
lora_train	10,000	LoRA fine-tuning data for the affiliation adapter (train / validation splits)

Nested structures are stored as JSON strings because record shape varies across 1952--2025; the flat affiliations and topics configs are the ones to query.

from datasets import load_dataset

aff = load_dataset("Molarix/anthem", "affiliations", split="train")

fulltext is the large config — load it only when you need the text itself.

Licensing

ANTHEM's own layer — the extracted affiliations, ROR links, CSO labels and evidence spans in papers, affiliations and topics — is released under CC-BY-NC-SA-4.0.

The license column is provenance, not the release licence: it records the terms the underlying paper carries in the Anthology. Per the ACL Anthology copyright policy, materials prior to 2016 are CC BY-NC-SA 3.0 and materials from 2016 onward are CC BY 4.0. Papers whose year could not be determined are assigned the more restrictive pre-2016 term.

9,470 papers were published at third-party (ELRA-published) venues such as LREC. The ACL copyright policy states that it does not cover third-party materials, so these rows carry third_party_venue = true; filter on it if your use needs only ACL-licensed material.

Source PDFs are not redistributed; the Anthology serves those. fulltext contains parsed renderings (TEI, Markdown, OCR text) produced by this pipeline.

Underlying licence	Papers
CC-BY-4.0	69,297
CC-BY-NC-SA-3.0	38,618
third-party venue (ELRA-published); see respective organisation	9,471

Construction

Affiliations were extracted from parsed PDFs by an LLM pipeline over combined parser outputs (GROBID TEI + PyMuPDF/LLM Markdown + OCR), then linked to ROR with country context. The model recorded per paper:

provenance_model	Papers
gemini-2.5-flash	79,666
gemini-2.5-flash-lite	22,261
gemini-1.5-flash	15,449
None	10

Personal and sensitive information

The dataset joins author names to institutions and publication years. All of it is already public in the ACL Anthology, and affiliations are as printed at publication time rather than current.

The fulltext config is a faithful parse of the published PDFs, so where a paper printed author e-mail addresses in its header, the TEI and OCR text contain them too. They are not separated into a column, and no contact details beyond what the papers themselves print are added.

Known limitations

Affiliation extraction and linking reach F1 0.760 (hierarchy-aware, parent and child ROR ids treated as equivalent) against a 781-paper decade-stratified gold set; ROR linking alone reaches 0.857 on hand-checked affiliation strings; CSO topic-label precision is 0.713. Roughly 90% of remaining false negatives are linking rather than extraction failures, concentrated in city-level institution clusters and geographically branched corporate entities.

This is a fixed snapshot (ANTHEM-202605); it is not updated in place.

Citation

Please cite the ANTHEM paper; the current reference is kept in the code repository.