--- pretty_name: 'LCRC: Local Citation Recommendation Corpus' license: cc-by-4.0 language: - en task_categories: - text-retrieval task_ids: - document-retrieval tags: - citation-recommendation - local-citation-recommendation - information-retrieval - astrophysics - earth-science - planetary-science - scientific-text size_categories: - 10K # LCRC: Local Citation Recommendation Corpus LCRC is a corpus for **local citation recommendation** in astrophysics, earth science, and planetary science. Each record is a single sentence drawn from the body of a review article, paired with the specific paper or papers its author cited at that point in the text. The cited papers are identified by DOI and by [ADS](https://ui.adsabs.harvard.edu/) bibcode. The sentence is provided in three forms — unmodified, with citation text deleted, and with citations replaced by a `[REF]` token — along with the position of the sentence in its source document and the publication date of the source document. Source documents are review articles spanning three subject areas — **astrophysics**, **earth science**, and **planetary science** — drawn from ADS by journal bibcode: `ARA&A`, `A&ARv`, `SSRv.`, `NewAR`, `LRSP.`, `P&SS.`, `RvGeo`, `ESRv.`, and `NRvEE`. No record cites a paper from those nine journals, so the source venues never appear as citation targets. Citation targets are predominantly primary research articles — `ApJ`, `MNRAS`, `A&A`, and `AJ` together account for 84% of the 9,980 distinct cited papers — but they are not exclusively so. Targets span 152 journal bibstems and include review articles in venues outside the nine excluded ones (18 in _Physics Reports_), conference proceedings (25 IAU Symposium, 15 ASP Conference Series), and 25 arXiv preprints. ## Dataset at a Glance | | | | --------------------------- | ---------------------------------------------------------- | | Records (citation contexts) | 14,735 | | Source documents | 354 review articles | | Distinct cited papers | 9,980 | | Source publication dates | 1976-01-01 – 2025-06-01 | | Cited papers per record | 1.28 mean, 1 median, 7 max | | Sentence length | 30 words mean, 28 median, 54 at p95, 455 max | | Files | `LCRC.jsonl` (single JSON Lines file, one record per line) | | Subject areas | Astrophysics, earth science, planetary science | | Language | English | Distribution of cited papers per record: | Cited papers | 1 | 2 | 3 | 4 | 5 | 6 | 7 | | ------------ | ------ | ----- | --- | --- | --- | --- | --- | | Records | 11,671 | 2,356 | 496 | 153 | 41 | 13 | 5 | ## Data Dictionary Every record has all eight keys; none are nullable. ### `source_doi` - **Type:** `string` - **Example:** `"10.1007/s00159-008-0010-0"` - DOI of the review article the sentence was taken from. Not unique across records — the 14,735 records come from 354 source documents, so a single DOI recurs across every sentence extracted from that article. ### `sent_original` - **Type:** `string` - **Example:** `"Novak et al. (2003) studied the dust emission at 450 μm in the galactic plane, and found the magnetic field to be well aligned parallel with the galactic plane (their Fig. 1), within 30 arcmin in longitude and 10 arcmin in latitude of the GC."` - The sentence as it appears in the source article, with inline citations intact. Contains author–year strings such as `Novak et al. (2003)`, which name the cited work directly. ### `sent_no_cit` - **Type:** `string` - **Example:** `"(), do show unusual, faint features in their surroundings."` - `sent_original` with the inline citation text deleted. Surrounding punctuation is left as-is, so the field commonly retains empty parentheses (`()`), doubled spaces, and leading whitespace where a citation began the sentence. No placeholder marks the deletion site. ### `sent_cit_masked` - **Type:** `string` - **Example:** `"([REF]), do show unusual, faint features in their surroundings."` - `sent_original` with each inline citation replaced by the literal token `[REF]`. Punctuation is otherwise preserved, so the token appears inside the original parentheses or other delimiters. A sentence citing several papers at one site carries a single `[REF]`; a sentence citing papers at distinct sites carries one `[REF]` per site, so the number of `[REF]` tokens need not equal `len(citation_dois)`. ### `sent_idx` - **Type:** `int` - **Example:** `1329` - Zero-based position of the sentence within the list of sentences segmented from the source article's body text, giving where in the document the citation context occurs. Indices are relative to that article only, and are contiguous per source document only to the extent that consecutive sentences were retained. Index values depend on the segmentation described in [Sentence Segmentation](#sentence-segmentation) and are not reproducible under a different segmenter or different parameters. ### `citation_dois` - **Type:** `list[string]` - **Example:** `["10.1086/311563"]` - DOIs of the papers cited at this point in the text — the ground-truth labels. Always non-empty. Order follows the order of appearance of the citations in `sent_original`. This is the complete set of citations in the sentence: records in which any citation could not be identified are absent from the corpus entirely, so no record carries an unlabeled citation. ### `resolved_bibcodes` - **Type:** `list[string]` - **Example:** `["1998ApJ...504L..23S"]` - ADS bibcodes for the papers in `citation_dois`, index-aligned and the same length. A bibcode encodes year, journal, volume, and first-author initial (`1998ApJ...504L..23S` is a 1998 _ApJ_ paper), so it doubles as a stable ADS identifier and a coarse description of the cited work. ### `pubdate` - **Type:** `int` - **Example:** `20080601` - Publication date of the source article as `YYYYMMDD`. Dates are approximate below month granularity: unspecified months and days in the ADS metadata appear as `01`, so a value such as `20080601` may encode "June 2008" rather than the first of the month, and `19760101` may encode only "1976". ## Sentence Segmentation Sentence boundaries, and therefore the values of `sent_idx`, were produced with [**`semantic-text-splitter`**](https://pypi.org/project/semantic-text-splitter/) **v0.22.0**, configured as: ```python from semantic_text_splitter import TextSplitter splitter = TextSplitter(capacity=1500, overlap=150) ``` Reproducing `sent_idx` requires this package at this version with these parameters; a different segmenter or different parameters could produce different sentence indexes. ## Data Characteristics Properties of the corpus that bear on how results should be read. - **Labels are complete per record.** `citation_dois` holds every citation in the sentence. Sentences in which any citation could not be identified were excluded wholesale rather than partially labeled, so a retrieved paper absent from the list is a true miss and not an unlabeled positive. - **Citations are resolved unambiguously.** A citation appears only where it matched exactly one entry in its source article's reference list, which keeps label precision high. The corresponding trade-off is coverage: citations whose first author and year collide with another reference in the same bibliography, and citations to papers lacking a DOI, are absent by construction. - **Review-article prose.** Source documents are reviews, which attribute individual claims to individual papers and cite more densely than primary research articles — the property that makes sentence-level citation contexts plentiful here. Citation behavior in reviews differs from that in research papers, so the corpus characterizes the former. - **Source documents are concentrated.** 14,735 records come from 354 articles. Records sharing a `source_doi` share topic, vocabulary, and a reference list, so they are not independent draws; partitioning the corpus by `source_doi` keeps reference lists from spanning partitions. - **Subject coverage.** All text comes from astrophysics, earth science, and planetary science, with vocabulary dense in instrument names, object and sample catalogs, and unit expressions. Records are not balanced across the three areas, and carry no subject label; subject can be approximated from the source journal by resolving `source_doi` through ADS. - **Cited papers follow a long tail.** 9,980 distinct papers are cited across 14,735 records. Most are cited once; influential papers recur many times. - **Every record cites at least one paper.** Sentences containing no citations are not part of this corpus, and sentences longer than 4,096 characters are not present. ## Licensing LCRC is released under [**CC BY 4.0**](https://creativecommons.org/licenses/by/4.0/). You may share and adapt the corpus, including commercially, with attribution. Bibliographic metadata derives from the [NASA Astrophysics Data System](https://ui.adsabs.harvard.edu/) and is included with ADS's permission. Work using this corpus should acknowledge ADS alongside the citation below. ## Citation LCRC is introduced and described in the following paper. If you use the dataset, please cite it: ```bibtex @inproceedings{basseri2026citation, title = {Local Citation Recommendation Without Graph Signals}, author = {Basseri, Benjamin and Protopapas, Pavlos and Zhang, Kemeng and Accomazzi, Alberto and Jarmak, Stephanie and Kelbert, Anna}, booktitle = {Proceedings of the (WASP 2026)}, year = {2026}, note = {To appear. Co-located with AACL-IJCNLP 2026.} } ```