--- name: acl-related-work description: Use when positioning an ACL submission against the NLP literature, covering ACL Anthology citation practice, arXiv-versus-published version citation, concurrent LLM-era preprints, prior-cycle ARR resubmission overlap, anonymity-preserving self-citation, and the fast-moving baseline problem in computational linguistics. --- # ACL Related Work Use this to audit positioning and citation hygiene. NLP moves fast enough that related-work sections age between ARR cycles; the section must prove the authors know the field *as of this cycle*, not as of the project's start. ## The Anthology-first rule - When a paper exists in the ACL Anthology (or another archival venue), cite that version, not its arXiv preprint — Anthology entries are the community's citation of record and carry stable BibTeX. - The Anthology is fully open access (post-2016 materials under CC BY 4.0), so "couldn't access the published version" is never a defense here. - For findings-tier work, cite it as Findings of ACL/EMNLP — it is archival and reviewers treat miscrediting it as sloppiness. ## Freshness pressure unique to this field - Expect at least one reviewer to name a baseline or dataset from the last two conference rounds. Sweep ACL, EMNLP, NAACL, EACL, TACL, and CL for the nearest neighbors published since your project began. - LLM-era concurrency is the norm: assume 2-4 similar preprints exist. Cite the ones findable at submission, contrast on technique rather than date, and avoid priority wars that reviewers cannot adjudicate. - If a newer model release makes your comparison set look dated mid-cycle, address it in the author response and plan the upgrade for camera-ready or resubmission — silence reads as unawareness. ## Positioning moves that work at ACL | Contribution type | The contrast reviewers want spelled out | |---|---| | Method | Nearest prior method + the mechanism-level difference, not just scores | | Resource | Existing datasets' coverage gaps: languages, domains, phenomena, license | | Analysis / evaluation | What prior evaluations measured incorrectly or could not see | | Metric | Failure cases of the incumbent metric, with correlation evidence | | Theme-track paper | How the work answers the year's theme question (ACL 2026: explainability) | One paragraph of typed contrast per lane beats an undifferentiated citation inventory; ACL's compressed format cannot afford taxonomy for its own sake. ## Self-citation and resubmission overlap - Cite your own prior work in third person ("Building on Smith (2024)...") so anonymity survives; heavy reliance on one group's unpublished thread is itself an anonymity leak. - If this is an ARR resubmission, the revision note — not the related-work section — is where you discuss the previous version; the paper itself should read cleanly to a fresh reviewer. - Workshop and non-archival versions (most \*ACL workshops are archival — check!) must be declared under the dual-submission policy rather than quietly extended. ## Preprint-era eligibility judgment calls - Your own arXiv preprint of this submission is allowed (no anonymity period since January 2024), but the declared preprint status must match reality. - A competing preprint that appeared after your ARR submission date is concurrent work: cite it in camera-ready, note the timing, contrast honestly. - Do not cite an inaccessible or anonymized document as supporting evidence — ARR forbids leaning on materials reviewers cannot read. ## Two-pass audit procedure ```text Pass 1 (coverage): for each claim of novelty, list the 3 nearest works from Anthology search + recent cycles; verify each is cited and contrasted, not just listed. Pass 2 (hygiene): for each bib entry, check venue-of-record vs arXiv, Findings attribution, third-person self-citation, and that no cited item is reviewer-inaccessible. ``` ## Where to actually search, in order 1. ACL Anthology full-text/venue search — the canonical record for every \*ACL main conference, Findings volume, TACL, CL, and most workshops. 2. The most recent two conference rounds' accepted-paper lists (each edition's site publishes them) — the freshness sweep reviewers run. 3. Semantic Scholar / Google Scholar citation graphs walked *forward* from your three nearest neighbors — who already built on them? 4. arXiv cs.CL for the concurrent-preprint scan, dated against your cycle's submission deadline so "concurrent" is defensible. 5. Leaderboards and dataset pages for your benchmarks — the SOTA row often names an uncited system. ## Related-work paragraph shape that survives compression For each literature lane, three sentences: (1) what the lane established, with its two or three best citations; (2) the specific gap or assumption it left; (3) how this paper differs mechanically. Delete any sentence that neither cites nor contrasts. In a short paper, the whole section may be one such paragraph plus contrasts woven into the introduction — that is normal and reviewers at ACL expect it. ## Red flags reviewers name explicitly - "The authors seem unaware of X" — the fatal version of staleness; one missed obvious neighbor outweighs fifty dutiful citations. - Citing a paper for a claim it does not make (checked more often than authors expect, since reviewers know their own subfield's papers). - A wall of "recently, LLMs have..." citations with no load-bearing role. - Comparing against numbers from a cited paper's arXiv v1 that changed in the published version. ## Output format ```text [Positioning verdict] sharp / inventory-like / stale / overlapping [Nearest neighbors] <3 works -> stated distinction each> [Freshness risks] [Citation hygiene] [Overlap declarations] ```